Professional Documents
Culture Documents
Jack Krupansky
https://medium.com/@jackkrupansky/how-close-is-ai-to-human-level-intelligence-
here-in-april-2018-9a6ceaff2f9d
Sumário
How Close Is AI to Human-level Intelligence Here in April 2018?................................................. 9
Topics to be covered in this paper .......................................................................................... 13
Task-specific and domain-specific AI....................................................................................... 14
Learning and machine learning ............................................................................................... 14
Learning concepts ................................................................................................................... 14
Robotics vs. intellectual capacities.......................................................................................... 14
Is it AI or just automation? ...................................................................................................... 15
Are all heuristics AI? ............................................................................................................ 24
Robotics ............................................................................................................................... 24
Conclusion ........................................................................................................................... 25
Is it AI for just a heuristic? ....................................................................................................... 26
Progress on gaming ................................................................................................................. 26
Is it AI or just machine intelligence ......................................................................................... 27
Tasks beyond what people and animals traditionally performed ...................................... 30
Big learning and little learning ............................................................................................ 30
Training................................................................................................................................ 30
Deep learning and guided learning ..................................................................................... 31
Advanced machine learning ................................................................................................ 31
Automation ......................................................................................................................... 31
Patterns ............................................................................................................................... 32
Computer vision .................................................................................................................. 33
Internet of Things (IoT)........................................................................................................ 33
Cybersecurity....................................................................................................................... 34
Data analytics and business intelligence ............................................................................. 35
Scientific and engineering calculation and modelling......................................................... 35
Very complex data patterns and connections within data ................................................. 35
Summary ............................................................................................................................. 36
What fraction of strong AI is needed for your particular app? ............................................... 36
Areas of intelligence ............................................................................................................ 37
Areas of human intelligence ............................................................................................... 37
Areas of AI research ............................................................................................................ 38
Levels of function ................................................................................................................ 39
Specific human-level functions of intelligence.................................................................... 40
Degree of competence ........................................................................................................ 44
Common use of natural language processing (NLP) ........................................................... 45
Autonomy, principals, agents, and assistants ..................................................................... 46
Intelligent agents................................................................................................................. 46
Intelligent digital assistants ................................................................................................. 47
The robots are coming, to take all our jobs? ...................................................................... 47
How intelligent is an average worker? ................................................................................ 47
No sign of personal AI yet (strong AI) ................................................................................. 48
AI is generally not yet ready for consumers........................................................................ 49
Meaning and conceptual understanding ............................................................................ 49
Emotional intelligence......................................................................................................... 50
Wisdom, principles, and values........................................................................................... 50
Extreme AI ........................................................................................................................... 51
Ethics and liability................................................................................................................ 51
Dramatic breakthroughs needed ........................................................................................ 52
Fundamental computing model .......................................................................................... 53
How many years from research to practical application? .................................................. 53
Turing test for strong AI ...................................................................................................... 54
How to score the progress of AI .......................................................................................... 56
Links to my AI papers .......................................................................................................... 56
Conclusion: So, how long until we finally see strong AI? .................................................... 56
What Is AI (Artificial Intelligence)? .............................................................................................. 57
What is intelligence? ............................................................................................................... 60
Artificial intelligence is what we don’t know how to do yet ................................................... 62
Emotional intelligence............................................................................................................. 63
Autonomy, agency, and assistants .......................................................................................... 63
AI areas and capabilities.......................................................................................................... 64
Neural networks and deep learning ........................................................................................ 65
Animal AI ................................................................................................................................. 65
Robotics ................................................................................................................................... 65
Artificial life ............................................................................................................................. 66
Ethics ....................................................................................................................................... 66
Historical perspective by John McCarthy ................................................................................ 66
Can machines think? ............................................................................................................... 66
What’s the IQ of an AI? ........................................................................................................... 67
Turing test ............................................................................................................................... 67
And so much more .................................................................................................................. 67
What Are Autonomy and Agency? .............................................................................................. 68
Dictionary definitions .............................................................................................................. 68
Intelligent entities ................................................................................................................... 70
Computational entities............................................................................................................ 70
Entities..................................................................................................................................... 71
Actions and operations ........................................................................................................... 71
Tasks, objectives, purposes, and goals.................................................................................... 71
Principals and agents............................................................................................................... 71
Delegation of responsibility and authority ............................................................................. 72
Principal as its own agent........................................................................................................ 72
Agent as principal for subgoals ............................................................................................... 72
Authority ................................................................................................................................. 72
Responsibility, expectation, and obligation ............................................................................ 72
General obligations ................................................................................................................. 73
Ethics ....................................................................................................................................... 73
Liability .................................................................................................................................... 73
Elements of a goal ................................................................................................................... 73
Relationship between principal and agent ............................................................................. 74
Contracts ................................................................................................................................. 74
Capacity for agency ................................................................................................................. 74
Assistants................................................................................................................................. 75
Full autonomy of a principal ................................................................................................... 75
Limited autonomy or partial autonomy of agents .................................................................. 75
Assistants have no autonomy ................................................................................................. 76
Assistants have responsibility but no authority ...................................................................... 76
Control..................................................................................................................................... 76
Robots ..................................................................................................................................... 76
Robots and computers out of control with full autonomy? ................................................... 76
Mission and objectives ............................................................................................................ 77
Mission and operational autonomy ........................................................................................ 77
Independence — mission and operational ............................................................................. 78
Luck and Mark d’Inverno: A Formal Framework for Agency and Autonomy ......................... 78
Motivation ............................................................................................................................... 80
Sociology and philosophy ........................................................................................................ 80
Agent-based modeling (ABM) and agent-based simulation (ABS) ......................................... 81
Definitions ............................................................................................................................... 81
Autonomous systems .............................................................................................................. 83
Lethal autonomous weapons (LAWs) ..................................................................................... 83
Sovereignty.............................................................................................................................. 84
Summary ................................................................................................................................. 84
What Is an Assistant? .................................................................................................................. 85
Definition................................................................................................................................. 85
Specializations of the term...................................................................................................... 86
Virtual assistant — remote or software .................................................................................. 87
Related terms .......................................................................................................................... 88
Principal ................................................................................................................................... 89
Level of expertise and responsibility — simple, specialized, executive .................................. 89
Tasks vs. goals ......................................................................................................................... 89
Primary types of assistant ....................................................................................................... 90
Personal services ..................................................................................................................... 91
Many other tasks..................................................................................................................... 91
Software service — intelligent digital assistant ...................................................................... 92
Synthesized definition ............................................................................................................. 92
Intelligent Entities: Principals, Agents, and Assistants................................................................ 92
What’s the point of an intelligent entity? ............................................................................... 94
How much intelligence is needed? ......................................................................................... 95
Is a dog an intelligent entity? .................................................................................................. 95
Solving bigger problems .......................................................................................................... 96
General meaning of entity ...................................................................................................... 96
Definition of sapient entity ..................................................................................................... 97
Definition of intelligent entity ................................................................................................. 97
How intelligent? ...................................................................................................................... 98
Definition of computational entity.......................................................................................... 98
Dictionary definitions of entity, principal, agent, and assistant ............................................. 98
Definitions of autonomy and agency .................................................................................... 100
More depth on autonomy and agency ................................................................................. 100
Degrees of autonomy ............................................................................................................ 100
Full autonomy for principals ................................................................................................. 101
Science fiction for robot and AI autonomy ........................................................................... 101
Limited autonomy for robots and AI in the real world ......................................................... 102
Dictionary definitions of independent .................................................................................. 102
Dictionary definitions of independence ................................................................................ 103
Freedom of action ................................................................................................................. 103
Definition of independent and independence ...................................................................... 103
Independence and autonomy as synonyms ......................................................................... 104
Dictionary definitions of dependence ................................................................................... 104
Dictionary definitions of dependent ..................................................................................... 105
Definition of dependent, dependence, and dependency ..................................................... 106
Dependence of a principal .................................................................................................... 106
Dependence of an agent ....................................................................................................... 106
Dependence of an assistant .................................................................................................. 106
Dictionary definitions of control ........................................................................................... 107
Definition of control .............................................................................................................. 107
Controlling entities ................................................................................................................ 107
Dictionary definitions of mission........................................................................................... 107
Dictionary definitions of objective ........................................................................................ 108
Dictionary definitions of goal ................................................................................................ 108
Dictionary definitions of task ................................................................................................ 108
Dictionary definitions of delegation...................................................................................... 109
Dictionary definitions of responsibility ................................................................................. 109
Dictionary definitions of contract ......................................................................................... 111
Dictionary definitions of capability ....................................................................................... 111
Dictionary definitions of reputation...................................................................................... 114
Dictionary definitions of requirement .................................................................................. 114
Definition of mission ............................................................................................................. 115
Mission of a principal ............................................................................................................ 115
Mission of an agent ............................................................................................................... 115
Mission of an assistant .......................................................................................................... 115
Definition of objective ........................................................................................................... 115
Definition of delegation ........................................................................................................ 116
Definition of responsibility .................................................................................................... 116
Responsibility of principal or contracting party .................................................................... 116
Responsibility of agent or contracted party.......................................................................... 116
Responsibility of an assistant ................................................................................................ 116
Definition of contract ............................................................................................................ 117
Definition of capabilities ....................................................................................................... 117
Working with capabilities ...................................................................................................... 117
Definition of reputation ........................................................................................................ 117
Definitions of task, purpose, goal, subgoal ........................................................................... 118
Definitions of motivation and intention................................................................................ 118
Definitions of actions and operations ................................................................................... 118
Goals, tasks, and actions ....................................................................................................... 118
Definitions of principal, agent, and assistant ........................................................................ 119
Distinctive roles of principals, agents, and assistants ........................................................... 119
Contract between principal and agent ................................................................................. 119
Definition of requirement ..................................................................................................... 120
Matching requirements with capabilities ............................................................................. 120
Assistant ................................................................................................................................ 120
Technicians ............................................................................................................................ 121
Organizations ........................................................................................................................ 121
Other categories of intelligent entities ................................................................................. 121
Interactions ........................................................................................................................... 122
Relationships ......................................................................................................................... 123
Partners, allies, friends, enemies, adversaries, competitors, and antagonists..................... 123
Connections........................................................................................................................... 124
Legal liability .......................................................................................................................... 124
Ethics ..................................................................................................................................... 125
Morality and ethics of an agent or assistant ......................................................................... 125
Future work ........................................................................................................................... 125
What Is an Intelligent Digital Assistant?.................................................................................... 125
Purpose ................................................................................................................................. 126
Key distinguishing features ................................................................................................... 126
Features................................................................................................................................. 126
How intelligent are they? ...................................................................................................... 127
The Big Four .......................................................................................................................... 128
Connected intelligence, Internet-enabled ............................................................................ 129
Privacy, security, and personal data ..................................................................................... 129
Software and hardware......................................................................................................... 130
Smart speakers ...................................................................................................................... 130
Equivalent terms ................................................................................................................... 130
Related terms ........................................................................................................................ 131
What is the proper term?...................................................................................................... 132
Personal digital assistant ....................................................................................................... 132
Tasks vs. goals ....................................................................................................................... 132
Proactive................................................................................................................................ 133
Online customer service........................................................................................................ 134
Plugin modules for websites and services ............................................................................ 134
Smart cars.............................................................................................................................. 134
Virtual assistant ..................................................................................................................... 135
The one question a digital assistant can’t answer ................................................................ 135
History ................................................................................................................................... 135
Future directions for digital assistants .................................................................................. 135
Human in the loop................................................................................................................. 136
Crowdsourcing ...................................................................................................................... 136
Crowdsourcing questions ...................................................................................................... 136
Crowdsourcing tasks ............................................................................................................. 136
Group crowdsourcing ............................................................................................................ 136
Video ..................................................................................................................................... 137
Therapeutic assistants........................................................................................................... 137
How Much of Robotics Is AI?..................................................................................................... 137
Reader’s choice ..................................................................................................................... 139
Robotic prosthetics ............................................................................................................... 140
How intelligent is your robot?............................................................................................... 141
What is intelligence? ............................................................................................................. 141
Dumb robots ......................................................................................................................... 142
Is it simply automation or is intelligence required? .............................................................. 143
Computer vision .................................................................................................................... 143
Robotic sonar ........................................................................................................................ 143
Specialized hardware ............................................................................................................ 144
Artificial life (A-Life)............................................................................................................... 144
Biologically inspired technology............................................................................................ 145
Criteria ................................................................................................................................... 146
Scoring? ................................................................................................................................. 146
Summary ............................................................................................................................... 147
Vocabulary of Knowledge, Thought, and Reason ..................................................................... 147
Meaning ................................................................................................................................ 148
Truth ...................................................................................................................................... 148
Reality.................................................................................................................................... 148
Popper’s three worlds for reality .......................................................................................... 149
Relation to intelligence ......................................................................................................... 149
Relation to logic..................................................................................................................... 149
Relation to science ................................................................................................................ 149
Relation to epistemology ...................................................................................................... 150
Relation to metaphysics ........................................................................................................ 150
Relation to ethics .................................................................................................................. 150
Relation to communication ................................................................................................... 150
Relation to media .................................................................................................................. 150
Relation to language ............................................................................................................. 151
Relation to knowledge representation and knowledge artifacts ......................................... 151
Artificial intelligence (AI) ....................................................................................................... 151
Domains of truth ................................................................................................................... 151
Entities................................................................................................................................... 151
Short of details ...................................................................................................................... 152
Entity details — metadata..................................................................................................... 154
Sapient entity — people and robots ..................................................................................... 154
Sentient entity — animals and dumb robots ........................................................................ 154
Why not simply quote from the dictionary? ......................................................................... 154
Basic terms for knowledge, thought, and reason ................................................................. 154
Terms related to knowledge, thought, and reason .............................................................. 157
Work in progress ................................................................................................................... 194
Frontier AI: How far are we from artificial “general” intelligence, really? ............................... 194
More AI research, resources and compute than ever to figure out AGI .............................. 195
AI algorithms, old and new ................................................................................................... 198
The fusion of AI and neuroscience ........................................................................................ 200
Conclusion ............................................................................................................................. 201
This informal paper won’t delve so much into specific AI projects or features, but will
endeavor to explore elements of a conceptual framework for judging progress of AI
towards full, human-level intelligence, also known as Strong AI.
So, here we are, still over a decade short of the super-optimist’s forecast for human-
level intelligence.
Are we behind?
First off, all bets are off, or maybe I should say that all bets are on since nobody, even
Kurzweil, has any clue where in that 5–11 year time horizon we really are.
Generally, progress proceeds in fits and starts, with occasional short bursts of
phenomenal breakthroughs, but interspersed with prolonged periods of disappointingly
slow progress.
Superficially, we seem to be in one of those rare periods of rapid advance, but who’s to
say how long it will last.
And who’s to say how long we will have to wait for the next burst.
And who’s to say how many bursts and breakthrough we will need to master before we
finally do break through to true human-level intelligence.
And, I would submit that there is another pair of questions which are more urgent:
1. What fraction of applications which are crying out for AI capabilities have those
needs met with off the shelf AI packages — or at least custom AI which could
be completed within no more than a few months by no more than a few people?
2. What fraction of off the shelf AI capabilities are truly ready for prime-time
deployment in production applications?
I won’t recount specific technical technical accomplishments myself, but I would refer
curious readers to a recent post entitled Frontier AI: How far are we from artificial
“general” intelligence, really? by venture capitalist Matt Turck of First Mark Capital,
which mentions quite a few of the recent accomplishments. His overall conclusion:
• So, how far are we from AGI [Artificial General Intelligence]? This high level
tour shows contradictory trends. On the one hand, the pace of innovation is
dizzying — many of the developments and stories mentioned in this piece
(AlphaZero, new versions of GANs, capsule networks, RCNs breaking
CAPTCHA, Google’s 2nd generation of TPUs, etc.) occurred just in the last 12
months, in fact mostly in the last 6 months. On the other hand, many the AI
research community itself, while actively pursuing AGI, go to great lengths to
emphasize how far we still are — perhaps out of concern that the media hype
around AI may lead to dashed hopes and yet another AI nuclear winter.
As far as a more abstract framework for criteria to judge progress, I’ll defer that for a
future paper, but many of the elements of such a framework will be explored in the
remainder of this paper, and are already discussed to a fair degree in a companion paper,
Untangling the Definitions of Artificial Intelligence, Machine Intelligence, and
Machine Learning.
1. Areas of intelligence.
2. Levels of function. In each area.
3. Degree of competence. At each level of function, in each area.
So, the basic problem we have today is that we can’t even begin to compare any
machine to the pathway of human intellectual development and ask the basic question:
And true general learning — as mastered by even small children — is well beyond
even the most advanced of AI systems today.
• The state of the art for AI today is primarily in task-specific and domain-specific
AI systems.
Granted, there is some preliminary research in unsupervised learning, but that is the rare
exception rather than the general rule, and the whole trust of this paper is on what is
general and common today rather than fringe or atypical, or coming further down the
road.
In short, AI is now quite common but still quite primitive, with rare exceptions.
If you have a problem which you wish to solve using AI, you absolutely cannot just go
out to the store (or Amazon) and order an off the shelf solution, except in a relatively
small number of areas.
Some of the areas where fairly sophisticated AI can be bought or downloaded for free
include:
• Basic natural language transcription. Speech recognition.
• Basic natural language commands.
• Basic but relatively primitive automatic natural language translation. And
detection of language. Google Translate, built into Google Search.
• Particular games. Play chess against the machine, online, for free.
• Intelligent digital assistants. Alexa, Siri, Google.
And when an advance does indeed transcend multiple tasks or multiple domains, it is
usually fairly narrow.
Learning concepts
As of today, there has been no real breakthrough on the ability of a machine to
independently and without human direction discover concepts, especially foundation
concepts.
Even the learning process for children is still well beyond the capabilities of even the
most capable AI systems with the most advanced machine learning algorithms.
Physical bodies and movement in the real world has a lot less to do with human-level
intelligence and more to do with the combination of:
This paper does not get into robotics per se, focusing more strictly on intellectual
capacities.
Is it AI or just automation?
Technically, just about anything a person can do in their head can be considered
intelligence, but do we really want to label relatively simple tasks such as the following
as artificial intelligence:
I think not.
Automation, yes, but rising to the level of higher-order human intelligence, no.
Yes, incredible progress has been made on automating tasks performed by people.
But I would assert that none of that constitutes progress towards Strong AI, automating
higher-order intellectual capacities.
A lot of digital devices, software, and features are being billed as AI (Artificial
Intelligence), but are they really true AI or just traditional automation, much as
computers have been doing for over 50 years (UNIVAC I in 1951) or since the 1890
census with Hollerith punched cards or even the punched cards of the Jacquard loom in
1804. At what point, level, or stage does computing cease being merely automation and
suddenly constitute Artificial Intelligence?
Alas, there is no magic answer, other than to say that it is a purely subjective
characterization to claim that a given form or level of automation is magically Artificial
Intelligence.
Six criteria will be used in this informal paper to try to distinguish mere automation
from more sophisticated AI:
1. Any function that a human or an animal can perform. That’s the starting point,
the initial filter.
2. Distinguish functions that a human can perform that are distinct from functions
that almost any animal can perform.
3. Distinguish animal functions that are not easy, simple, and straightforward to
program on a computer.
4. Distinguish higher-order human intellectual functions that are not easy, simple,
and straightforward to program on a computer.
5. Distinguish difficult human intellectual functions which are merely complex
information processing rather than complex from an intellectual perspective.
6. Distinguish manual tasks performed exclusively by humans which are non-trivial
to automate.
I stick with the focus of that definition on intellectual capacity, but for purposes in this
paper we’ll consider a somewhat broader and looser spectrum of possibilities:
Part of the intent there is to include robots, even those which mimic only animals rather
than limiting the focus to the mimicking of only higher-order human intellectual
capacities.
Just to get a few simpler cases out of the way, the following forms of automation would
probably not be considered AI:
• Washing machine.
• Dish washer.
• Vacuum cleaner.
• Thermostat.
• Garage door opener.
• Textile and clothing power looms capable of being programmed for patterns.
• Numerical control machines (CNC — Computer Numerical Control).
• Player piano. Or any form of playing recorded music, speech, or video.
Granted, any of these could also be supplemented with AI features, but absent some
specific AI features, their main functions don’t suggest AI per se.
A device or appliance such as a Roomba would represent the crossover from mere
automation to the integration of AI due to its navigation abilities — interpreting sensory
perception of the environment.
Those items constitute the menu for both automation and AI. The rest of this paper
focuses on sorting out how those items can or should be categorized as mere automation
in contrast with true artificial intelligence.
Some things that aren’t considered to be part of that list and aren’t really relevant to AI
and automation include some but not all basic biological functions, otherwise
considered part of behavior:
1. Breathing.
2. Eating.
3. Hydration.
4. Excretion.
5. Bone structure.
6. Blood.
7. Secretions.
8. Organs.
9. Cell metabolism.
10. Regulation of metabolic function.
11. Reproduction and mating.
Although the goal of AI overall is not to simulate animals per se, it seems clear that
doing so is an essential precursor to simulating humans, in terms of behavior (except the
exclusions listed above), social behavior, movement, sensory perception, and many
aspects of intelligence.
Granted, simulating animals will not get you even close to simulating distinctly human
intelligence and certainly not higher-order human intellectual functions, but it will
provide a solid and rich foundation as a starting point, especially when considering
robotics.
The items on that main list that are especially human include:
Again, we’re back at the starting point that unless you accept that a computer
performing any task that a human is capable of is implicitly and by definition AI, we
need to consider which tasks or functions may be mere automation rather than AI per
se.
Many of the devices, machines, and software to handle or facilitate the tasks on that list
can be reasonably sophisticated, but commonly won’t be considered AI, although
elements of AI can be included in any of them.
The primary assertion of this informal paper is to focus attention on intellectual activity
as opposed to manual activity, although a lot of the interest, especially for robotics,
includes manual activity, a lot of which is shared with animals.
But what about the smart car feature of self-parking (parallel parking)? How intellectual
is that really? Seems like mostly a manual process, right? But I think most people
would currently consider it an AI feature.
Or a self-driving car. It simply goes from point A to point B with lots of turning,
accelerating, and braking, which are primarily manual operations that hardly require
any significant intellectual effort when a human is driving a normal car. But a self-
driving vehicle would certainly be considered AI, at least today, although much of that
may really be due primarily to its novelty and sophistication rather than any true
intelligence or higher-order intellectual capacity.
Or look at the movement and actions of animals, whether dogs, higher primates,
rodents, small birds, or even miniscule insects. How much intellectual effort do they
need, all without the need of big brains like us humans? Still, building machines to
mimic such movements and actions are commonly accepted as being in the domain of
AI.
Robots are almost uniformly considered AI when so much of what they do, try to do, or
hope to eventually do is hardly more the the behavior of relatively simple animals, even
small rodents and insects.
Somehow, we currently consider movement and navigation in the real world as being
AI, despite the fact that little in the way of higher-order intellect is required.
• Perception
• Attention
• Recognition
• Communication
• Processing
• Memory
• Following rules
• Decision
• Volition or will
• Movement and motor control
• Behavior
Even relatively simple animals (or machines) require many of these mental functions
and mental processes.
Most people would accept or expect that these capabilities should be considered AI,
even if the behavior is not substantially more than that of animals and even insects.
Even then, it is a bit of a stretch to assert that these rudimentary functions and processes
constitute intelligence per se.
Hmmm… can it really be AI if it is not intelligence? Good point. Artificial life (A-Life)
would be a better term for a lot of what is being considered with robots and driverless
vehicles, especially when it comes to robotics and sensory perception, but it’s probably
easier to accept that AI covers a fair bit of A-Life, rather than confuse a lot of people
who have enough trouble understanding artificial intelligence.
Maybe the real point is that although that list of functions and processes isn’t
intelligence alone, they are precursors or requirements in order for a machine to
interact and behave in an intelligent manner. After all, what good is intelligence without
the ability to perceive, communicate, and act in some intelligent manner?
The remaining mental functions and mental processes from the What is AI list are more
clearly higher-order intellectual capacities, which appear to distinguish human
intelligence from animal intelligence:
Granted, designers and engineers may use such intellectual capacities to produce
machines, devices, and software which automate tasks, but it is only when the created
system exhibits such higher-order intellectual capacities that we would feel obligated to
characterize the system as possessing AI.
Finishing off that thought, we will define higher-order intellectual activity as the use of
the higher-order intellectual capacities on that list, those special capacities which
distinguish humans from animals.
Beyond basic manual tasks and automation of basic information processing, more
complex tasks which are clearly automation and clearly involve very sophisticated
processing but just don’t seem to warrant classification as AI per se include:
Granted, specific, niche AI features can find application within any of those automated
tasks, but the tasks themselves don’t seem to provide evidence that the computer is
engaging in higher-order intellectual activity as would be expected for an AI system.
There are plenty of borderline tasks which aren’t necessarily AI per se, but are more
than mere automation of simple or straightforward tasks, including:
And then there are borderline tasks which are now commonly accepted as automation
but at least have roots in AI, including:
And finally we have tasks which are not so dissimilar from some of these mere
automation tasks but are commonly accepted as AI, at least at the present moment:
Heuristics have the allure that they appear or seem to be comparable to intelligence, but
they have limitations that a true, human-level intelligence would not.
They constitute a gray area between mere automation and true intelligence.
Robotics
Much of robotics revolves around sensors and mechanical motions in the real world,
seeming to have very little to do with any intellectual activity per se, so one could
question how much of robotics is really AI.
Alternatively, one could say that sensors, movement, and activity enable acting on
intellectual interests and intentions, thus meriting coverage under the same umbrella as
AI.
In addition, it can be pointed out that a lot of fine motor control requires a distinct level
of processing that is more characteristic of intelligence than mere rote mechanical
movement.
In summary, the reader has a choice as to how much of robotics to include under the
umbrella of AI:
Conclusion
Personally, I prefer the first three choices, including robots which mimic animals, but I
also accept that some may prefer the other choices, as stated or with additional nuances.
So, I personally would not consider most traditional software or even a lot of modern or
even smart software unless it exhibits higher-order intellectual capacities.
I accept robots and driverless vehicles as being AI. Although we should start calling
them artificial life (A-Life.)
Intelligent digital assistants are in a gray area. They are generally more heuristic or use
specialized, niche AI features rather than being broad AI systems capable of a wide
range of higher-order intellectual activity. They’re more about automation than higher-
order intelligence. That said, I’ll accept that they fall under the AI rubric, at least for
now.
Of course, what current AI systems evolve into in the coming years and decades is
another matter.
Ten to twenty years from now people will look back and probably call current AI
systems mere toys, and laugh that we considered them to be intelligent.
But thirty years from now people won’t be laughing anymore about our current
technology at all. That’s because (at least according to Ray Kurzweil’s Singularity) the
robots will have taken over. Instead, it will be the robots laughing that we considered
ourselves to be intelligent.
But I’ll try to limit my analysis and speculation to the present and near future.
But merely getting comparable results to a human for a given task does not necessarily
mean that the machine has human-level intelligence.
Besides, a lot of tasks performed by people are relatively simple in the first place, so
that they aren’t necessarily tapping into the core of the higher-order intellectual
capacities which may be present but not necessarily used.
Heuristics and mental shortcuts are highly valued and to be applauded, but they most
certainly are not the same as higher-order intellectual capacities.
Progress on gaming
Some of the highest profile and most impressive advances in AI have been on the
gaming front, including:
• Chess.
• Go.
• Learning classic video games.
1. These are niches, or what I call towers of intelligence. Excellence in one of these
areas implies nothing about general intelligence in disparate areas.
2. Humans must preprogram basic knowledge and basic logic, such as ground
rules. These AI systems are not strictly learning from a completely blank tabula
rasa.
3. Heuristics and statistics rather than true, higher-order human-level intellectual
capacities are being exploited.
In short such advances are significant progress in machine intelligence, but not artificial
higher-order human intelligence per se.
Is it AI or just machine intelligence
Although some (many) people treat machine intelligence and artificial intelligence as
synonyms, I would strongly advise treating the terms as distinct.
We can and should strongly applaud advances in machine intelligence without any need
or obligation to assert that such advances necessarily constitute advances in artificial
intelligence of the higher-order human-level intellectual capacities kind.
Artificial Intelligence (AI) and machine intelligence are commonly used synonymously,
but there is a nuanced difference — AI is more properly intended to focus on simulating
the features of the human mind, while machine intelligence can also include intelligent
information processing that is distinct from or beyond the capabilities of the human
mind. Machine intelligence is also sometimes used to refer to machine learning (ML), a
subset of machine intelligence — and AI as well, depending on how you interpret the
terms AI and machine intelligence. This informal paper will endeavor to distinguish AI
and machine intelligence.
Again, learning is properly part of intelligence but worth calling out since learning
commonly occurs distinctly from applying learned knowledge. As discussed later,
learning occurs on an ongoing basis in daily life — so-called little learning — while
more complex learning of concepts, principles, and methods — so-called big learning
— tends to require a more specialized and more dedicated effort. The main reason for
calling out learning here is since we have this specialized field of AI and machine
intelligence called machine learning.
There is no requirement that machines directly parallel or directly mimic people, but
that is a common interest and noteworthy endeavor. It is worth calling out the
distinction:
A purist would argue that AI should only be used when the intelligence is directly
comparable to human intelligence.
A different class of purist could assert that everything a machine does is artificial, so
that any purported machine intelligence is by definition artificial intelligence.
Another purist could also assert that any intelligence of a machine is by definition
machine intelligence, so that even AI which directly mimics human intelligence should
be called machine intelligence even if another purist calls it machine intelligence.
The reader is free to take their pick. My job is simply to highlight the issues and the
distinctions.
1. Human physical activity. Things people do. Mostly manual tasks, guided by
intelligence.
2. Human intelligence. Human mental activity. Things people do in their minds.
Thinking, reasoning, calculating, planning, imagining, learning, communicating,
and guiding physical activity.
3. Human learning. Little learning, such as names and faces of people we meet,
places we go, things we observe, and activities we engage in, as well as big
learning, such as school and other forms of education, studying, reading, and
research to learn and develop concepts, principles, and methods.
4. Robotics. Physical activity of machines.
5. Artificial intelligence. Machines simulating human intelligence.
6. Machine intelligence. May be artificial intelligence, a subset of artificial or
human intelligence, or forms of reasoning, calculating, planning, imagining,
learning, communicating, guiding machine activity, which are distinct from the
mental activities of people.
7. Machine learning. Specialized forms of pattern and rule recognition that either
mimic analogous human recognition, or are distinct from human recognition.
Generally, a very limited subset or minimal approximation of human
recognition, or fairly distinct from human recognition. More advanced forms can
develop concepts, principles, and methods, but that is generally beyond the
abilities of most current machines.
The remainder of this paper will focus on specific aspects of machine intelligence and
machine learning. The final summary offers some questions that will help the reader
decide whether any given piece of technology is better classified as AI or machine
intelligence — or machine learning.
And for a more brief, light introduction to AI, see What Is AI (Artificial Intelligence)?
The label AI makes most sense when considering automated tasks which people (or
animals) have traditionally performed.
The label machine intelligence makes most sense when considering tasks which are not
traditionally associated with human (or animal) mental or physical activity, including:
For the purposes of clarification, I would distinguish what I call big learning from what
I call little learning.
The former (big learning) is the difficult form of learning where complex concepts,
principles, and methods are learned with great effort. This is the world of school and
other forms of education, studying, tests, research, and struggling. It is usually a distinct
activity from daily life.
The latter (little learning) is the learning we do in everyday daily life, like meeting new
people, learning their names and interests, and remembering their faces. Or facts we
learn as we travel and engage in activities in unfamiliar places. Not as much effort or as
much of a struggle as big learning. No new concepts, principles, or methods involved.
In between we have a hybrid, focused more on categorical distinctions and nuances. The
world of patterns and rules.
Training
A lot of what passes for machine learning is focused on a combination of the latter three
subsets of learning — little learning, hybrid of little and big learning, and training.
Sometimes the machine can learn completely on its own, but very commonly some
degree of human intervention and so-called training is required.
Then, the machine can begin to discover patterns and infer rules for that focused
domain.
The machine may indeed succeed at ferreting out what is happening (what the patterns
and rules are), but not why the activity is happening or what the objective of the activity
is. Very limited learning.
So, big learning focuses on deep concepts, broad principles, complex methods, and the
bigger picture context, while little learning is much more superficial, either being
relatively trivial or requiring hard-wired a priori knowledge (including human
genetically-encoded abilities or hard-coded knowledge for a machine) or a limited
ability to discover patterns and rules.
Although much of the emphasis of machine learning today is on relatively basic patterns
and rules rather than the concepts, principles, and methods of human learning, it is
actually quite possible that machines can or could perform some forms of learning even
better than many humans, such as:
As usual, a hybrid is possible, so that the best of both worlds, man and machine, can be
combined in a synergistic manner.
Automation
Adding a column of numbers or sorting a list of names does require some intelligence,
but we normally don’t consider arithmetic and basic information processing to be AI or
even machine intelligence.
Some forms of automation that are in a gray area where it’s a fielder’s choice whether to
consider it just automation or true machine intelligence include:
1. Optimization.
2. Scheduling.
3. Data analytics.
4. Business intelligence.
5. Simulation.
6. Scientific and engineering calculation and modeling.
Patterns
For the purposes of this paper, AI would relate to patterns that a typical person could
recognize, like names, faces, objects, and concepts, while machine intelligence would
related to patterns more easily recognized by machines than people, as well as the
simpler forms of patterns that even an average person can recognize quite quickly.
Right now, people are generally pleasantly surprised whenever machines are able to
recognize even relatively simple patterns.
Granted, people are amazed by specialized cases such as DNA and fingerprint
matching, facial recognition, voice recognition, chess, Go, ping pong, and Jeopardy.
But, that’s the point here, that each of those is very specialized and requires careful and
complex hard-wiring of basic knowledge.
It may be another few years or more before people begin to be wowed and blown away
by machines recognizing concepts and principles.
Computer vision
In truth, most people would not consider visual recognition of objects, scenes, and
qualities to require much in the way of intelligence — even small children can do it. Or
even animals, for that matter.
Personally, I’d consider computer vision more associated with machine intelligence
than AI.
Generally, perception and recognition are inputs to intelligence rather than exemplifying
intelligence itself — reasoning, working with concepts and principles, planning, guiding
activity, etc.
For sure, there are plenty of applications for human-like vision, whether for robots in
the home or driverless vehicles, but there are whole other broad categories for computer
vision that are more distinct from or beyond the vision that people generally possess. I
would classify the latter under machine intelligence, although technically some would
insist that this is still AI.
Again, technically, this could still be considered AI, but you gain no conceptual benefit
from labelling it AI rather than what it really is — machine intelligence.
Still, if the processing directly parallels human intellectual activity, then by all means
label it as AI as well.
Cybersecurity
Although skilled individuals can do a fair amount of interesting work in the area of
cybersecurity, automation is desperately needed. Some of that may be relatively simple
automation, but a fair amount could easily be classified as AI.
But at some stage, the types of data, its volume, its complexity, and the relationships
within and between the data begins to take on a character that is distinctly beyond what
even a highly-motivated technical specialist could muster. Enter the world of machine
intelligence.
Again, technically, it can still be considered AI, but it begins to look so different from
what any normal person would do that it just makes a lot more sense to label it what it
really is — machine intelligence.
That said, I wouldn’t want to go so far as to label all forms of automated cybersecurity
as machine intelligence. I’d prefer to reserve the term for cases such as discovery of
new patterns of data, new patterns of activity and behavior, and new threats, rather than
detecting instances of data, activity, behavior, and threats which are already known and
manually hard-coded by skilled human operators.
Generally, processing in cybersecurity will fall into one of four general categories.
The point here for machine intelligence is to emphasize the brain of the computer, not
just its brawn.
As usual, a hybrid is possible and welcome as well — man and machine, each
contributing their own strengths in intelligence.
Data analytics and business intelligence
Again, we don’t want to merely label all forms of automation as machine intelligence,
but there are indeed more advanced forms of data analytics and business intelligence
which contribute more than just raw number crunching and basic statistical processing.
But any time that the machine can discover new, previously unknown patterns, the label
of machine intelligence can become warranted.
That said, if the machine capabilities are simply aiding a human user in their own
detection of patterns, I’d be more reluctant to trot out the label of machine intelligence.
And if the types of patterns recognized look very little like the patterns that a human
would normally recognize, I’d be more cautious about labeling it AI.
Again, I lean towards labeling an automated activity as AI when it has a fairly direct
analog to human mental activity.
But if the calculations endeavor to automatically recognize new patterns, rules, concepts
and principles, or any other activity normally associated with higher-order intellectual
activity, then AI and machine intelligence might become appropriate labels. That is not
commonly the case.
Usually it is up to the researcher to use their own mind to recognize patterns, rules,
concepts, and principles from the data output from the calculations or modelling
process.
There are a number of applications, domains, and forms of data which have a fairly
dramatic level of complexity, such as:
1. Graphs.
2. Networks.
3. Time series.
4. Complex database joins.
5. Relationships.
6. Complex relationships.
The mere complexity would not automatically confer the label of machine intelligence,
but to the degree that the algorithms working with such data are discovering new
patterns, rules, or even concepts, principles, and methods, it becomes more appropriate
to label such processing as machine intelligence.
Although merely labeling such processing machine intelligence can also bring it under
the larger umbrella of AI, I’d opt to stick with labelling it machine intelligence rather
than labelling it AI unless there are distinctly human patterns, rules, concepts, and
principles involved.
If the machine is doing something that a human would traditionally have done (before
computers were invented), then it could well be considered AI proper.
Summary
The questions one should ask when considering whether a technology is AI or machine
intelligence are:
When in doubt, unless it’s basic automation, it’s probably AI unless it’s processing data
that no mere mortal would relate to.
This is not unlike the simple fact that even for jobs staffed by people, not all positions
require a genius or even more than a very modest fraction of what the staff are really
capable of. Like, Einstein working as a patent clerk.
1. Area of intelligence.
2. Level of intelligence. In a given area of intelligence.
3. Degree of competence. For a given level in a given area of intelligence.
Not every application requires all areas of human-level intelligence. Not every app
needs to be a chess grandmaster. Not every app requires facility with quantum
mechanics.
Even in a given area of intelligence, not every app requires all levels of function in that
area. An auto mechanic doesn’t need to be able to design a new engine. A roadside
assistance technician doesn’t need to be able to tear down and rebuild an engine.
Even for a given level of function in a given area of intelligence, a given app doesn’t
require the maximum level of competence. Basic competence may be quite sufficient
and readily achievable, while expert or genius level competence may be expensive,
difficult, or even impractical.
Areas of intelligence
1. Natural language.
2. Spoken word.
3. Written word.
4. Gestures. Hand, finger, arm.
5. Facial expressions. Smile, frown.
6. Nonlinguistic vocal expression. Grunts, sighs, giggles, laughter.
7. Body language.
8. Images.
9. Music.
10. Art.
11. Movement.
12. Creation and consumption of knowledge artifacts — letters, notes, books,
stories, movies, music, art.
13. Ability to engage in discourse. Discussion, conversation, inquiry, teaching,
learning, persuasion, negotiation.
14. Discerning and conveying meaning, both superficial and deep.
1. Objects
2. Faces
3. Scenes
4. Places
5. Names
6. Voices
7. Activities
The measure of progress in AI in the coming years will be the pace at which additional
areas from those lists are ticked off, as well as improvements in the level of competence
in the levels of function in each area.
Progress in AI will likely continue to be uneven, with both strength and weakness in
distinct areas, levels of functions, and degrees of competence.
Areas of AI research
1. Reasoning.
2. Knowledge and knowledge representation.
3. Optimization, planning, and scheduling.
4. Learning.
5. Natural language processing (NLP).
6. Speech recognition and generation.
7. Automatic language translation.
8. Information extraction.
9. Image recognition.
10. Computer vision.
11. Moving and manipulating objects.
12. Robotics.
13. Driverless and autonomous vehicles.
14. General intelligence.
15. Expert systems.
16. Machine learning.
17. Pattern recognition.
18. Theorem proving.
19. Fuzzy systems.
20. Neural networks.
21. Evolutionary computation.
22. Intelligent agents.
23. Intelligent interfaces.
24. Distributed AI.
25. Data mining.
26. Games (chess, Go, Jeopardy).
For more depth in these areas, see Untangling the Definitions of Artificial Intelligence,
Machine Intelligence, and Machine Learning.
Levels of function
In a given area of intelligence, we can also discern levels of function — what is the AI
system actually accomplishing, relative to what a human might be able to accomplish.
Here are a list of generalized, abstract, but informal levels of function that can be
applied to any area of intelligence:
At a more detailed level, the mental functions and mental processes of intelligence or
intellectual capacity include:
That’s a very large amount of intellectual capacity which an AI system will need to
possess to be truly classified as Strong AI.
But as this paper suggests, we can partition intelligence into distinct areas, distinct
levels of function in each area, and distinct degrees of competence for each level of
function in each area.
Degree of competence
A given AI system will have some degree of competence for some level of function in
some area of intelligence. That AI system may have differing or even nonexistent
competence for other levels of function or in other areas of intelligence.
Levels of competence for current AI systems are all over the map.
In some towers of intelligence current AI systems may indeed be at the expert, elite
expert, or even super-expert level.
But as a general proposition, current AI systems tend to be in the minimal to rich subset
range of competence for most functions.
One of the brighter spots of AI in recent years has been the widespread use of fairly
competent natural language processing (NLP).
I expect these blemishes to be overcome without too much difficulty, but it may still be
another two to five years before natural language processing becomes a slam dunk and
second nature for computing in general. For now, it remains more of a special feature
rather than a presumed general feature.
Autonomy, principals, agents, and assistants
Autonomy is a key requirement for true strong AI. this means that the AI system would
be able to set its own goals, not merely do the bidding of a human master.
1. Principals. Full autonomy. Entity can set its own goals without approval or
control from any other entity.
2. Agents. Limited autonomy. Goals are set by another entity, a principal. Entity
has enough autonomy to organize its own time and resources to pursue tasks
needed to achieve the goals which it has been given.
3. Assistants. No significant autonomy to speak of. Unable to set its own goals. In
fact an assistant is given specific, relatively narrow tasks to perform, with no
real latitude as to how to complete each task.
A strong AI system would have full autonomy. It would be able to act as a principal.
Note that a driverless car would be an agent. It cannot decide for itself where to go, but
given a destination, it is free to choose the route to get there.
Technically, we shouldn’t expect full autonomy for AI systems in the foreseeable future.
That would mean citizen robots which control their own destiny rather than merely
doing our bidding, which would have limited them to being agents. Think Hal in 2001
or Skynet in Terminator.
For more on autonomy, principals, agents, and assistants see these two papers:
Intelligent agents
Although agents by definition do not have full autonomy, it is very helpful if they have
a significant degree of autonomy so that the user can request that a goal be pursued
without needing to expend any significant energy detailing how the agent should
achieve that goal.
Intelligent agents don’t really exist today. Rather, we are seeing significant activity and
progress with intelligent assistants, but these AI systems are focused on narrow,
specified tasks rather than broader goals with much less latitude as to how to achieve
them.
Although we don’t have much in the way of principals and agents (besides driverless
cars), we are seeing significant activity and progress with intelligent assistants or
intelligent digital assistants, such as Alexa and Siri, able to complete relatively simple
tasks requested in spoken natural language.
For more on intelligent digital assistants, see What Is an Intelligent Digital Assistant?
Are massive waves of workers about to lose their jobs due to automation and intelligent
robots.
Yes, robots are getting incrementally more sophisticated as every year goes by, but they
are still quite primitive.
We are definitely seeing significant progress in machine intelligence, but not so much
progress yet on higher-order human-level intellectual capacities.
Yes, incrementally, small numbers of workers will be displaced by robots, but nothing
major anytime soon.
Ongoing innovation will tend to create new forms of employment as quickly as older
jobs are eliminated. Granted, training and relocation may be required, but that’s the
world we live in.
Although robots are nowhere close to being capable of replacing large numbers of
workers, at some point in the more distant future that may indeed be the case.
Besides, an average worker doesn’t really utilize more than a tiny fraction of their
intelligence for the tasks they are commonly assigned.
Not very.
But, still, probably significantly more intelligent than the current crop of robots.
Besides, even if 90% to 99% of the tasks performed by an average worker require
minimal intelligence, that other 1% to 10% of their work may require significantly more
intelligence, such as:
Granted, even many of those areas are also significant opportunities for automation, but
progress will continue to be inconsistent, undependable, and problematic.
Incrementally, more and more workers will have their full jobs automated, but that
gradual process will likely be compatible with the incremental appearance of new forms
of work.
Even if it is indeed theoretically possible to innovate all existing jobs away, there will
always be practical reasons that this process will not occur rapidly.
And to the extent to which average workers are only using a small fraction of their
intellectual capacity, they have excellent potential to be trained for new jobs.
The appearance of the personal computer (PC) was a major revolution. Ditto for the cell
phone and smartphone. We haven’t seen such a revolution for personal AI yet, in the
sense of strong AI. Weak AI, yes; strong AI no.
Okay, we have Alexa and Siri and other intelligent digital assistants, but they are too
minimal and too specialized to constitute a true revolution to what I would call personal
AI.
Alexa and Siri remind me more of the very early personal computers, such as the Altair,
Commodore, and Atari computers, which were truly mere toys and more suitable for
playing games and having fun than any serious computing. Even the Apple II was in
that category.
It wasn’t until the advent of the IBM PC and the Apple Macintosh that personal
computers could finally be counted on to do serious work.
That’s the kind of transition we are still waiting for for AI. From toylike, simple, single-
task features to broader and richer goal-oriented activities.
More significantly, we need AI systems that can figure out our needs and automatically
address them without requiring us to explicitly and carefully detail individual tasks.
AI systems and features currently provide plenty of automation, but are not yet offering
any significant higher-order human-level intellectual capacities.
Another benefit of the personal computer was that they offered a fairly rich set of
features right out of the box without requiring an expensive connection to an external
service. Alexa and Siri are interesting, but most of their function is accomplished in the
networked cloud rather than locally.
The four main qualities which I am looking for for personal AI, which is a breakthrough
comparable to the personal computer (IBM PC and Apple Macintosh) are:
1. Fairly rich set of features. 10X to 100X what Alexa provides. Covers a much
broader swath of the average person’s daily life.
2. Very easy for average user to set up, configure, control, monitor, and
understand. No degree in rocket science required. No technical sophistication
required.
3. No network connection required. Yes, a network connection may provide
additional features and power, but would not be required. Or at least not always
be required.
4. Based on open source software. Users should not be held hostage by vendors
and should be able to view and even enhance their systems. Even if a user
doesn’t wish to do this themselves, they can at least take advantage of the work
of other users who are willing and capable of taking advantage of open source
capabilities.
AI systems are currently more oriented towards high-end applications than consumer-
oriented features except for more basic features.
Sure, web sites may use lots of AI under the hood or have AI chat bots for automated
customer service, but none of this is a direct benefit for the consumer.
Generally, AI is not yet ready for consumers for any higher-order human-level
intellectual capacities.
Only limited, narrow niches seem particularly ripe for consumer AI, such as:
1. Simple robotic animals. May be fun, amusing, and interesting, but offer little in
the way of intellectual capacity.
2. Task-specific AI or domain-specific features.
3. Broader but shallow features, such as intelligent digital assistants.
4. Photo manipulation and management.
5. Gaming.
6. Driving automation.
Driving automation has significant impact, but is generally weak to moderate AI, well
short of strong AI, with relatively narrow, specific functions such as:
Emotional intelligence
One important area where most AI systems are severely lacking is emotional
intelligence.
The mere fact that emotional intelligence does not even get mentioned for most AI
systems focuses attention on the fact that these systems are not focused on strong AI.
That will be a key indicator of true progress towards strong AI — that emotional
intelligence begins to play a more significant and even essential role.
1. Data
2. Information
3. Knowledge
4. Wisdom
The first two levels are handled very adequately by current digital computing.
Knowledge is a mixed bag, with some fairly decent advances, but still some gaps. Facts
are a slam-dunk for the most part. Know-how can still be problematic, depending more
on hardcoded preprogramming more than true machine learning of concepts and
human-level meaning.
But wisdom is a whole other category, virtually untouched by current AI systems. And
considered essential for a mature human being.
Beyond basic facts and practical know-how, wisdom includes principles and values,
both more abstract than concrete. And the ability to apply abstract principles and values
in other domains where concrete knowledge, facts, and know-how may be minimal.
A companion paper has a proposed list of core principles of general wisdom for any AI
system which wishes to qualify as Strong AI:
• Proposal for Level 0 General Wisdom
We may well be more than a few years away from AI systems that exhibit human-level
wisdom, or even a small fraction of human-level wisdom.
Extreme AI
For more on extreme AI, see the companion paper Extreme AI: Closing the Loop and
Opening the Spiral, as well as the larger paper, Untangling the Definitions of Artificial
Intelligence, Machine Intelligence, and Machine Learning.
This paper is not intended to delve into ethical and legal issues of technology, focusing
primarily on the technology itself. These matters are covered a little in the companion
paper Untangling the Definitions of Artificial Intelligence, Machine Intelligence, and
Machine Learning.
Other than to simply say that yes, there will be lots of ethical issues that will arise as we
begin to dawn on the age of strong AI, but we’re not even close yet, so not to worry.
There will be legal liability issues as well, but many of them already apply to existing
software systems.
Lethal autonomous weapons (LAWs) will present interesting ethical and legal
challenges, including international law and the law of war.
Fully autonomous or near-fully autonomous AI systems will present serious ethical and
legal challenges as well.
I don’t want to downplay these matters, but they are somewhat beyond the scope of this
paper. They warrant a separate paper or papers, but the unfortunate fact is that we won’t
be able to address ethical and legal issues in any meaningful depth until we understand
much better the capabilities of such systems, which we don’t.
Attempting to solve a problem which doesn’t yet exist is always problematic. Yes, we
can and should anticipate potential problems, but over-anticipating could cause more
problems than it might solve.
There is no question that a wide variety and range of dramatic breakthroughs are needed
to achieve true strong AI.
1. Unpredictable pace.
2. Unpredictable timing.
3. Unpredictable impact.
The best and most valuable breakthroughs tend to come out of nowhere, when you least
expect them.
1. Manhattan-style projects.
2. Moonshot programs.
Those approaches do work, but only on rare occasion, when there is a critical mass and
all of the fundamental elements are essentially in place.
Yes, significant research is still needed, and that means money, people, and priority, but
the emphasis should always be on patient effort, not some insane belief that if we throw
enough money at the problem a magical solution will appear overnight.
That said, when breakthroughs do eventually come, they always come fast and
furiously.
But when?
But with AI as it is today, far too few of the fundamental elements are in place, or
anything close to it.
Fundamental computing model
All of our great progress in digital computing has been based on the power of Turing
machines, but there is no great clarity as to whether a Turing machine has sufficient
conceptual power to simulate the human brain and mind.
Turing himself hypothesized what he called a B-type u-machine which could rival the
computational power of the human mind. The operations of his unorganized machine
(u-machine) more closely parallel the neurons of the human brain. A B-type machine
has the ability to dynamically reconfigure the connections between those computing
elements.
Whether today’s neural network computing models have sufficient power to simulate
the human brain is unclear, especially as many of them are simulated on Turing
machines. There have been efforts to do specialized hardware for AI and neural
networks in particular, but there is no clarity about the effectiveness or limitations of
such efforts. They may be faster than a Turing machine, but the essential question is
whether they can compute anything different than a Turing machine. Certainly there is
great hope, but hope itself is not a solution.
Some researchers believe there is a lot more going on within neurons and their
connections that we do not yet fully fathom.
In the end, the fundamental computing model will matter, but we are not there yet, here
in April 2018.
When we do achieve conceptual breakthroughs, how long will it take to get them from
the lab to the hands of consumers?
Unfortunately, the answer is that the time to market is unknown, unknowable, and
highly variable.
There is always the chance that some guy in his garage might engineer a breakthrough
that can be taken directly to market, but that is more of a fluke than something we
should depend upon. And usually the guy in his garage is building upon a lot of
foundational work which was done previously for high-end applications.
1. Standardized IQ tests.
2. The Turing test.
Human intelligence has traditionally been measured using standardized IQ tests. People
commonly recognize that an IQ of 140 is a genius and 160 is a super-genius.
Granted, there are a variety of disputes over standardized IQ tests, but they are the gold
standard.
But today’s AI systems could not even take an IQ test, let alone score high. That alone
says something about where we are relative to strong AI.
But, that exemplifies the state of affairs for AI today, namely that an AI system can be
fairly readily preprogrammed for a specific, relatively narrow application, but that does
not mean that such an AI system is capable of general intelligence applicable to a wide
range of problems.
There have been some proposals for specialized IQ tests designed specifically for AI
systems, but that only proves the point, that machine intelligence is rather distinct from
higher-order human-level intelligence.
The so-called Turing test, which Turing himself referred to as The Imitation Game, is
more of a metaphor for arriving at a true/false answer as to whether a human or a
machine is on the other end of a communication link (in another room from the
observer.)
The basic concept of the Turing test is to construct questions such that the answer will
provide a clue as to whether the responding entity is intelligent or not. So, ask a bunch
of questions, evaluate the responses, and decide whether the entity might be a person or
likely merely a machine which is imitating a person.
I haven’t heard of any AI system that can credibly pass, consistently. There have been
some claims of success, but there have also been rebuttals to such claims. So, it remains
a matter of dispute.
And even if a claim to pass a given Turing test were to hold up, it merely means that the
entity was intelligent for that particular test, with no guarantee that the entity would
respond intelligently for other tasks.
Again, there is the risk that the AI system might be preprogrammed to pass the test
rather than truly intelligent and able to learn on its own.
In short, the traditional Turing test is too weak and vague to be a technically robust test
of true, higher-order human-level intelligence.
CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans
Apart) is a variation on Turing’s original test which focuses more on visual perception
rather than being an intellectual challenge. The test presents a small amount of arbitrary
text that is artificially distorted to make it very difficult to apply traditional optical
character recognition techniques. It does indeed work fairly well, although it has some
limits and there have been a number of recent efforts that do seem to be able to defeat or
pass CAPTCHA tests automatically, or by diverting the test query to a pool of users
who earn a small fee for correctly responding to the challenge.
But just because an algorithm does indeed defeat particular CAPTCHA tests does not
mean that it can defeat all CAPTCHA tests.
Worse, CAPTCHA was never really a test of intelligence in the sense of higher-order
intellectual capacities. At best, you could claim a tower of intelligence which works
well for a narrow range of tasks but has no real applications to a wide range of tasks.
Worse, the CAPTCHA system doesn’t even comprehend the human-level meaning of
what it is doing, whereas a true human-level intelligence would in fact fathom the
nature of its tasks.
There has also been some work in training AI systems to pass various college-level
tests.
Once again, that is impressive as a heuristic or machine intelligence, but the AI system
doesn’t comprehend the meaning and concepts behind the subject matter being tested.
The AI system may well be able to pass the test, but wouldn’t necessarily be able to
succeed at applying the subject matter for solving real-world problems.
A better test would be to provide the AI system with only a raw PDF of the textbook
and any related materials, and then take the test. Or even better, answer detailed
questions on the subject from a sophisticated user.
Even that would not assure that the AI system actually comprehended the full depth of
the meaning and concepts of the subject matter.
Maybe that’s the indicator of our status with AI, that current AI systems mostly depend
on preprogrammed knowledge and limited, domain-specific learning so that they are not
truly facile with the concepts and their true meaning.
In short, we are not yet in a position to have reliable tests to evaluate the intelligence of
an AI system.
I haven’t worked out any precise, numerical scoring system for AI progress towards
strong AI, but there are some possibilities.
1. A score for each degree of competence in each level of function for each area of
intelligence. This would be the finest grain of scoring. Very specific.
2. An overall score for each level of function for each area of intelligence.
3. An overall score for each area of intelligence.
4. An overall score across all areas of intelligence. The total score. The IQ of the
AI system.
How to score AI systems which are towers of intelligence, focus on particular tasks or
domains, or are deficient in some areas while excelling in others is problematic. Having
separate scores for each area would make it easier to tell what the real story was.
Links to my AI papers
Ten years from now would be 2028, a year short of Kurzweil’s target of 2029.
So, I’ll say ten to fifteen years, or 2028 to 2033. That’s not grossly out of step with
Kurzweil’s target.
Written by
Jack Krupansky
Freelance Consultant
Jack Krupansky
Jan 5, 2018
First, the obvious. AI means that the following elements are involved:
1. A machine.
2. A computer.
3. Computer software.
4. Some degree of intelligence that is suggestive of the intelligence of a human.
What about robotics, so much of which is merely mechanical and seemingly unrelated
to any intellectual activity — is it really AI per se? There is a section on Robotics later
in this paper to explore this question a little deeper. The short answer is that it’s a
fielder’s choice how much of robotics should be considered AI. If it enables or supports
intellectual activity or the carrying out of intellectual intentions, then it’s fair to be
considered under the rubric of AI.
The operative word there is suggestive — meaning that AI doesn’t require achieving the
full range of human cognitive and behavioral capabilities, but merely enough of a
fraction of the full range that at least hints at or gives the appearance of human-level
intelligence.
You can also read articles about superintelligence, far beyond even human intelligence.
But that’s more the realm of science fiction and speculation at this stage.
In fact, even strong AI remains far beyond our technological reach at this stage.
Current intelligent digital assistants have achieved a minimal level of Moderate AI,
but they still fall far short of Strong AI.
Many current consumer and industrial products have some level of Weak AI,
occasionally bordering on minimal Moderate AI. It is common now to use the adjectives
intelligent or smart to indicate the presence of Weak AI in a product, system, service, or
feature, such as:
Again, these systems and devices exhibit some fraction of a human-level function, but
usually only in some relatively modest sense. And certainly nothing approaching
human-level Strong AI.
My longer paper also discusses levels of competence or how robust and capable a given
implementation is in any particular area of function, relative to a fully-functional
human. I call this competent AI. Levels of automation competence range from:
That’s it, the starting point for an understanding of AI. Continue reading if you need a
little more depth.
My longer paper also discusses the spectrum of functional behavior, to categorize how
functional a system is overall. The point of this model is that:
What is intelligence?
Unfortunately there is no concise, crisp, and definitive definition for intelligence,
especially at the human level. But a number of elements of intelligence are readily
identified.
When we refer to human intelligence we are referring to the intellectual capacity of a
human being.
See my longer AI paper for a lot more depth, but at a superficial level intelligence
includes a significant variety of mental functions and mental processes:
Communication includes:
1. Natural language.
2. Spoken word.
3. Written word.
4. Gestures. Hand, finger, arm.
5. Facial expressions. Smile, frown.
6. Nonlinguistic vocal expression. Grunts, sighs, giggles, laughter.
7. Body language.
8. Images.
9. Music.
10. Art.
11. Movement.
12. Creation and consumption of knowledge artifacts — letters, notes, books,
stories, movies, music, art.
13. Ability to engage in discourse. Discussion, conversation, inquiry, teaching,
learning, persuasion, negotiation.
14. Discerning and conveying meaning, both superficial and deep.
Recognition includes:
1. Objects
2. Faces
3. Scenes
4. Places
5. Names
6. Voices
7. Activities
8. Identities
9. Intentions
10. Meaning
Only a Strong AI system would possess all or most of these characteristics. A Weak or
Moderate AI system may only possess a few or a relatively narrow subset.
The measure of progress in AI in the coming years will be the pace at which additional
elements from those lists are ticked off, as well as improvements in the level of
competence in these areas of function.
When capabilities seem beyond what a computer can easily do, it is easy to ascribe it to
being a matter of intelligence. As if the tasks we have already automated didn’t require
intelligence.
Once we do manage to figure out how to automate some seemingly difficult task, we
assert that this is artificial intelligence. Or at least until it becomes widely accepted that
computers can obviously do a particular task and do it quite well. Only then will we
gradually and quietly cease using the AI label for those tasks that we no longer have a
need to refer to explicitly.
Maybe the issue is that since we have already automated so much of the low-hanging
fruit that we are finally bumping into the knee of the difficulty curve so that it takes
increasingly intense levels of effort and resources to advance up the intelligence
spectrum, so that each advance comes more slowly and therefore seems so much more
spectacular.
Robots, driverless cars, and even intelligent digital assistants certainly seem spectacular
right now, but once they get all the wrinkles worked out and they become common and
mundane rather than rare and special, the urge to label them AI will quickly fade.
Anti-locking brakes, optical character recognition, spelling and grammar checkers and
correctors, and auto-focusing cameras were once quite unusual and exceptional and
hence noteworthy as AI, but these days they are assumed and not so notable common
features, no longer warranting a label of AI.
Emotional intelligence
Usually, when people are discussing AI or even Strong AI they are referring to
relatively mechanical operations and calm, dispassionate reasoning. The non-emotional
side of reasoning, including the ability to read the emotional state of another intelligent
entity, whether human or machine.
Although there have been experimental efforts to imbue machines with some sense of
emotive capabilities, that is still more of a science fiction fantasy than current or
imminent reality. It may exist in some relatively narrow or specialized niches, but not in
any broad and general sense.
Yes, someday AI systems will have at least some emotive capabilities, so-called
emotional intelligence, but not in the near future.
Short of full autonomy, agency is the capacity for an AI system to pursue a goal on
behalf of another entity, whether it be a human or some other digital system. An AI
system with agency is free to decide on its own how to achieve its given goal, but is not
free to set goals of its own, other than in a manner that is subsidiary to its assigned goal.
Note: In philosophy and sociology agency is taken to mean the same as autonomy is
used here, while computer science and AI use this alternative meaning for agency —
acting on behalf of another.
A driverless vehicle would fit the definition of agency but not true, full autonomy. The
vehicle might choose between alternative routes, but wouldn’t have the autonomy to
choose its own destination or to decide whether not to do as it is told by its owner or
operator.
An assistant has even more limited autonomy and agency, being given a specific task
and instructions and having very little freedom.
An intelligent digital assistant fits this definition for assistant.
For more on autonomy, agency, and assistants see the companion papers:
1. Reasoning
2. Knowledge and knowledge representation
3. Optimization, planning, and scheduling
4. Learning
5. Natural language processing (NLP)
6. Speech recognition and generation
7. Automatic language translation
8. Information extraction
9. Image recognition
10. Computer vision
11. Moving and manipulating objects
12. Robotics
13. Driverless and autonomous vehicles
14. General intelligence
15. Expert systems
16. Machine learning
17. Pattern recognition
18. Theorem proving
19. Fuzzy systems
20. Neural networks
21. Evolutionary computation
22. Intelligent agents
23. Intelligent interfaces
24. Distributed AI
25. Data mining
26. Games (chess, Go, Jeopardy)
Animal AI
We tend to focus on human intelligence when discussing AI, but AI can be applied to
the animal world as well, such as a personable robot dog, a robotic bird, or a robotic
flying insect, although the focus in these cases is far less about higher-level cognitive
abilities such as reasoning, mathematics, and creativity, and more about the physics,
physiology, sensory perception, object recognition, and motor control of biological
systems.
Robotics
Much of robotics revolves around sensors and mechanical motions in the real world,
seeming to have very little to do with any intellectual activity per se, so one could
question how much of robotics is really AI.
Alternatively, one could say that sensors, movement, and activity enable acting on
intellectual interests and intentions, thus meriting coverage under the same umbrella as
AI.
In addition, it can be pointed out that a lot of fine motor control requires a distinct level
of processing that is more characteristic of intelligence than mere rote mechanical
movement.
In summary, the reader has a choice as to how much of robotics to include under the
umbrella of AI:
In short, it’s not too much of a stretch to include virtually all of robotics under the rubric
of AI — provided there is at least some element of intelligence in the system, although
one may feel free to be more selective in specialized contexts.
Artificial life
Technically, some day scientists may be able to create artificial life forms in the lab that
have many of the qualities of natural biological life, but have possibly rather distinct
chemical bases, structures, and forms. Such artificial life could conceptually be imbued
with some form of intelligence as well — artificial intelligence for artificial life.
But, for now, such artificially intelligent artificial life remains the realm of speculation
and science fiction. Still, it would be very interesting and potentially very useful.
Granted, there might be more than a few ethical considerations.
An exception is virtual reality (VR), where even the laws of physics can be conveniently
ignored, if desired. Traditional chemistry and biology present no limitations to the
creativity of designer worlds in the realm of VR. In fact, one could say that all forms of
life in a VR world are artificial, by definition. One can even imbue otherwise inanimate
objects with any degree of life one chooses.
Ethics
Consult my longer paper for a discussion of ethical considerations for AI.
• What is AI?
• What is Artificial Intelligence?
• Basic Questions
I have some comments and questions on that topic in a companion paper, Can
Machines Think?
Diving even deeper, I have a longer list of questions designed to spur thought on this
matter in another companion paper, Questions about Thinking.
What’s the IQ of an AI?
Next question. Seriously, there is no clarity as to how how the human concept of IQ
could be adapted to machines. Some people have ideas about how to do it, but there is
no consensus. It’s almost kind of moot until we actually achieve Strong AI or something
fairly close.
Besides, given the malleable nature of software, the code of an AI system could be
quickly revised to adapt to whatever new test came along so that an AI would score
significantly higher than if the code hadn’t been tuned to the test.
But that’s the nature of AI today — it is relatively easy to identify specific and
relatively narrow niche cases and code up heuristics that work fairly well for those
narrow niches, making the software appear quite intelligent, even while it is far more
difficult or even near impossible with today’s technology to achieve true, full, Strong AI
which works equally well for all niches.
Still, it would be good to have a more objective measure of the level of intelligence of
an AI than simply weak or strong, or even my moderate level or my spectrum of
functional behavior and levels of competence.
Turing test
In theory, the so-called Turing test (also called The Imitation Game) can detect whether
a machine or AI is able to interact in such a human-like manner that no human observer
could tell that it was a machine by asking a finite set of questions.
There is some significant dispute about both whether the test is indeed a valid binary
test of intelligence (always arrive at the correct conclusion whether the test subject has
human-level intelligence or not) and whether claims to have passed the test are truly
valid.
The real bottom line is that as a thought experiment the test highlights the great
difficulty of definitively defining human-level intelligence in any deeply objective and
measurable sense.
That’s really only an issue for defining and testing for Strong AI. Weak AI has no such
strong testing requirements — even if only a fraction of human-level capability, or
seeming only partially human-like, that’s good enough for many applications.
Jack Krupansky
Dec 4, 2017
When considering robots, intelligent agents, and intelligent digital assistants, questions
of autonomy and agency arise. This informal paper attempts to define these key
concepts more clearly and explore questions of what are they, how they are different,
and how they are related.
Synthesized definitions for autonomy and agency will be provided after discussing all
the relevant aspects of these concepts.
Note that this paper is more focused on people, robots, and software agents rather than
on countries or autonomous regions within countries (e.g., Catalonia in Spain or
Kurdistan in Syria and Iraq), although the basic definition of autonomy still applies to
those cases as well.
Also note that sociology, philosophy, and agent-based modeling and simulation use the
terms agency and agent as the terms autonomy and autonomous entity are used in this
paper (freedom of choice and action, unconstrained by any other entity.)
For quick reference, see the section entitled Definitions of autonomy and agency for
the final definitions of these terms.
Dictionary definitions
A later section of this paper will come up with synthesized definitions for autonomy and
agency that are especially relevant to discussion of intelligent agents and intelligent
digital assistants, but the starting point is the traditional dictionary definitions of these
terms.
There are other meanings for entity, but those are the senses relevant to this paper.
Intelligent entities
Autonomy and agency are all about intelligent entities and their freedom to make
decisions and take actions, and their authority, responsibilities, and obligations.
Generally, an entity is any person, place, or thing. In the context of autonomy and
agency, an intelligent entity is a person or thing which is capable of action or operation
and at least some fraction of perception and cognition — thought and reason, coupled
with memory and knowledge.
More specifically, an intelligent entity has some sense of intelligence and judgment, and
is capable of making decisions and pursuing a course of action.
Computational entities
An intelligent entity can be a person or a machine or software running on a machine.
The latter are referred to as computational entities or digital entities. They include:
• Robots
• Driverless vehicles
• Smart appliances
• Software agents
• Intelligent agents
• Digital assistants
• Intelligent digital assistants
• Apps
• Web services
How much autonomy or agency a given computational entity has will vary greatly, at
the discretion of the the people who develop and deploy such entities based on needs,
requirements, desires, preferences, and available resources and costs.
Sometimes people want more control over their machines, and sometimes they value
greater autonomy, agency, or automation to free themselves from being concerned over
details.
Entities
As a convenience and for conciseness, this paper will sometimes use the shorter term
entity as implicitly referring to an intelligent entity, either a person or a computational
entity.
1. Action. Something that can be done by an entity. An observable effect that can
be caused in the environment.
2. Operation. Generally a synonym for action. Alternatively, an action that
persists for some period of time.
For example flipping a switch to turn on a machine is an action, while the ongoing
operation of the machine is an operation. The flipping of the switch was an operation
too, only of a very short duration.
If a machine would operate only while a button was depressed, the pressing and holding
of the button as well as the operation of the machine would both be actions and
operations.
1. Principal. An intelligent entity which has the will and desire to formulate an
objective or goal.
2. Agent. An intelligent entity which has the capacity and resources to pursue and
achieve an objective or goal on behalf of another intelligent entity, its principal.
1. A principal but not an agent. Does all actions itself, without any delegation to
agents.
2. An agent but not a principal.
3. Both a principal and an agent. A principal for subgoals.
4. Neither a principal nor an agent. Possibly an assistant for specific tasks, but not
any goals.
Authority
The authority of an intelligent entity is the set of actions that the entity is permitted to
take.
An agent would have limited authority related to the goal(s) that the principal is
authorizing the agent to pursue.
In the real world, many principals are in fact agents since they act on behalf of other
principals. A company has a board of directors, investors, and shareholders. Robots
have owners.
An agent has responsibility, expectations, and obligations as set for it by its principal.
An agent may act as it sees fit, provided that its actions satisfy any limitations or
constraints set by its principal.
In the real world, many principals are in fact agents since they act on behalf of other
principals. A company has a board of directors, investors, and shareholders. Robots
have owners. So a company or robot may have responsibilities, expectations, and
obligations set by somebody else.
General obligations
Regardless of obligations which result from autonomy and agency, all intelligent
entities will have general obligations which spring from:
• Physics. Obey the laws of physics. Reality. The real world. Natural law. For
example, gravity, entropy, and the capacity of batteries.
• Limited resources and their cost. For examples, the availability and cost of
electricity, storage, computing power, and network bandwidth.
• Laws. Obey the laws of man. Including regulations and other formalized rules.
• Ethics. Adhere to ethical codes of conduct. Including professional and industry
codes of conduct.
Ethics
Just to reemphasize from the previous section, that intelligent entities will have to
adhere to ethical considerations in the real world.
Liability
A principal may be exposed to liability to the extent that it enlists the aid of an agent
and that agent causes harm or loss or violate laws or rules while acting on behalf of the
principal.
Requested goals might have unintended consequences which incur unexpected liability.
An agent may be exposed to liability if it naively follows the guidance of its principal
without carefully reviewing whether specified goals, expectations, or obligations, might
cause harm or loss or violate laws or rules when carried out.
Elements of a goal
A goal must be:
1. Formulated. Clearly stated.
2. Planned. A strategy developed. A plan developed. Resources allocated. Tasks
identified.
3. Pursued. Individual tasks performed. Decisions may need to be made or revised
and the original plan adapted based on results of individual tasks.
4. Achieved or not achieved. The results or lack thereof.
1. Power. The principal has the power to set the objectives and goals to be
pursued. The agent has only the delegated power to select tasks to achieve the
objectives and goals set by the principal and to pursue them through actions, but
no power to change the objectives or goals themselves.
2. Action. The agent is responsible for performing the actions or tasks needed to
achieve the objectives and goals set by the principal. The agent is also
responsible for deciding what tasks and actions must be performed to achieve
the objectives and goals, and for coming up with a plan for performing them
3. Control. The principal controls what objectives and goals are to be pursued.
The agent controls what tasks and actions must be performed to achieve the
objectives and goals and how to perform them.
Contracts
Generally there is a contract of some form between a principal and its agent, which
clearly sets out the objectives and goals, responsibilities, expectations, and obligations
of both parties, both the principal and the agent.
The contract details what is expected of the agent, what the agent is expected to deliver,
what the agent needs to pursue the specified goals, including resources, and what
compensation the agent will receive in exchange for achieving the goals.
The contract also details any limitations or restrictions that will apply to the agent and
its work.
Contracts are needed both for human entities and for computational entities.
The latter requires that there is a principal involved, doing the empowerment, the
authorization to act on its behalf.
The former can exist even if there is no principal present. An intelligent entity can act
on its own interests, on its own behalf, being its own principal. An entity can be self-
empowering. That’s what it means for an entity to have agency in a traditional,
sociological or philosophical sense.
The first sense is true in both instances, where either a principal is present as an external
entity, and when no principal is present.
In the context of intelligent agents and intelligent digital assistants, agency usually
refers to the latter sense, that the agent is acting on behalf of the principal, which is
commonly a human user, but may also be some other computational entity, such as
another intelligent agent or a robot.
Assistants
A separate companion paper, Intelligent Entities: Principals, Agents, and Assistants,
will introduce the concept of an assistant, which is quite similar to an agent in the sense
that it is capable of performing the tasks needed to achieve goals, but can only perform
specific tasks as dictated by its principal without any sense of any larger goal or
objective that the task is needed to achieve, and with much less room for discretion as to
how to perform the tasks.
An assistant has limited agency in that it performs tasks on behalf of a principal but it
lacks the authority or capacity to decide which tasks to perform in the context of a goal
or objective.
Okay, technically, assistants do have a modest degree of autonomy, but very modest
and very minimal. Any system that doesn’t require a principal to be directly controlling
every tiny movement by definition is delegating at least a small amount of autonomy.
But not enough for the term autonomy to have any significant relevance to the freedom
of action of such a system. That’s the point of distinguishing assistants from agents —
to indicate the almost complete lack of autonomy.
An assistant has no authority, simply the responsibility for a specified task or action, as
specified, with little or no room for discretion or decision.
Control
A principal always has control over agents to which it has delegated responsibility for
goals, and control over assistants to which it has assigned specific tasks.
A principal could change or revise or even cancel goals, instructions which agents
would be obligated to comply with.
A principal can at any time request a status report on progress that an agent is making
on a goal or objective.
Robots
Superficially, robots would seem to be fully autonomous, but in reality they have the
more limited autonomy or partial autonomy of agents. After all, robots are owned and
work on behalf of their owners, performing tasks and pursuing goals as their owners see
fit, and dictate.
That said, as with an agent, a robot can be granted a significant level of autonomy and
be given fairly open-ended goals, so that they could actually be fairly autonomous even
if not absolutely fully autonomous.
But it’s also possible that someone might mistakenly grant a robot complete autonomy
and it might be difficult to regain control over the robot. Although, it would be possible
to make it illegal to grant a robot full autonomy.
The HAL computer in the 2001: A Space Odyssey movie and the Skynet AI network of
computers and machines in the Terminator movies were in fact machines which
somehow gained full autonomy — with quite scary consequences.
It would be interesting to see a science fiction movie in which fully autonomous robots
have a strictly benign and benevolent sense of autonomous responsibility. But maybe
that violates the strict definition of autonomy — if they act as if to serve people, then
they aren’t truly autonomous.
Maybe robots would need to exist in colonies or countries or planets or space stations of
their own, with full autonomy there, rather than coexisting within our human societies.
Robot societies and human societies could coexist separately and could interact, but
respecting the autonomy of each other, with neither in charge or dominating the other.
Maybe.
The mission will break down into objectives, which will break down into discrete goals.
The enterprise or organization may periodically review and adjust, revise, or even
radically change its mission and objectives. At its own discretion. That’s autonomy.
An agent is given a discrete goal to pursue. A small part of a larger mission and its
objectives. An agent does indeed have a mission and objective, but they are set by its
principal. An agent has no control over its mission or objective.
A principal has a larger mission and associated objectives for which discrete goals are
periodically identified and assigned to discrete agents. A principal sets its own mission
and objectives.
For more discussion of mission and objectives, see the companion paper, Intelligent
Entities: Principals, Agents, and Assistants.
So:
1. Mission independence. The entity can choose and control its own missions
rather than be constrained to pursue and follow a mission set for it by another
entity, a principal. This is closer to true autonomy.
2. Operational independence. The entity can decide for itself how to accomplish
operational requirements. This is independent of control of the overall mission.
This is characteristic of an agent, although an autonomous entity would tend to
also have operational independence as well.
So:
• Abstract: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.51.4431
• PDF:
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.25.8941&rep=rep1&t
ype=pdf
The abstract:
• With the recent rapid growth of interest in MultiAgent Systems, both in artificial
intelligence and software engineering, has come an associated difficulty
concerning basic terms and concepts. In particular, the terms agency and
autonomy are used with increasing frequency to denote different notions with
different connotations. In this paper we lay the foundations for a principled
theory of agency and autonomy, and specify the relationship between them.
Using the Z specification language, we describe a three-tiered hierarchy
comprising objects, agents and autonomous agents where agents are viewed as
objects with goals, and autonomous agents are agents with motivations.
I cite this reference here neither to blindly accept it nor to quibble with it, but simply to
provide a published foundation which readers can reference.
That said, I’ll offer a couple of relatively minor quibbles, more along the lines of how to
define terms:
1. I’d prefer to use the term entity or even intelligent entity rather than object. To
my mind, objects include trees, rocks, and mechanical machines, but generally
not include peoples and animals per se. Technically, intelligent entities are
indeed objects, but the term object doesn’t capture the essential meaning of an
intelligent entity.
2. The cited paper defines an object as an entity that comprises a set of actions and
a set of attributes.
3. This notion of having a set of actions or capable of acting is a bit more than the
traditional, real-world, non-computer science sense of the meaning of the
concept of an object.
4. A machine is capable of acting in some sense, but unless it has some sort of
robotic brain, it has no sense of sensing its environment and making decisions
about how to interact with its environment. A sense of agency is needed.
5. A washing machine or refrigerator would fit the meaning of an object in the
sense of the cited paper, although I would refer to them as assistants rather than
mere objects in a real-world sense. They have no agency, with no ability to
choose how to pursue a goal rather than to blindly perform a specific task. That
ability to perform tasks does fit the definition of an assistant used in this paper.
6. A driverless car would be a good fit for what I would call an intelligent entity
and would fit the concept of agent used in the cited paper. You tell the car where
you want to go and it figures out the rest, coming up with a plan and figuring out
what tasks are needed to get you to your objective.
7. Said driverless car would superficially seem to have a sense of autonomy, in that
it can move around without a person at the controls, but it lacks the ability to set
its goals. It can pursue and follow goals given to it, but not set them. In that
sense, both mine and the cited paper, said driverless car does not have
autonomy.
8. Driverless cars did not exist back in 1995, but I think even now the authors of
the cited paper would likely agree that a driverless car lacks the motivation or
ability to set goals that is required to meet the definition for autonomy.
9. As the paper would seem to agree, goals are set from motivations.
10. As the paper would seem to agree, being an agent does not automatically confer
the presence of motivations. Agents don’t need to be motivated. They just need
to be able to pursue and achieve goals.
11. In the context of software agents, which was indeed the context of that 1995
paper, I’d refer to degree of autonomy, meaning the extent to which the agent is
free to make its own choices, as opposed to the degree to which the agent’s
principal has already made choices and has decided to constrain the choices or
autonomy of the agent.
12. An upcoming companion paper, Intelligent Entities: Principals, Agents, and
Assistants, will explore this notion of principal with respect to autonomy.
13. The cited paper uses the term motivation to essentially mean that the agent has
the ability to set its own goals.
14. I agree with the cited paper that agents are all about goals.
15. The open issue is who sets the goals for a given agent.
16. In my terms, it is the principal which sets goals. That could be a person, or some
piece of software or even a robot. And this paper does allow for the prospect of
subgoals so that an agent can act as principal for a subgoal.
17. In the terms of the cited paper, an autonomous agent would correspond to my
concept of principal.
18. A key difference between the terminology of the cited paper and of this paper, is
that this paper first seeks to ground the terms in the real world of human entities
or people before extending the terms and concepts to the world of machines and
software.
Motivation
Motivation is a greater factor in autonomy, but can be relevant to agency as well.
A principal should clear have some good reason for its choices in setting objectives and
goals. Its motivation.
An agent might have some minor motivation for its choices as to what tasks to perform
to pursue and achieve the goals given to the agent by its principal, but those minor
motivations pale in significance to the larger motivation for why the goal should be
pursued at all, something only the principal can know.
The contract between principal and agent may likely express the motivation for each
goal or objective, although that expression may have dubious value to the agent.
One exception is when the specification for the objectives might be technically weak
and too vague, incomplete, or ambiguous, leaving the agent with the job of deducing the
full specification of objectives by parsing the motivation. That’s not the best approach,
but may be the only viable or sane approach.
With no mention of any principal or other external intelligent entity setting objectives
for the agent to follow.
That would be more compatible with the sense of principal as agent used in this paper,
where the agent is indeed setting its own objectives and goals.
ABM/ABS is a hybrid field, a mix of computer science and social science, and not
limited to computer science or social science, either. In fact, it can be applied to other
fields as well. Anywhere that there are discrete, autonomous entities that interact and
can have some sort of aggregated effect. ABM/ABS is more of a tool or method than a
true field per se.
For all intents and purposes, ABM/ABS could be considered part of social science and
sociology.
Definitions
As promised, here are the synthesized definitions of autonomy and agency as used in
this paper:
1. Autonomy. Degree to which an intelligent entity can set goals, make decisions,
and take actions without the approval of any other intelligent entity. The extent
to which an entity is free to exert its own will, independent of other entities. Can
range from the full autonomy of a principal to the limited autonomy or partial
autonomy of an agent to no autonomy for an assistant. The entity can decide
whether to take action itself or delegate responsibility for specific goals or
specific tasks to other intelligent entities, such as agents and assistants.
2. Agency. Ability of an intelligent entity, an agent, to plan, make decisions, and
take actions or perform tasks in pursuit of objectives and goals provided by a
principal. The agent has limited autonomy or partial autonomy to decide how to
pursue objectives and goals specified by its principal. A contract between
principal and agent specifies the objectives and goals to be pursued, authorizing
action and obligations, but leaving it to the agent to decide how to plan, define,
and perform tasks and actions. The agent may decompose given objectives and
goals into subgoals which it can delegate to other agents for whom this agent is
their principal. Note: In sociology and philosophy agency refers to autonomy or
the extent to which an entity is free to exert its own will, independent of other
entities.
These definitions should apply equally well to human and computational entities, or at
least be reasonably compatible between those two domains.
Terms used within those definitions are defined elsewhere in this paper, including:
• Entity
• Intelligent entity
• Principal
• Agent
• Assistant
• Objective
• Goal
• Task
• Action
• Subgoal
• Responsibility
• Authority
• Delegation
• Contract
Autonomous systems
Generally and loosely speaking, people speak of autonomous systems, whether it be a
robot, a software application, a satellite, a deep space probe, or a military weapon.
This is not meant to imply that such systems are fully, completely, and absolutely
autonomous, but simply that they have a high degree of autonomy. Or what we call
limited autonomy or partial autonomy in this paper.
And to draw a contrast to directly or remotely controlled systems such as drones where
every tiny movement is controlled by a human operator.
As noted for autonomous systems in general, even so-called lethal autonomous weapons
will not typically be fully, completely, and absolutely autonomous.
They may have a significantly higher degree of autonomy, but not true, full autonomy.
There is some significant effort to assure that at least some minimal degree of human
interaction occurs, what they call meaningful human control. That’s still a somewhat
vague term, but the concept is still in the early stages.
Even an automatic rifle or machine gun has a trigger, causing it to stop firing when a
person decides to stop holding the trigger. That’s meaningful human control.
Even before we start getting heavily into artificial intelligence (AI), there are already
relatively autonomous systems such as the Phalanx CIWS close-in weapon system gun
for defense against anti-ship missiles. It is fully automated, but with oversight by a
human operator. It can automatically detect, track, and fire on incoming missiles, but
the operator can still turn it off.
A big ethical concern for lethal autonomous weapons is the question of accountability
and responsibility. Who is responsible when an innocent is harmed by such an
autonomous weapon when there is no person pulling the trigger?
The only truly fully autonomous lethal weapon I know of is the minefield. Granted it
has no AI or even any digital automation, and the individual mines are not connected,
but collectively it acts as a system and is fully, absolutely autonomous. It offers both the
best and worst of military and ethical qualities. It has no discrimination. It is fully
autonomous. It is quite reliable. It is quite lethal. It is quite humane. It has absolutely no
compassion. It has no accountability. No responsibility. And no human operator can
even turn it off other than by laboriously and dangerously dismantling the system one
mine at a time. Somebody put the mines there, but who?
Now, take that rather simple conception of a minefield and layer on robotics, digital
automation, and even just a little AI, and then you have mountains of technical,
logistical, and ethical issues. That’s when people start taking about killer robots and
swarms.
Sovereignty
Another related term which gets used in some contexts as a rough synonym for both
autonomy and independence is sovereignty.
Summary
To recap:
Autonomy refers to the freedom of an intelligent entity to set its own objectives and
goals and pursue them, either by acting directly itself or delegating goals to agents.
Agency refers to the freedom of an intelligent entity (agent) to pursue goals delegated to
it by its principal as it sees fit, although subject to expectations and obligations
specified by its principal in the contract which governs their relationship.
An agent owes its allegiance to its principal.
Although in sociology, philosophy, and agent-based modeling and simulation the terms
agency and agent are used and defined as the terms autonomy and autonomous entity
are in this paper.
One can also refer to degree of autonomy, so that an agent has some limited degree of
autonomy and so-called autonomous systems have a fair degree of autonomy even
though they do no have full, complete, and absolute autonomy.
Lethal autonomous weapons? Coming, but not here yet, and not in the very near future.
What Is an Assistant?
Jack Krupansky
Definition
Unfortunately, there is no single, universal definition for what it means to be an
assistant.
From Google:
From Merriam-Webster:
From Dictionary.com:
From Wikipedia:
• Administrative assistant
• Artificially intelligent assistant
• Assistant account executive
• Assistant coach
• Assistant commissioner
• Assistant facility manager
• Assistant manager
• Assistant operations manager
• Assistant professor
• Assistant program director
• Assistant secretary
• Call center customer care assistant
• Customer care assistant
• Chatbot/virtual assistants
• Community assistant
• Customer care assistant
• Dental assistant
• Deputy assistant secretary
• Digital assistant
• Digital virtual assistant
• Editorial assistant
• Executive assistant
• Executive virtual assistant
• Finance assistant
• Health assistant
• Intelligent automated virtual assistant
• Intelligent digital assistant
• Intelligent personal assistant
• Intelligent virtual assistant
• Integration assistant
• Lab assistant, laboratory assistant
• Legal assistant
• Marketing administrative assistant
• Medical assistant
• Nursing assistant
• Outside advertising sales assistant
• Program and executive assistant
• Project assistant
• Recruiter assistant
• Retail assistant
• Sales assistant
• Staff assistant
• Student assistant
• Teacher assistant
• Team Assistant
• Virtual assistant
• Virtual digital assistant (VDA)
• Virtual office assistant
The first is a real live person, a human, who just happens to work from a location other
than an office of the company or organization being served, such as from home or at a
third-party firm providing such services. They may work using the telephone, email, or
online chat.
The second is a digital simulation of a person, able to respond to a subset of the requests
that a normal person would be able to handle.
Alternatively, the software may be able to handle a much deeper or broader range of
requests than any single human being could be readily trained to handle.
Related terms
Synonyms from Thesaurus.com:
• abettor
• accessory
• accomplice
• adherent
• adjunct
• aide
• aide-de-camp
• ally
• appointee
• apprentice
• associate
• attendant
• auxiliary
• backer
• backup
• coadjutor
• coadjutant
• collaborator
• colleague
• companion
• confederate
• cooperator
• deputy
• fellow worker
• flunky
• follower
• friend
• gofer
• help
• helper
• helpmate
• mate
• partner
• patron
• peon
• representative
• right-hand person
• secretary
• subordinate
• supporter
• temp
• temporary worker
• Chatbot
• Customer care representative
• Customer service representative
• Medical scribe
• Personal digital assistant
• Project management specialist
• Receptionist
• Registration and scheduling specialist
• Representative
• Scribe
• Social bot
• Software assistant
• Technical assistance
Principal
This paper will use the term principal to refer to the boss or manager to whom an
assistant reports and who assigns them work or tasks to be completed. The person for
whom the assistant works. The person whom the assistant assists.
1. Simple assistant. Your basic, run of the mill assistant. Minimal education,
minimal prior experience, minimal technical knowledge. And minimal level of
responsibility, a well-defined collection of administrative-type tasks. The
administrative assistant is representative of this level.
2. Specialized or technical assistant. May require a more specialized degree or
training. Requires the ability to master particular subject matter. Requires the
ability to perform specialized tasks, beyond simple administrative tasks. Medical
assistants and research assistants are representative of this level.
3. Executive assistant. Not so much a matter of education or specialized
knowledge, but rather able to handle more significant responsibilities. Able to
accomplish tasks or pursue goals of their principal without being explicitly
directed to perform each task. Almost literally able to read their principal’’s
mind, or at least be able to frequently and commonly anticipate predictable
requests. More significantly, is able to serve their principal better than their
principal could directly and explicitly instruct them.
Goals are more complex collections of tasks that require some significant level of
complex reasoning, judgment, and careful decision. And significant planning.
A goal is specified by stating the objective to be achieved. The objective itself rather
than the details of how to achieve the objective.
Tasks generally don’t require much deep thought, just slogging through the work.
Goals tend to require deeper, more careful, and more insightful thought.
Personal services
Performing services of a personal rather than professional nature for the principal is a
primary function of a personal assistant, although specialized assistants as summarized
in the preceding section may focus more heavily on professional services.
• Clothing
• Appearance
• Hygiene
• Housekeeping
• Food
• Drink
• Dining
• Transportation and travel
• Entertainment
About the only limitation is that the nature of the task needs to be clearly and simply
articulated so that the task can be performed in a direct, straightforward manner by the
assistant, who may not have the detailed education and knowledge to understand the
deeper why of the task. Their function is simply to excel at the how of the task.
Software service — intelligent digital assistant
A subsequent paper will dive into greater detail, but an intelligent digital assistant is a
software service, possibly coupled with a specialized hardware device, such as a smart
speaker, or merely a feature offered on a general purpose computing device such as a
personal computer, tablet, smartphone, or wearable computer (such as a digital
wristwatch), which offers some interesting set of the abilities of a traditional, human
assistant, most notably answering questions and performing tasks using voice and
natural language processing (NLP) backed by artificial intelligence (AI).
Examples include Amazon Alexa/Echo, Apple Siri, Google Assistant, and Microsoft
Cortana.
Synthesized definition
After digesting available material on the topic, I have come up with the following
synthesized definition for assistant:
Jack Krupansky
Not all entities are created equal, whether they are people, robots, software agents, or
intelligent digital assistants. This informal paper explores the relative roles of
principals, agents, and assistants. These respective roles apply to both the real world of
people as well as the digital world of intelligent agents and intelligent digital assistants.
A prior paper, What Are Autonomy and Agency, explored autonomy and agency to
some degree. This paper goes deeper, focusing on the roles of the entities themselves as
well as their autonomy and agency.
Although the ultimate goal of this paper is to get at the computational aspects of
intelligent entities, principals, agents, and assistants, most of the concepts should apply
equally well to the human world and the world of computers.
The essential concepts explored in this paper relate to agency, which itself relates to
autonomy.
Much of the motivation for the depth in this paper is to enable the concepts of
intelligent entities, agency, and autonomy to be successfully implemented in artificial
intelligence systems. Outside of AI systems, people can get by with casual and intuitive
interpretations of these powerful concepts, but no computer system is going to figure
out all of the nuances of these concepts on its own without sufficient depth being
programmed in from the get-go.
Again, agency is the central focus of this paper, from two aspects:
Not all intelligent entities are locked into strict roles of principals, agents, and assistants,
but they are not related to the concept of agency, which is the focus of this paper. See
the section on Other categories of intelligent entities.
Other key concepts and terms that will be defined and explored in this paper:
1. Agency
2. Autonomy, autonomous
3. Independence, independent
4. Dependence, dependent, dependency
5. Control
6. Mission
7. Purpose
8. Objective
9. Contract
10. Capability
11. Reputation
12. Requirement
13. Delegation
14. Goal
15. Task
16. Action
17. Operation
Most concepts and terms will be provided with both their traditional dictionary
definitions as well as refined definitions which more closely capture the essential
meaning of the concepts and terms in the context of intelligent entities and their
relationships, especially agency.
Simply put, intelligence is required to do interesting things that cannot be done with a
mere pre-programmed sequence of rote, mechanical steps.
1. Defining a mission.
2. Defining objectives to achieve the mission.
3. Planning.
4. Making decisions.
5. Dealing with the unexpected.
6. Dealing with ambiguity.
7. Coping with the vagaries of human nature.
8. Coping with the vagaries of weather and other natural phenomena.
9. Dexterity that cannot be easily or cheaply replicated by a machine.
10. Pattern recognition that cannot be easily or cheaply replicated by a machine.
11. Creativity.
12. Imagination.
In some (or even many) cases it may be theoretically possible to arrange for a dumb
machine (or person) to be trained to accomplish a task or pursue a goal, but the cost or
risk of doing so might cause one to fall back on an intelligent entity, human or machine,
rather than deal with the complexities and vagaries of dumb entities, whether human or
machine.
How much intelligence is needed?
Enough to respond to common obstacles and variations in the operating environment.
In defense of dogs, it is worth noting that even some human assistants are not very
competent at some of those tasks and activities that are beyond the abilities of a dog, so
we shouldn’t necessarily hold those limitations against dogs per se.
I hesitate to grant them the full status of intelligent entity and assistant, but I don’t want
to completely dismiss their value either.
Maybe I’ll just leave it up to the discretion of the reader — you have my permission to
confer or deny full status of intelligent entity and assistant, as you see fit.
Me, I’ll grant our canine friends provisional status as intelligent entities until such time
as someone can offer a convincing and satisfying argument against such status.
Solving bigger problems
The essential rationale for principals, agents, and assistants is to enable entities to
address significantly larger problems than they could if working alone.
The principal defines the larger mission and breaks it down into manageable objectives,
which can in turn be broken down into narrower goals, each of which can be delegated
to an agent using a contract which specifies the details of the goal, or possibly into
more discrete tasks, each of which can be delegated or assigned to an assistant.
Each agent studies and analyzes the goal or objective it was assigned by its principal,
comes up with a strategy for how to achieve the goal, comes up with a plan to pursue
that strategy by decomposing the goal into individual tasks, and then parcels each task
out to an assistant, or if simple enough, performs the individual tasks itself.
Each assistant then focuses on a single task, sequencing through the specific actions or
operations needed to complete its task.
One principal can employ any number of agents. And possibly assistants as well.
Even beyond that, an agent can decompose its assigned goal into subgoals and then
delegate each subgoal to yet another agent, each with its own contract for its specific
subgoal.
As well, an assistant can partition a large assigned task into smaller tasks and then
delegate or assign the smaller tasks to yet other assistants.
That said, it is instructive to look at the full range of meaning of the term entity first.
• Object. Something that exists or at least appears to have form, substance, shape,
or can be detected in some way, or can be experienced with the senses or
imagination, or manipulated by a computer, either as a real-world object or an
imaginary object, such as a media object, mental object, or computational
object, and can be distinguished from its environment. See also: entity, a subset
of which are objects. Whether liquid and gaseous matter should be considered to
be objects is debatable, but they are under this definition. A storm could
certainly be treated as an object even though it consists only of air and water.
Alternatively, the entity or matter at which an action is being directed — see
also: subject.
Technically, an entity does not even have to be a person or smart machine, so for the
context of this paper we need to restrict the definition to the subset of entities that are
people and smart machines — sapient entities, alternatively known as intelligent
entities. And smart machines are also known as computational entities.
The intention in the context of this paper is that principals, agents, and assistants are all
sapient entities.
Granted, wisdom is a bit of a stretch for current AI, but anything better than really dumb
machines is worth at least partial credit.
This paper focuses on sapient entities, but for convenience and conciseness simply
refers to them as simply entities with sapience implied, or as intelligent entities.
How intelligent?
How exactly should we distinguish dumb entities (human or machine) from intelligent
entities? That’s an open matter of great debate.
For starters, review the extensive explorations of the nature of intelligence (both human
and machine) in the companion paper, Untangling the Definitions of Artificial
Intelligence, Machine Intelligence, and Machine Learning, particularly the sections
Levels of Artificial Intelligence and Spectrum of Functional Behavior.
To oversimplify, you have Weak AI and Strong AI, with plenty of shades of gray
between.
In short, you can credibly claim that you or your computer software is intelligent if it is
at least somewhat intelligent, exhibiting behavioral qualities that are at least quasi
human-like even if not all that sophisticated.
Put another way, an agent or assistant really only needs to be able to do something,
anything useful so that you can feel that it has taken some interesting, significant
burden off of your shoulders and made your life at least a non-trivial degree of better or
at least easier. And even if it is only a trivial degree of improvement, that’s likely good
enough as well.
Sure, ten years from now intelligent entities are going to be really intelligent, but we
should be content to crawl before we walk let alone run and sprint.
There are other meanings for entity, but those are the senses relevant to this paper.
1. Autonomy. Degree to which an entity can set goals, make decisions, and take
actions without the approval of any other entity. Can range from the full
autonomy of a principal to the limited autonomy of an agent to no autonomy for
an assistant. The entity can decide whether to take action itself or delegate
responsibility for specific goals or specific tasks to other entities, such as agents
and assistants.
2. Agency. Ability of an entity, an agent, to plan, make decisions, and take actions
or perform tasks in pursuit of objectives and goals provided by a principal. The
agent has limited autonomy to decide how to pursue objectives and goals
specified by its principal. A contract between principal and agent specifies the
objectives and goals to be pursued, authorizing action and obligations, but
leaving it to the agent to decide how to plan, define, and perform tasks and
actions. The agent may decompose given objectives and goals into subgoals
which it can delegate to other agents for whom this agent is their principal.
Degrees of autonomy
In the real world, autonomy is not a binary all or nothing proposition. It’s a spectrum
with unlimited gradations, the most common and significant in the context of this paper
being:
But can real principals in the real world ever have absolutely full autonomy?
In a technically purist sense, no. Real, practical principals will be constrained by:
But other than that, we can consider principals to have full autonomy.
That’s fine for people, but what about robots and AI? Well…
The HAL computer in the 2001: A Space Odyssey movie and the Skynet AI network of
computers and machines in the Terminator movies were in fact machines which
somehow gained full autonomy — with quite scary consequences.
It would be interesting to see a science fiction movie in which fully autonomous robots
have a strictly benign and benevolent sense of autonomous responsibility. But maybe
that violates the strict definition of autonomy — if they act as if to serve people, then
they aren’t truly autonomous.
Maybe robots would need to exist in colonies or countries or planets or space stations of
their own, with full autonomy there, rather than coexisting within our human societies.
Robot societies and human societies could coexist separately and could interact, while
respecting the autonomy of each other, with neither in charge or dominating the other.
Maybe. But no time soon.
Even if you did manage to put a robot on an uninhabited island, on an unmanned ship,
or even launched into space never to return, it’s not clear how you could give up legal
ownership and responsibility so that the robot was truly autonomous. There would have
to be a change in our laws to permit such an emancipation of property, ala the concept
of emancipation of minors (children.) Merely abandoning or freeing a robot would not
address the legal aspect of ownership.
For now, and the indefinite future, robots and AI systems will have owners, who have
control over them, which is inconsistent with full autonomy.
For now, and the indefinite future, robots will not be citizens or have the rights of
citizens.
For now, and the indefinite future, driverless cars will go where their owners or
occupants tell them to go. As such, a driverless car would be more of an agent rather
than a principal.
So, the proper characterization is that robots and AI systems can be semi-autonomous
with limited autonomy.
1. not dependent
2. not subject to control by others — self-governing
3. not affiliated with a larger controlling unit
4. not requiring or relying on something else — not contingent
5. not looking to others for one’s opinions or for guidance in conduct
6. not bound by or committed to a political party
7. not requiring or relying on others (as for care or livelihood)
8. being enough to free one from the necessity of working for a living
9. showing a desire for freedom
10. not determined by or capable of being deduced or derived from or expressed in
terms of members (such as axioms or equations) of the set under consideration
11. having the property that the joint probability (as of events or samples) or the
joint probability density function (as of random variables) equals the product of
the probabilities or probability density functions of separate occurrence
12. main
13. neither deducible from nor incompatible with another statement
14. one that is independent
15. one that is not bound by or definitively committed to a political party
16. someone or something that is not connected to others of the same kind
17. a person who does not belong to a political party
18. not under the control or rule of another
19. not connected with something else
20. not depending on anyone else for money to live on
21. thinking freely : not looking to others for guidance
Freedom of action
Autonomy and independence are terms for referring to the degree of freedom of action
of an entity.
That includes free will or the ability to make decisions without external constraint as
well.
Generally, agents have only limited independence since they are acting on behalf of a
principal.
That said, autonomy and independence are quite distinct in the political domain, with no
sense of degrees, levels, or gradations. For example, an autonomous region does not
have independence regardless of how autonomous it is. If it had full autonomy in the
sense used in this paper, it would be considered independent rather than merely
autonomous.
1. the state of relying on or needing someone or something for aid, support, or the
like
2. reliance
3. confidence
4. trust
5. an object of reliance or trust
6. the state of being conditional or contingent on something, as through a natural
or logical sequence
7. subordination or subjection
1. hanging down
2. determined or conditioned by another — contingent
3. relying on another for support
4. subject to another’s jurisdiction
5. subordinate
6. not mathematically or statistically independent
7. equivalent
8. one that is dependent
9. a person who relies on another for support
10. relies on someone else for most or all of his or her financial support
11. decided or controlled by something else
12. needing someone or something else for support, help, etc.
13. a person (such as a child) whose food, clothing, etc., you are responsible for
providing
14. determined by something or someone else
15. relying on someone else for support
16. a person who depends upon another for support
Dependence of a principal
A principal can be dependent on:
• Other principals.
• Agents to which it has delegated goals.
• Assistants to which it has delegated tasks.
• Resources needed to pursue its mission.
• Customers, clients, and users for their business or patronage.
Dependence of an agent
An agent can be dependent on:
Dependence of an assistant
An assistant can be dependent on:
• A principal or agent for its task assignments, the resources needed to complete
its tasks, and for its overall employment.
Dictionary definitions of control
Definition entries from Merriam-Webster definition of control:
Definition of control
1. Control. To limit, restrain, direct, guide, or influence the decisions or actions of
another entity.
Controlling entities
A principal controls agents and assistants.
1. a definite piece of work assigned to, falling to, or expected of a person — duty
2. any piece of work
3. a matter of considerable labor or difficulty
1. to entrust to another
2. to appoint as one’s representative
3. to assign responsibility or authority
4. to give (control, responsibility, authority, etc.) to someone
5. to trust someone with (a job, duty, etc.)
6. to choose (someone) to do something
7. to make responsible for getting something done
8. to entrust or transfer (as power, authority, or responsibility) to another
9. to transfer (one’s contractual duties) to another
10. to empower a body (as an administrative agency) to perform (a governmental
function)
1. an agreement between two or more parties for the doing or not doing of
something specified
2. an agreement enforceable by law
3. the written form of such an agreement
4. the formal agreement of marriage — betrothal
1. the quality or state of being able, especially physical, mental, or legal power to
do something
2. natural aptitude or acquired proficiency, natural talent or acquired skill
3. the power or skill to do something
1. having suitable or sufficient skill, knowledge, experience, etc., for some purpose
2. having sufficient skill, knowledge, etc. — capable
3. properly qualified
4. adequate but not exceptional
5. having legal competence, as by meeting certain minimum requirements of age,
soundness of mind, or the like
6. proficient
7. suitable or sufficient for the purpose
8. properly or sufficiently qualified
9. capable of performing an allotted or required function
10. legally qualified or fit to perform an act
11. able to distinguish right from wrong and to manage one’s affairs
1. something required
2. something wanted or needed — necessity
3. something essential to the existence or occurrence of something else —
condition
4. something that is needed or that must be done
5. something that is necessary for something else to happen or be done
6. something that is necessary
Definition of mission
1. Mission. The larger and core purpose, general focus, and target audience for an
intelligent entity, primarily a principal, beyond immediate and specific
objectives, goals, and tasks. Objectives follow from the mission. Alternatively,
the immediate objective, goal, or task for an intelligent entity.
In the context of this paper, mission is generally used in the former sense, the larger
purpose and core purpose of a principal.
Mission of a principal
Generally, a principal will have a clearly defined mission. It’s objectives follow from
that mission.
Mission of an agent
Generally, an agent will not have a mission of its own with regard to its work on a
specific goal, deferring to the mission of the principal on whose behalf it is pursuing the
goal.
That said, an agent will commonly have the implied mission of serving principals,
satisfying their requirements, successfully completing contracts, and otherwise
establishing a solid track record of satisfying the requirements of principals.
Mission of an assistant
Generally, an assistant does not have a mission of its own. The mission of an assistant
in the traditional sense is simply to serve its principal by successfully completing tasks.
Definition of objective
1. Objective. Larger or more general target or aim to be pursued or achieved by an
entity as part of its mission. Not as specific as a goal, but not uncommonly goal
and objective are used as synonyms. The intent here is that a goal has a narrower
scope while an objective has a broader scope. Sometimes referred to as a
strategic objective.
A principal will translate its mission into objectives, and then translate each objective
into one or more goals or tasks that can then be delegated to agents and assistants.
In a simpler sense, the objective of an agent is simply the goal for which the agent has
been contracted.
Definition of delegation
1. Delegation. The process of an entity identifying a subset of its goals and tasks
that can be offloaded to another entity, or to multiple entities.
Definition of responsibility
1. Responsibility. Obligation or expectation of delivery of goods, services,
information, or guidance or performance or achievement of an entity by another
entity, typically as a result of delegation under a contract.
Responsibility of an assistant
Assistants would generally not operate under a formal contract for a given work request,
other than a general contract of employment that covers all of their work.
Rather, their responsibilities would be informally given with each task that they are
assigned.
Definition of contract
1. Contract. Agreement between two intelligent entities specifying a relationship
in which goods, services, or payments are exchanged according to agreed upon
terms and conditions, with the intention of pursuing and achieving a specified
goal. Commonly between a principal and agent, or between a principal or agent
and an assistant.
Definition of capabilities
1. Capabilities. Detailed list of what an intelligent entity can do and accomplish.
And in some cases non-intelligent entities. What types of objectives, goals,
tasks, actions, and operations the entity can perform. What abilities, skills,
knowledge, expertise, talents, aptitudes, competencies, and proficiencies the
entity has. Also what education, training, experience, credentials, qualifications,
and licensing the entity has.
Definition of reputation
1. Reputation. Track record for an entity delivering on the terms and conditions of
contracts for its capabilities. Knowledge shared by entities which have
previously contracted with the entity.
Reputation includes:
For example flipping a switch to turn on a machine is an action, while the ongoing
operation of the machine is an operation. The flipping of the switch was an operation
too, only of a very short duration.
If a machine would operate only while a button was depressed, the pressing and holding
of the button as well as the operation of the machine would both be actions and
operations.
A higher level of intellectual effort is needed for the first two levels.
Definition of requirement
1. Requirement. A capability that is needed by an entity to achieve some goal.
Requirements are the foundation of a contract between two entities. They specify what
the contracting party of the contract needs and expects.
One difference between requirements and capabilities is that a requirement may specify
a range of acceptable capabilities such that any entity with a capability that falls
somewhere in that range will be acceptable.
Or, a requirement may be optional so that even if entities with that capability might be
preferred, entities without that capability would still be acceptable.
Assistant
Synthesized definition of assistant from What Is an Assistant?:
Technicians
A technician is simply an assistant who has specialized technical training for a set of
tasks in some specialized area.
Organizations
In additions to individual intelligent entities acting in the roles of principal, agent, and
assistant, entire organizations can act in each role as well.
A business may contract out to a service company such as a law firm, accounting firm,
or marketing firm to act as its agent.
A service firm could hire or contract out to another service firm to perform specific
tasks.
Each firm would of course have individual staff members performing the duties of
principals, agents, and assistants.
Interactions
Intelligent entities can interact both with other intelligent entities and non-intelligent
entities as well.
In the context of principals, agents, and assistants, the interactions can be:
In the context of principals, agents, and assistants, the relationships can be between:
1. With animals
2. With pests.
3. With disease.
4. With one’s own body.
5. With a patient’s body. Medical treatment.
6. With nature.
1. Partners. Two or more principals can be partners who work very closely
together for specific objectives or a shared mission.
2. Allies. Two or more principals or agents can be allies, having shared interests,
and related objectives, missions, or goals.
3. Friends. Entities whose interests are roughly compatible even if they do not
directly facilitate achieving their respective missions, objectives, goals, or tasks.
4. Enemies. Two or more principals can be enemies seeking the same mission or
objective but with opposing interests.
5. Adversaries. Less severe synonym for enemy.
6. Competitors. Less severe synonym for adversary.
7. Antagonists. Two or more principals, agents, or assistants can be antagonists
whose missions, objectives, goals, or tasks conflict in ways that makes it
difficult for each entity to achieve what it seeks even if not outright enemies,
competitors, or adversaries. For example, landlords, vendors, criminals.
Connections
In addition to interactions and relationships, entities can have connections, such as:
Legal liability
Legal liability can arise in various ways:
Future work
The theory, technology, implementation, and practice of intelligent entities, principals,
agents, assistants, autonomy, and agency will remain works in progress for the
foreseeable future.
Jack Krupansky
Examples include Amazon Alexa/Echo, Apple Siri, Google Assistant, and Microsoft
Cortana.
This informal paper will briefly explore the nature and capabilities of intelligent digital
assistants.
Technically, a digital assistant does not need to use voice or even natural language, but
in the context of this paper, the term digital assistant will be used as a shorthand for
intelligent digital assistant and presume that it is voice-enabled with natural language
processing.
Purpose
What is the purpose of a digital assistant? As Google puts it:
• Request information.
• Perform tasks.
• Voice input.
• Natural language processing (NLP).
• Voice output.
Two other distinguishing qualities are that execution of requests can be based on not
only the raw input request or command from the user, but also:
And, as with most devices and services, personal preferences of the user will be taken
into account.
Features
This generic list of features of digital assistants is not intended to be absolutely
comprehensive, but should be fairly representative:
For the purposes of this paper, the following points can be made about intelligence and
intelligent digital agents:
In short, the current crop of intelligent digital assistants exhibit some significant
qualities normally associated with intelligence, and even seem human-like, but only in a
fairly minimal and superficial sense. It is certainly better than nothing, but just a start
rather than anywhere near the finish line.
• Amazon Alexa/Echo
• Apple Siri
• Microsoft Cortana
• Google Assistant
It is beyond the scope of this paper to delve into specific product features or
recommendations for such products.
• Amazon Alexa/Echo
• Apple Siri
• Microsoft Cortana
• Google Assistant — Google Home devices
Connected intelligence, Internet-enabled
A key aspect of the design of this latest wave of digital assistants is that they are
services running on servers in the cloud, where most of the AI capabilities are in the
cloud, with the connected device seen and used by the user simply serving as an input
and output device.
• Who actually owns the user’s data and records of all requests and actions made
by the user?
• What exactly can and can’t the digital assistant vendors do with any of that user
data?
• Can the vendor give any third-parties access to that user data?
• How secure is that user data, really? Says who?
• Are user interactions with digital assistants vulnerable to man in the middle
attacks or using malware installed in the user device?
• How often is security and privacy of user data audited, and by what technical
means?
• What level of technical skill might be sufficient to hack into user data?
• Might government, foreign government, or intelligence services possess the
technical skills and means to hack user data?
• What assurances does a user have that vendor staff could not theoretically hack
user data as an inside job? For financial gain for to pursue a social or political
agenda.
• Can any of that user data be sold?
• Does the user have any way to get access to all of the data on them?
• Can a user move their data, including complete usage history to another vendor
or different type of device?
• Does the user have any way to scrub or delete some or all of the data on them?
• Is there a retention policy for user data?
• What rights does the user retain or forfeit with regard to court orders to access
their data? Both criminal and civil.
• How vigorously will vendors defend the rights of the user in the face of court
orders? Says who?
• In what legal jurisdiction(s) does the the user data reside? Servers and data
centers.
• Does the user have any control or ability to select a jurisdiction? Especially with
regard to court orders and actions of law enforcement in those jurisdictions.
• Might the user data be kept in more than one legal jurisdiction? Multiple copies
or distributed between servers in different data centers.
• Is location data given the same protection as interaction data?
• Can a user shield their location even if their interaction data is accessed, such as
through a court order?
• Can a parent or legal guardian get access to user data of children or relatives?
• Can a user allow another user to access their data?
• Can users share data?
• Desktop computers
• Laptop computers
• Tablet computers
• Smartphones
• Smart wristwatches
• Wearable computers
• Smart speakers
• Smart TVs
• Smart appliances, smart kitchen appliances
As mentioned in the previous section, the real intelligence is off in the cloud, with the
user’s device or computer used only to communicate with the cloud-based services.
Smart speakers
Smart speakers are the rage right now, with Amazon Echo, Google Home, and soon
Apple HomePod.
It’s a bit of a misnomer to say that the speakers themselves are smart since the actual
speakers are simply output devices and the real smarts is driven by microphones
included in the same physical box as the speakers.
The microphones pick up your voice and send it off to servers in the cloud to be
processed by the actual AI algorithms, before sending audio back to the actual speakers
for you to listen to the result.
Equivalent terms
As with any new and evolving technology, the terminology around intelligent digital
assistants is fluid, in a state of flux, and still unsettled.
All of the following terms are roughly equivalent to intelligent digital assistant, or at
least used as if equivalent despite nuances of differences:
• AI assistant
• AI digital workforce platform
• AI voice assistant
• AI-powered virtual agent
• AI-powered voice assistant
• Artificial intelligence voice assistant
• Artificial-intelligence assistant
• Artificially intelligent assistant
• Bot
• Chatbot
• Chatterbot
• Connected assistant
• Connected intelligent assistant
• Digital agent
• Digital assistant
• Digital virtual assistant
• Digital voice assistant
• Intelligent assistant
• Intelligent digital assistant
• Intelligent personal assistant
• Intelligent virtual assistant
• Personal AI assistant
• Personal assistant
• Personal assistant voice apps
• Personal digital assistant
• Smart assistant
• Smart digital assistant
• Socialbot
• Virtual assistance
• Virtual assistant
• Virtual customer assistant
• Virtual digital assistant
• Virtual personal assistant
• Voice AI capabilities
• Voice AI–capable device
• Voice assistant
• Voice-enabled digital assistant
• Voice-powered digital assistant
Not all bots, chatbots, socialbots, or digital or virtual assistants are necessarily voice-
activated or use voice response. They may use text.
Not all bots or socialbots recognize natural language. They may simply act in a way that
mimics human behavior using a variety of heuristics such as recognizing keywords that
are significant for the particular subject matter domain which the bot is designed for.
Related terms
Some other terms that might sometimes be used to refer to digital assistants:
• Agent
• Digital agent
• Intelligent agent
• Software agent
What is the proper term?
Alas, there is no single, widely acknowledged proper term for the products and services
covered by this paper. To wit, here are the common characterizations of the Big Four
products:
Those are the terms used in the respective Wikipedia articles for those products and
services.
Given how fluid and unsettled the use of the terminology is, this paper arbitrarily settled
on the use of the term intelligent digital assistant, or digital assistant for convenience
and conciseness when the context is reasonably clear.
It’s primary function was contact management with names, phone numbers, addresses,
and notes. A vest-pocket rolodex and notebook, to be used in conjunction with a non-
smart cell phone. No question/answer or task capabilities. Actually, there were a variety
of apps, games, and the like that could be downloaded to the device, but nothing like a
voice or natural language interface for those functions.
Maybe with time the term will be reclaimed as a synonym for intelligent digital
assistant.
In fact, the current web page for Microsoft Cortana uses the term at one point:
Although on a support page for Cortana they use the term digital agent:
Thus illustrating how fluid and unsettled the terminology is for this new product/service
category.
As discussed in the preceding paper, What Is an Assistant?, tasks are relatively simple
operations that may require a lot of effort, but generally do not require much in the way
of complex reasoning, judgment, careful decision, and planning, while goals are more
complex collections of tasks that require some significant level of complex reasoning,
judgment, careful decision, and planning.
Granted, as that paper pointed out, much of the work of many assistants really is simply
task-oriented, but more specialized or capable assistants are capable of goal-oriented
work.
The AI in the current wave of digital assistants has barely enough capability to parse
basic natural language and recognize an interesting but rather limited set of patterns of
meaning, well short of the more complex meaning of the more advanced capabilities of
human assistants.
A goal is specified by stating the objective to be achieved. The objective itself rather
than the details of how to achieve the objective. In fact, and generally, the specific tasks
needed to achieve a goal might not be known in detail in advance and only become
apparent as work towards the goal progresses.
Current digital assistants are generally performing a single operation at a time. Google
do this. Alexa do that. One question or task at a time.
Tasks generally don’t require much deep thought, just slogging through the work.
Goals tend to require deeper, more careful, and more insightful thought. And planning.
Current digital assistants can handle relatively simple tasks, but not more complex tasks
or complex reasoning.
Proactive
Current digital assistants have only limited capabilities at best for being proactive,
doing things for you without being explicitly asked. Reminders and alerts, and learning
from personal data and usage is about the best they can currently muster.
That said, future iterations of digital assistants are likely to become much more
proactive, even to the point of providing us with information and services before we are
even consciously aware that we might want or need them.
This is a close cousin to the technology utilized by the kind of intelligent digital
assistants covered by this paper, but websites are focused more on commercial
customer service types of questions and tasks rather than consumer-oriented questions
and tasks.
That said, a website chat may offer more and deeper insight into narrow niches of your
online life than one of the general purpose intelligent digital assistants.
That’s not common today, but will likely become more common as adoption of
intelligent digital assistants grows, not unlike the fact that many websites also offer apps
for smartphones.
Smart cars
Even before the advent of driverless vehicles, cars in recent years have incorporated
quite a few smart features that automate actions that previously had to be done
manually by the human driver and do involve some degree and sensing and judgment by
the vehicle itself. Whether or not these features constitute intelligence per se is a matter
of debate, but at a minimum they do assist the driver, so in a very real sense they can be
considered a digital assistant, so it’s no great stretch to consider them an intelligent
digital assistant, especially when these features perform the kind of proactive tasks that
even home digital assistants do not currently muster.
Driverless vehicles are too new and unproven to draw any strong conclusions about yet.
In fact, part of the problem is that they still are relatively dumb, limited, and more
focused on heuristics and other weak AI capabilities rather than anything even remotely
resembling strong AI.
But in coming years, AI, smart car features, and driverless vehicles will each be
evolving so that it is not that big a stretch to consider cars of the future to be intelligent
digital assistants. After all, personal transportation is a personal service, traditionally
performed by a human assistant called a driver.
Virtual assistant
One nit is that the term virtual assistant is ambiguous. In addition to referring to one of
these new voice-activated digital assistants, the term also refers to a human assistant
who works remotely, such as from home or for a third-party contractor. That latter
usage is common for job listings.
History
The long history of digital assistants is interesting but beyond the scope of this paper.
In addition, there are likely to be a wide range of enhancements that are based on the
unique capabilities of digital computing which are very different from human
capabilities.
Still, it is likely to be quite some time before digital assistants can surpass human
assistants. But since so many consumers are unable to hire an army of human assistants,
enhanced digital assistants are a promising future even at only a small fraction of
human-level capabilities.
A key gating factor for the evolution of intelligent digital assistants will be the pace of
advances in AI itself, which is discussed at much greater length in my paper Untangling
the Definitions of Artificial Intelligence, Machine Intelligence, and Machine
Learning.
Human in the loop
One prospect not exploited in the current crop of digital assistants is the ability to
integrate the human into the loop, not the user, but a third party, an expert or company
representative who can add value from their human intellect and subject matter
expertise that current digital assistant technology can’t quite muster at this stage.
Put simply, the digital assistant would do the vast bulk of the easy tasks, falling back to
human intervention only for the harder tasks.
Crowdsourcing
Crowdsourcing is another way of putting people in the loop, lots of them rather than
one, to answer more complex or subjective or current questions that a simple lookup or
real-time data reference can’t answer. Or to perform tasks.
Crowdsourcing questions
To the best of my knowledge, there are no digital assistants on the market using
crowdsourcing to respond to questions.
The best you can do today is post a question on a question/answer website such as
Quora or StackExchange and patiently wait for an answer.
Crowdsourcing tasks
There are a variety of crowdsourcing services on the Internet for tasks, but none that are
integrated with the top intelligent digital assistants at this time to the best of my
knowledge, but it’s probably only a matter of time before they start springing up.
There is a skill module that can be added to Amazon Alexa to invoke TaskRabbit, but
the integration seems a bit primitive.
What is needed is for each of the major digital assistants to have generic features for
crowdsourcing tasks that don’t require user knowledge of specific task services.
Crowdsource the crowd sourcing.
Even further, the user should simply be able to state the nature of their task or problem
and the digital assistant should be able to deduce what task is needed. Like:
Group crowdsourcing
Crowdsourcing is generally very open-ended and unlimited — anybody anywhere can
participate, but for some activities there may be a desire or even an advantage to restrict
participation to a smaller or more select group.
Video
The current voice-activated consumer digital assistants don’t incorporate video in their
operation, but that is likely to change in the coming years. They will eventually sense
motion and activity in the room, and eventually be able to recognize objects, people,
and pets and incorporate that information into their actions.
Therapeutic assistants
One promising avenue for the future are specialized digital assistants, therapeutic
assistants, to help people with mental and behavioral problems, to guide them towards
more useful thinking and behavior, and also to monitor them and alert their mental
health professionals or guardians to any problematic symptoms. And also to permit
mental health professionals to guide the actions of the digital assistant as well.
Granted, this area has lots of ethical issues, a virtual minefield. Still, it does have real
potential, to actually help people lead better lives.
A more wide-ranging but more ethically challenging application would be for a plug-in
module for everyday digital assistants to detect when an otherwise normal user might be
exhibiting mental or behavioral symptoms which should be brought to the attention of a
mental health professional. In this scenario there would not be any explicit action by the
user or some health authority to enable such monitoring, although the setup for the
digital assistant might have a simple opt-out or opt-in configuration setting to monitor
for potential health concerns.
Parents with at-risk children could more readily make a decision to explicitly enable
such monitoring for their children or family members or relatives who they worry might
be at risk.
Beyond this paper, these and many related issues are explored in greater depth in a
much larger (but still informal) paper, Untangling the Definitions of Artificial
Intelligence, Machine Intelligence, and Machine Learning. This paper is primarily an
excerpt from that paper, although some new material is presented as well.
1. Human-like intelligent robots. More the realm of science fiction for now.
2. Human-like semi-smart machines. On the near horizon, a few here already.
3. Human-like dumb machines. Stiff, mechanical movements. Very limited
intellectual capacity.
4. Animal-like semi-smart machines. Some examples today. More on the near
horizon.
5. Animal-like dumb machines. Stiff, mechanical movements, although getting
better. Minimal function compared to comparable animals.
6. Non-mammal machines. No intelligence required. Varying degrees of
competence for motion and activity compared to comparable animals. Birds,
insects, reptiles, fish.
7. Industrial robots. Including numerical control (CNC) robotic arms. Precision
motor control, but not the dexterity of true AI robots.
8. Warehouse robots. Great for moving material over fixed, gridlike patterns.
9. Household convenience robots. Niche functions. Precision movement. Precision
sensors. Mostly heuristic rather than true, human-level intelligence.
10. Driverless vehicles.
11. Smart cars. Automation and AI for selected, niche functions, from anti-locking
brakes and cruise control to parallel parking, lane monitoring, and collision
avoidance.
12. Prosthetic limbs. Arms, hands, legs, feet. No higher-order intellectual function,
but precision motor control, sensor feedback, and end effector control (e.g.,
grasping with fingers.)
13. Remotely-piloted vehicles. Including drones. Human is in the loop. Not truly
autonomous.
14. Robotic arms. For amplifying human motions and actions, or working in
physically challenging environments or at a distance. Including surgery. Human
in the loop. Not truly autonomous. But can require fine motor control and
dexterity.
And more. But that’s a fairly comprehensive range. At least it’s representative.
Whether or not a robot possesses intelligence, there is so much more to robotics than
intellectual capacity. Just to mimic the movement capabilities of simple animals or even
insects is a lot of effort, most of it not directly associated with the kind of intellectual
effort associated with the human mind.
Motion, movement, and manipulation of objects in the real world is a real challenge.
They require sophisticated software. They certainly qualify as artificial life, but whether
to consider them AI as well is a matter of debate. The mental aspects are more likely AI
(what do you wish to move, to where, and why), but the physical aspects not so much,
or at least a very gray area.
A key distinction of robotics from traditional, non-robotic AI systems is the fact that the
robotic system is continuously monitoring and reacting to the environment on a real-
time basis.
Much of robotics revolves around sensors and mechanical motions in the real world,
seeming to have very little to do with any intellectual activity per se, so one could
question how much of robotics is really AI.
Alternatively, one could say that sensors, movement, and activity enable acting on
intellectual interests and intentions, thus meriting coverage under the same umbrella as
AI.
In addition, it can be pointed out that a lot of fine motor control requires a distinct level
of processing that is more characteristic of intelligence than mere rote mechanical
movement.
Reader’s choice
In summary, the reader has a choice as to how much of robotics to include under the
umbrella of AI:
In short, it’s not too much of a stretch to include virtually all of robotics under the rubric
of AI — provided there is at least some element of intelligence in the system, although
one may feel free to be more selective in specialized contexts.
The section of the Untangling paper entitled Artificial Life (A-Life) also has some
interesting insight on this matter.
Robotic prosthetics
How to categorize robotic prosthetic limbs is any interesting edge case.
Since a limb, even for a sapient creature such as a human, has no direct role in
intellectual activity it’s quite a stretch to call it AI by itself.
But when a robotic prosthetic limb is attached to a (presumably) human body, it acts as
if it were a real limb.
Since movement, positioning, fine motor control, end effector control (e.g., use of
fingers), touch, pushing, hitting, kicking, grasping, and carrying are activities
peripherally related to activities that result from intellectual activity it is not too much of
a stretch to associate them with that intellectual activity, at least indirectly.
This makes it a fielder’s choice whether you want to categorize robotic prosthetic limbs
as AI per se.
Subsidiary mechanical and electrical activities that are needed to support the intellectual
intentions of the attached body could be considered at least tangential to the intentional
intellectual activity.
For the purposes of this paper, it is reasonably fair to conclude that robotic prosthetic
limbs could qualify as AI. Unless they are particularly simplistic and very limited in
their function and reaction to stimulus.
They could be considered AI at least to the extent they they contain and depend on some
sort of embedded computer chip that participates in the interaction between sensors and
fine motor control.
How intelligent is your robot?
Intelligence of machines is rather limited today, so we haven’t had to fully grapple with
measurement of machine intelligence, but one can usually get a quick sense of how
intelligent an AI system or robot at least appears to be.
The point here is simply that if the robot is grossly lacking any significant sense of
intelligence, then it is questionable whether it could be considered AI.
On the flip side, if the robot has some interesting fraction of human level intelligence, it
would then qualify as AI.
What is intelligence?
Intelligence is discussed at much greater depth in the Untangling the Definitions of
Artificial Intelligence, Machine Intelligence, and Machine Learning paper.
Just to quickly summarize, there are a variety of levels of intelligence and a variety of
elements of intelligence. Elements include:
That full list is way too much to expect from robots at this juncture, but even a very
modest fraction of those elements could still yield a robot that can act reasonably
intelligently, by today’s standards.
Dumb robots
Beyond intelligent and seemingly smart robots and relatively dumb robotic prosthetic
limbs, we also have relatively dumb robots.
For our purposes here, a dumb robot is simply one that does not have any interesting
degree of intelligence or intellectual capacity.
It’s a fielder’s choice whether robots designed to mimic animals should be considered
intelligence since they may or may not indeed possess the ability to mimic animal
intelligence.
For our purposes here, a robot capable of mimicking some significant degree of animal
intelligence could be considered under the category of AI.
For example, if the robot can recognize, interact with, or navigate around objects, they
could indeed be considered a reasonable subset of intelligence.
But if the robot’s movements and actions are strictly pre-programmed, rote, and
mechanical, that would not seem sufficient to warrant being considered AI.
A numerical control (CNC) robotic arm would not seem to qualify as AI by itself.
But if a CNC robotic arm had additional features, such as recognizing different types of
objects or being able to deftly manipulate fragile objects, the door to characterization as
AI would be opened. It would still all depend. It’s a gray area. But absent such
advanced features, a CNC robotic arm would not seem to qualify as AI.
That said, if the same CNC robotic arm could be used in different applications and
deployments as either a dumb, rote, mechanical arm or alternatively as a smart, human-
like hand, then the underlying robotic arm/hand technology could credibly be
considered AI since it is enabling the AI application of the technology.
In short, in general, a dumb robot is not AI per se, but the addition of even a single AI-
related function can indeed suddenly make an otherwise dumb robot or component of
robot technology a candidate for being considered AI.
That said, it isn’t always possible for a mere mortal to judge what is happening under
the hood of a robot. Sometimes even sophisticated intelligence can seem trivial, while at
other times a function that seems fairly intelligent may be based on simple, mechanical,
heuristic technology that doesn’t have any true AI at all.
In short, we are forced to accept the twin propositions that sometimes a dumb machine
will seem smart and sometimes a smart machine will seem dumb.
The reality is simply that if the robot seems smart it will be considered AI, and if the
robot seems dumb it will be considered to not be AI.
A traditional industrial robot or even a modern warehouse robot doesn’t require any
obvious intelligence. Some specific instances may, but not as a general proposition.
But if the tasks require fine motor control, dexterity, handling of fragile objects, or
visual recognition of objects, the line is being crossed from mere automation to
intelligence.
A robotic vacuum cleaner is in the gray zone, not having any dramatic level of
intelligence, but the ability to detect and navigate around obstacles suggests at least a
borderline level of intelligence.
A lot of judgment will still be required and a lot of cases will still amount to a fielder’s
choice whether the robot in question amounts to only mere automation of manual tasks
or amounts to at least minimal intelligence.
Computer vision
Any machine which has a video camera with with software to detect and identify
objects or scenes and autonomously taking action based on what it detects and identifies
is a decent candidate for being classified as AI.
Robotic sonar
Any machine which has a sonar sensor that can be used to detect objects using sound
waves and then take action based on location and distance of objects is a decent
candidate for being classified as AI.
Some applications may be too simple to consider AI, but to the degree that a significant
level of processing is performed using the sonar data the AI label would be more
warranted.
Specialized hardware
Generally speaking, most hardware components would not be considered AI, but an
exception might be appropriate for any components which were specially designed to
facilitate robotic applications, especially when these specialized components fairly
directly facilitate intelligence.
1. Better sensors which provide the robot’s mind with more refined sensory data.
2. Better motors and motor controls, such as facilitating finer motor control for
greater dexterity and handling of fragile objects.
3. Lighter weight and stronger structural materials which permit more intelligent
motion.
4. More efficient or cheaper batteries and motors which facilitate more adventurous
activity.
5. Customized processor, memory, and control chips which enable greater
intellectual activity.
6. Smaller, cheaper, and more efficient processor, memory and control chips that
enable design, construction, and operation of robots that might be too expensive
or unwieldy with off the shelf electronic components.
Most of the questions and points made in this paper about robotics would apply to the
full category of A-Life.
A-Life is generally considered a branch of AI, at least to the extent that concerns about
the role of intelligence are considered.
Generally, the point of A-Life is to support intelligent activity that merely happens to be
artificial in some sense.
For more on A-Life, see the Artificial Life (A-Life) section in the Untangling the
Definitions of Artificial Intelligence, Machine Intelligence, and Machine Learning
paper.
Of course, even a two-legged walking robot with two arms, two hands, and two eyes
can be said to be biologically inspired.
The main thrust of bio-inspired technology is to look at, model, or mimic the muscles,
cells, materials, structures, and control systems of natural biological systems.
And the mind as well, both the human mind and the minds or at least brains of lesser
animals.
And social systems and social behavior of the natural biological world as well.
More properly, bio-inspired technology should be associated with the larger A-Life
category (artificial life), of which robotics (mechanical or electromechanical robots) is
only one part.
Generally, any technology which was specially designed to mimic or model biological
systems of the natural world would get a free pass to be considered under the category
of A-life. Whether it should also be categorized as AI would depend on the degree to
which it is either directly involved with intelligence (intellectual activity) or aids or
facilitates intellectual activity or carrying out actions that resulted from intellectual
activity. And of course there is plenty of room for discretion as to where to draw the
line on how closely related the bio-inspired activity is to any intellectual activity.
Criteria
To be categorized as AI, the robotic technology would need to apply to at least one of:
Generally, the answer is that there needs to be a fairly obvious, direct, or at least not too
indirect connection between the technology in question and at least some sort of
intellectual activity.
Scoring?
Conceptually, one could render a score for how close a given piece of technology comes
to being considered AI.
Basic, off the shelf hardware components, like bolts, batteries, and pieces of metal
would get a score of zero.
A fully-functional, human-like robot including natural language functions comparable
to an intelligent digital assistant would get a score of 100%.
Or even a dog-like robot could score near 100% if its function were close enough to that
of a real dog.
A relatively dumb robotic prosthetic limb might get a score down near 25%, while a
more advanced limb with dexterity and smooth movement that is easily controlled by
the wearer might score 50% or even 75% or higher.
This idea of scoring hasn’t been pursued any deeper, to the best of my knowledge, but
seems promising.
Summary
The mere labeling of a technology as robotic does not immediately inform us as to
whether the technology should be categorized as AI.
In short, if the technology directly results in intelligent behavior, directly informs such
behavior, or fairly directly enables such behavior, then it seems fair to consider it under
the umbrella of AI.
Jack Krupansky
This informal paper attempts to clarify the meanings and usages of the terms related to
knowledge, thought, and reason. Too frequently, public discourse is horribly marred by
really sloppy vocabulary and misuse of terms.
Part of the problem is that traditional definitions are somewhat vague and sloppy, even
in a decent dictionary.
Actual usage has been even sloppier, with a lot of these terms treated and used as
synonyms.
It is not the intention of this paper to give a full treatment of thought and reason, but
only enough to support the full treatment of knowledge, since knowledge, the primary
focus of this paper, is so dependent on thought and reason.
Meaning
Knowledge, by definition, includes meaning, at least basic meaning and various levels
or layers of meaning.
There may be additional levels or layers beyond those included in knowledge per se,
such as subjective meaning that is not shared by everyone who shares the basic,
objective meaning of particular knowledge.
Or layers of subjective meaning that are shared by some individuals or groups but not
by others.
This paper takes the position that all of those layers or levels of meaning are still by
definition part of knowledge, even though they may not be shared by all who possess
that same knowledge. That may feel a little odd, but the simple fact of life is that
knowledge possessed by more than one person is not necessarily exactly 100.000%
identical for each of those persons. Even if they all read the same exact definition, they
may each interpret it slightly differently.
Truth
Truth itself is a very slippery topic. It is indeed touched on in a variety of ways in this
paper, but not in great depth. At best, beliefs, facts, and knowledge seek to approximate
truth, but frequently fall short or even entirely miss the mark. In any case, aspects of
truth are included in this paper to the extent that they hinge on knowledge itself.
It is worth noting that truth and knowledge are not strict synonyms. The nuances are
beyond the scope of this paper. Sometimes ultimate truth is not accessible by even the
best of human intentions. And sometimes interpretations are more important in human
discourse than actual reality.
Reality
Reality refers to all that exists, the physical world, the natural world, life, human life,
human social structures, and any artifacts created by the efforts of humans.
Technically, reality would include human knowledge, but simply as any other physical
artifacts, rather than asserting that the ideas embedded within human knowledge are
real per se.
Generally, the best we can hope for is that our knowledge approximates reality.
Sometimes we can get very close or even occasionally happen to be exactly correct, but
very frequently we are far off base.
1. World 1. The physical world. Reality. Animals and people exists in the physical
world, but merely as objects which happen to move around and their knowledge
exists only as electrical signals and markings on objects, devoid of any meaning
in World 1 pr se.
2. World 2. The mental world. Our models of what we perceive the physical world
to be and how we think about it. Our beliefs, models, theories, feelings, creative
urges, and imagination.
3. World 3. Knowledge and media artifacts which represent our World 2 beliefs,
which we believe correspond to the objective reality of the physical world. Real-
world objects which are products of the human mind.
Relation to intelligence
Knowledge, thought, and reason are inexorably intertwined with intelligence. As such, a
fair amount of terms related to intelligence are included in this paper. But not all terms
related to intelligence will be included here, only those which reasonably intersect with
knowledge, thought, and reason per se.
Relation to logic
Knowledge and reason certainly relate to logic. A modest amount of terms related to
logic are included in this paper, but not all terms related to logic will be included here,
only those which are reasonably necessary to understand and discuss knowledge and
reason per se.
Relation to science
Knowledge, thought, and reason and science are also very intertwined. A modest
amount of terms related to science are included in this paper, but not all terms related to
science will be included here, only those which are reasonably necessary to understand
and discuss knowledge, thought, and reason per se.
Relation to epistemology
Epistemology is the branch of philosophy concerned with the nature of knowledge. The
terms discussed in this paper will certainly overlap with epistemology, but there is no
intention to fully explore epistemology here, just to the extent that it hinges on the
vocabulary of knowledge, thought, and reason.
Relation to metaphysics
Metaphysics is the branch of philosophy concerned with the nature of existence. The
terms discussed in this paper will certainly overlap with metaphysics, but there is no
intention to fully explore metaphysics here, just to the extent that it hinges on the
vocabulary of knowledge, thought, and reason.
Deeper discussion of existence and essence are contained in a companion paper, Model
for Existence and Essence.
Relation to ethics
Ethics is the branch of philosophy concerned with the nature of human nature and social
interaction. The terms discussed in this paper will certainly overlap with ethics, but
there is no intention to fully explore ethics here, just to the extent that it hinges on the
vocabulary of knowledge, thought, and reason.
Relation to communication
Communication is essential to the sharing of knowledge, but the means and methods of
communication are neutral with respect to the actual knowledge and meaning
transmitted and received via communication. A subset of the terms related to
communication will be included in this paper, but only to the extent that they directly
hinge on the knowledge, thought, and reason itself.
Relation to media
Media and communication are tightly related.
Media has only one real purpose, to enable and facilitate communication.
Actually, media has two roles, to communicate knowledge and to record knowledge, but
the recording of knowledge is for the purpose of communicating that knowledge.
Books, papers, audio recordings, and videos simultaneously record and communicate
knowledge, content, and meaning.
This paper does not endeavor to provide a full treatment of media, but simply to cover it
enough to discuss its relationship to knowledge.
Relation to language
Language is clearly essential to sharing of knowledge, but it is meaning that is most
relevant here, not the syntax, grammar, punctuation, or spelling and pronunciation of
words. A subset of the terms related to language will be included in this paper, but only
to the extent that they directly hinge on issues related to meaning and knowledge.
But a full treatment of knowledge representation and knowledge artifacts is well beyond
the scope of this paper. Knowledge representation and knowledge artifacts are treated
here only to the degree needed to provide a full treatment of knowledge.
Domains of truth
There are many distinct domains of knowledge, each having its own distinctive
concepts and vocabulary, even if they all still share the same basic natural language
(English or whatever.) Truth and meaning in one domain don’t necessarily mean the
same thing in different domains.
This matter is discussed and detailed in much greater detail in a companion paper,
Domains of Truth.
Entities
This paper takes an expansive view of entity, treating the concept as referring to
anything that could be referred to in the propositions of knowledge. More than solely
tangible objects, entity is used to refer to intangibles as well, including ideas and
concepts — anything someone might seek to refer to in a statement or proposition.
A partial list of things that would be considered entities in the vocabulary of this paper
include:
• Object
• Inanimate object
• Machine
• Living creature
• Person
• Place
• Thing
• Idea
• Concept
• Thought
• Decision
• Plan
• Topic
• Area
• Event
• Matter
• Action
• Phenomenon
• Situation
• Environment
• Conditions
• Computational entity. Computer software program, object, or data. Robotic or
artificially intelligent computer software.
• Anything of unspecified or even vaguely specified nature that has some sort of
significance.
• A group of closely related entities can also be considered collectively as a larger
entity, such as a family, partnership, team, business, nonprofit organization, or a
country.
• Entities with some characteristics in common can constitute a category or class,
which itself is an entity.
• Qualities, characteristics, attributes, details, and metadata of entities are
themselves entities.
All of that said, the term entity has a more strict meaning for entities which have a
significant sense of independence, in contrast with subsidiary entities which have a
significant meaning only within the context of a larger, umbrella entity.
Short of details
The goal here is to define the vocabulary needing to talk about knowledge, thought, and
reason, but not to detail knowledge about anything other than knowledge, thought, and
reason themselves. As such, the vocabulary stops at the level of the concept of detail, so
that language needed to elaborate detail is excluded, such as:
• Color
• Size
• Height
• Width
• Depth
• Weight
• Texture
• Shape
• Structure
• Purpose
• Function
• Substance
• Emotions
• Intentions
• Gender
• Types of objects
• Types or forms of life
• Details of life
• Specific actions
• Specific activities
• Specific values
• Specific senses
• Specific intentions
• Specific emotions
• Specific feelings
• Specific drives
• Specific attitudes
• Details of logic
• Details of truth
• Structure of the universe.
• Nature of the physical world.
Deeper discussion of existence and essence are contained in a companion paper, Model
for Existence and Essence.
• Specific entities.
• Specific classes or types of entities.
• Specific details or metadata of classes of entities.
• Specific details of specific entities.
• Any domain-specific entities.
• Any domain-specific metadata.
The point of this paper is not to define the details of entities, but to treat such details in
an abstract manner and simply to acknowledge that entities have details which
themselves are referred to as entities and that those details are referred to as entity
metadata in an abstract sense — no detail to be described in this paper.
Many terms defined in this paper will use this term, sapient entity, rather than human
being, person, people, or individual.
People and intelligent robots are sapient (intelligent and wise) as well, but some forms
of perception and basic information about the world only require sentience rather than
full-blown sapience.
Your personal robot could whip out an umbrella for you when it starts raining, but that
requires only sentience rather than sapience.
Work in progress
The analysis described here remains a work in progress. It is as complete as I know at
this time, but it will be enhanced and revised as I become aware of new information.
Some call it “strong” AI, others “real” AI, “true” AI or artificial “general” intelligence
(AGI)… whatever the term (and important nuances), there are few questions of greater
importance than whether we are collectively in the process of developing generalized AI
that can truly think like a human — possibly even at a superhuman intelligence level,
with unpredictable, uncontrollable consequences.
This has been a recurring theme of science fiction for many decades, but given the
dramatic progress of AI over the last few years, the debate has been flaring anew with
particular intensity, with an increasingly vocal stream of media and conversations
warning us that AGI (of the nefarious kind) is coming, and much sooner than we’d
think. Latest example: the new documentary Do you trust this computer?, which
streamed last weekend for free courtesy of Elon Musk, and features a number of
respected AI experts from both academia and industry. The documentary paints an
alarming picture of artificial intelligence, a “new life form” on planet earth that is about
to “wrap its tentacles” around us. There is also an accelerating flow of stories pointing
to an ever scarier aspects of AI, with reports of alternate reality creation (fake celebrity
face generator and deepfakes, with full video generation and speech synthesis being
likely in the near future), the ever-so-spooky Boston Dynamics videos (latest one:
robots cooperating to open a door) and reports about Google’s AI getting “highly
aggressive“
However, as an investor who spends a lot of time in the “trenches” of AI, I have been
experiencing a fair amount of cognitive dissonance on this topic. I interact daily with a
number of AI entrepreneurs (both in my portfolio and outside), and the reality I see is
quite different: it is still very difficult to build an AI product for the real world, even if
you tackle one specific problem, hire great machine learning engineers and raise
millions of dollars of venture capital. Evidently, even “narrow” AI in the wild is
nowhere near working just yet in scenarios where it needs to perform accurately 100%
of the time, as most tragically evidenced by self-driving related recent deaths.
A lot of my blog posts on AI have been about how to build AI applications and
startups. In this post, I look a bit upstream at the world of AI research to try and
understand who’s doing what work, and what may be coming down the pipe from the
AI research labs. In particular, I was privileged to attend an incredible small-group
workshop ahead of the Canonical Computation in Brains and Machines held at NYU a
few weeks ago, which was particularly enlightening and informs some of the content in
this post.
Those are just my notes, destined to anyone in tech and startups generally curious about
AI, as opposed to a technical audience. Certainly a work in progress, and comments are
most welcome.
NIPS, a highly technical conference started in 1987, once a tiny and obscure event, had
8,000 participants in 2017 .
The other major recent trend has been that fundamental AI research has been
increasingly conducted in large Internet companies. The model of the company-
sponsored lab, of course, is not new – think Bell Labs. But it’s taken a new dimension
in AI research recently. Alphabet/Google have both DeepMind (a then startup acquired
in 2014, now a 700-person group focused largely on fundamental AI research, run by
Demis Hassabis) and Google Brain (started in 2011 by Jeff Dean, Greg Corrado and
Andrew Ng, with more focus on applied AI). Facebook has FAIR, headed up by Yann
LeCun, one of the fathers of deep learning. Microsoft has MSR AI. Uber has the Uber
AI Labs, that came out of their acquisition of New York startup Geometric
Intelligence. Alibaba has Alibaba A.I. Labs, Baidu has Baidu Research and Tencent has
the Tencent AI Lab. The list goes on.
Those industry labs have deep resources and routinely pay millions to secure top
researchers. One of the recurring themes in conversations with AI researchers is that, if
it is hard for startups to attract students graduating with a PhD in machine learning, it’s
even harder for academia to retain them.
In addition, AI research, particularly in those industry labs, has access to two key
resources at unprecedented levels: data and computing power.
The ever-increasing amount of data available to train AI has been well documented by
now, and indeed Internet giants like Google and Facebook have a big advantage when it
comes to developing broad horizontal AI solutions. Things are also getting
“interesting” in China where massive pools of data are being aggregated to train AI for
face recognition, with unicorn startups like Megvii (also known as Face++) and
SenseTime as beneficiaries. In 2017, a plan called Xue Liang (“sharp eyes”) was
announced and involved pooling and processing centrally footage from surveillance
cameras (both public and private) across over 50 Chinese cities. There are also rumors
of aggregation of data across the various Chinese Internet giants for purposes of AI
training.
Beyond data, another big shift that could precipitate AGI is a massive acceleration in
computing power, particularly over the last couple of years. This is a result of progress
both in terms of leveraging existing hardware, and building new high performance
hardware specifically for AI, resulting in progress at a faster pace than Moore’s law.
To rewind a bit, the team that won the ImageNet competition in 2012 (the event that
triggered much of the current wave of enthusiasm around AI) used 2 GPUs to train their
network model. This took 5 to 6 days, and was considered the state of the art. In 2017,
Facebook announced that it had been able to train ImageNet in one hour, using 256
GPUs. And a mere months after it did, a Japanese team from Preferred Networks broke
that record, training ImageNet in 15 minutes with 1024 NVIDIA Tesla P100 GPUs.
But this could be a mere warm-up, as the world is now engaged in a race to produce
ever more powerful AI chips and the hardware that surrounds them. In 2017, Google
released the second generation of its Tensor Processing Units (TPUs), which are
designed specifically to speed up machine learning tasks. Each TPU can deliver 180
teraflops of performance (and be used for both inference and training of machine
learning models). Those TPUs can be clustered to produce super-computers – a 1,000
cloud TPU system is available to AI researchers willing to openly share their work.
There is also tremendous activity at the startup level, with heavily-funded emerging
hardware players like Cerebras, Graphcore, Wave Computing, Mythic and Lambda, as
well as Chinese startups Horizon Robotics, Cambricon and DeePhi.
Finally, there’s emerging hardware innovation around quantum computing and optical
computing. While still very early from a research standpoint, both Google and IBM
announced some meaningful progress in their quantum computing efforts, which would
take AI to yet another level of exponential acceleration.
The massive increase in computing power opens the door to training the AI with ever
increasing amounts of data. It also enables AI researchers to run experiments much
faster, accelerating progress and enabling the creation of new algorithms.
One of the key point that folks at OpenAI (Elon Musk’s nonprofit research lab) make is
that AI already surprised us with its power when the algorithm were running on
comparatively modest hardware a mere five years ago – who knows what will happen
with all this computing power? (see this excellent TWiML & AI podcast with Greg
Brockman, CTO of OpenAI)
AI algorithms, old and new
The astounding resurrection of AI that effectively started around the 2012 ImageNet
competition has very much been propelled by deep learning. This statistical technique,
pioneered and perfected by several AI researchers including Geoff Hinton, Yann LeCun
and Yoshua Bengio, involves multiple layers of processing that gradually refine results
(see this 2015 Nature article for an in depth explanation). It is an old technique that
dates back to the 1960s, 1970s and 1980s, but it suddenly showed its power when fed
enough data and computing power.
Deep learning powers just about any exciting AI product from Alexa to uses of AI in
radiology to the “hot dog or not” spoof product from HBO’s Silicon Valley. It has
proven remarkably effective at pattern recognition across a variety of problems – speech
recognition, image classification, object recognition and some language problems.
From an AGI perspective, deep learning has stirred imaginations because it does more
than what it was programmed to do, for example grouping images or words (like “New
York” and “USA”) around ideas, without having been explicitly told there was a
connection between such images or words (like “New York is located in the USA”). AI
researchers themselves don’t always know exactly why deep learning does what it
does.
Interestingly, however, as the rest of the world is starting to widely embrace deep
learning across a number of consumer and enterprise applications, the AI research
world is asking whether it is hitting diminishing returns. Geoff Hinton himself at a
conference in September 2017 questioned back-propagation, the backbone of neural
networks which he helped invent, and suggested starting over, which sent shockwaves
in the AI research world. A January 2018 paper by Gary Marcus presented ten concerns
for deep learning and suggested that “deep learning must be supplemented by other
techniques if we are to reach artificial general intelligence”.
Much of the discussion seems to have focused on “supervised” deep learning – the form
of deep learning that requires being shown large amounts of labeled examples to train
the machine on how to recognize similar patterns.
The AI research community now seems to agree that, if we are to reach AGI, efforts
need to focus more on unsupervised learning – the form of learning where the
machine gets trained without labeled data. There are many variations of unsupervised
learning, including autoencoders, deep belief networks and GANs.
Another related area that has seen considerable acceleration is reinforcement learning
– a technique where the AI teaches itself how to do something by trying again and
again, separating good moves (that lead to rewards) from bad ones, and altering its
approach each time, until it masters the skill. Reinforcement learning is another
technique that goes back as far as the 1950s, and was considered for a long time an
interesting idea that didn’t work well. However, that all changed in late 2013 when
DeepMind, then an independent startup, taught an AI to play 22 Atari 2600 games,
including Space Invaders, at a superhuman level. In 2016, its AlphaGo, an AI trained
with reinforcement learning, beat the South Korean Go master Lee Sedol. Then just a
few months ago in December 2017, AlphaZero, a more generalized and powerful
version of AlphaGo used the same approach to master not just Go, but also chess and
shogi. Without any human guidance other than the game rules, AlphaZero taught itself
how to play chess at a master level in only four hours. Within 24 hours, AlphaZero was
able to defeat all state of the art AI programs in those 3 games (Stockfish, elmo and the
3-day version of AlphaGo).
How close is AlphaZero from AGI? Demis Hassabis, the CEO of DeepMind, called
AlphaZero’s playstyle “alien”, because it would sometime win with completely
counterintuitive moves like sacrifices. Seeing a computer program teach itself the most
complex human games to a world-class level in a mere few hours is an unnerving
experience that would appear close to a form of intelligence. One key counter-argument
in the AI community is that AlphaZero is an impressive exercise in brute
force: AlphaZero was trained via self-play using 5,000 first generation TPUs and 64
second generation TPUs; once trained it ran on a single machine with 4 TPUs. In
reinforcement learning, AI researchers point out that the AI has no idea what it is
actually doing (like playing a game) and is limited to the specific constraints that it was
given (the rules of the game). Here is an interesting blog post disputing whether
AlphaZero is a true scientific breakthrough.
When it comes to AGI, or even the success of machine learning in general, several
researchers have high hopes for transfer learning. Demis Hassabis of DeepMind, for
example, calls transfer learning “the key to general intelligence”. Transfer learning is a
machine learning technique where a model trained on one task is re-purposed on a
second related task. The idea is that with this precedent knowledge learned from the
first task, the AI will perform better, train faster and require less labeled data than a new
neural network trained from scratch on the second related task. Fundamentally, the
hope it that it can help AI be more “general” and hop from task to task and domain to
domain, particularly those where labeled data is less readily available (see a good
overview here)
For transfer learning to lead to AGI, the AI would need to be able to do transfer learning
across increasingly far apart tasks and domains, which would require increasing
abstraction. According to Hassabis “the key to doing transfer learning will be the
acquisition of conceptual knowledge that is abstracted away from perceptual details of
where you learned it from”. We’re not quite there as of yet. Transfer learning has been
mostly challenging to make work – it works well when the tasks are closely related, but
becomes much more complex beyond that. But this is a key area of focus for AI
research. DeepMind made significant progress with its PathNet project (see a good
overview here), a network of neural networks. As another example of interest from the
field, just a few days ago, OpenAI launched a transfer learning contest that measures a
reinforcement learning algorithm’s ability to generalize from previous experience. The
algorithms will be tested against 30 SEGA “old school” video games.
Recursive Cortical Networks (RCN) are yet another promising approach. Developed
by Silicon Valley startup Vicarious, RCN were recently used to solve text-based
CAPTCHAs with a high accuracy rate using significantly less data than its counterparts
much – 300x less in the case of a scene text recognition benchmark, (see Science article,
December 8, 2017)
There are many more methods being contemplated, developed or re-explored in light of
the most recent technological progress, including in no particular order: Geoff Hinton’s
capsule networks or CapNets (approachable explanation involving Kim Kardashian
here), neural attention models (approachable explanation without Kim Kardashian
here), one shot learning, differentiable neural computers (DNC), neuroevolution,
evolutionary strategies,… the list goes on, as further testament to the explosive vitality
of AI research.
Teaching a machine of how to learn like a child is one of the oldest ideas of AI, going
back to Turing and Minsky in the 1950s, but progress is being made as both the field of
artificial intelligence and the field of neuroscience are maturing.
This intersection of AI and neuroscience was very much the theme of the “Canonical
Computation in Brains and Machines” workshop I alluded to earlier. While both fields
are still getting to know each other, it was clear that some of the deepest AI thinkers are
increasingly focused on neuroscience inspired research, including deep learning
godfathers Yann LeCun (video: What are the principles of learning in newborns?) and
Yoshua Bengio (video: Bridging the gap between deep learning and neuroscience).
While MIT just launched in February an initiative called MIT Intelligence Quest to help
“crack the code of intelligence” with a combination of neuroscience, cognitive science,
and computer science, all of this is still very much lab research and will most likely
require significant patience to produce results applicable to the real world and industry.
Conclusion
So, how far are we from AGI? This high level tour shows contradictory trends. On the
one hand, the pace of innovation is dizzying — many of the developments and stories
mentioned in this piece (AlphaZero, new versions of GANs, capsule networks, RCNs
breaking CAPTCHA, Google’s 2nd generation of TPUs, etc.) occurred just in the last
12 months, in fact mostly in the last 6 months. On the other hand, many the AI research
community itself, while actively pursuing AGI, go to great lengths to emphasize how
far we still are – perhaps out of concern that the media hype around AI may lead to
dashed hopes and yet another AI nuclear winter.
Regardless of whether we get to AGI in the near term or not, it is clear that AI is getting
vastly more powerful, and will get even more so as it runs on ever more powerful
computers, which raises legitimate concerns about what would happen if its power was
left in the wrong hands (whether human or artificial). One chilling point that Elon
Musk was making the “Do you trust this computer?” documentary was that AI didn’t
even need to want to be hostile to humans, or even know what humans are, for that
matter. In its relentless quest to complete a task by all means, it could be harmful to
humans just because they happened to be in the way, like a roadkill.
Leaving aside physical harm, progress in AI leads to a whole series of more immediate
dangers that need to be thoroughly thought through – from significant job losses across
large industries (back offices, trucking) to a complete distortion of our sense of reality
(when fake videos and audio can be easily created).
i
The Singularity Is Near: When Humans Transcend Biology is a 2005 non-fiction book about artificial
intelligence and the future of humanity by inventor and futurist Ray Kurzweil.