History of AI
It is hard to generalise about the complex history of AI,
which goes back to the 1950s.
We will attempt some broad overviews.
Classic AI
One statement that is probably true about the early days of AI,
what is sometimes called "classic AI",
is that it concentrated on
adult human-level intelligent skills,
not adult animal skills
or child human skills.
Classic AI concentrated on:
- Human-level intelligence. Abstract reasoning.
- Linguistic intelligence.
- Symbol using and knowledge representation.
- Rule-following and logic.
Archetypal problem - Chess
- AI and chess
- Making a program to beat humans at the game of chess became an early goal for AI.
- Later we will consider the long history of
AI and games
- Games are popular because a metric exists
to prove one algorithm is better than another.
Chess on Ancient Brain
"Ancient Brain" is an online coding platform
we will use for this course.
Here is an initial demo of an AI program on Ancient Brain.
A chess "World" has a built-in chess player.
You can plug in different other programs to play against it. These are called "Minds".
Click to run Mind:
Cloned Simple AI - Depth 1
in World:
Chess : Mind vs Simple AI (a bit of random) with more minds
at
Ancient Brain.
Chess gets solved
After 40 years, a milestone:
AI beat best human at chess in 1997.
The "Chess is easy" argument
- An interesting argument developed since (and even before) 1997.
Which is:
Chess is the kind of thing that
is easy for computers.
Tennis is hard.
Or football.
Walking/running over rough ground is hard
(which adult humans, children and animals can all do).
- Later in reading we will extend this argument.
Chess impresses us because we can't do it.
Walking doesn't impress because we can all do it.
Doesn't mean walking is easy.
Rather it means all humans have solved a hard problem.
The history of AI: A repeated cycle of excitement and disillusion
Slower than expected progress in classic AI led to
what might be called a
"biologically-inspired" movement in AI research,
with new algorithms and new approaches,
which really took off in the 1980s.
The
"biologically-inspired" movement
itself made
slower than expected progress.
It fell into some disillusion around 2000.
It has been revived again recently with "deep learning".
There is a new cycle of excitement.
Conclusion: AI is hard!
Don't get too excited too soon.
Biologically-inspired AI
What might be called "Biologically-inspired AI"
took off in 1980s.
Researchers taking a new look at
why AI is so hard.
And in particular:
How does nature do it?
After all, nature solved all the AI problems before we were born.
Themes of Biologically-inspired AI
Idea in Biologically-inspired AI |
As opposed to |
Action Taker and Evaluator |
Knows in advance what to do |
Learning, Evolution, Self-organisation |
Design, Search |
Robustness, Duplication, Multi-minds |
Brittleness, single thread |
Non-Symbolic AI, Sub-Symbolic AI |
Symbolic AI |
Sensorimotor skills |
Higher-level cognitive skills |
(Whole) Animals |
(Parts of) Humans |
Situated (in stream of IO) |
Isolated |
A-life, A-animals, bacteria, insects, dogs, chimps, H.erectus and infants |
Adult H.sapiens |
Symbol-grounding, origin of representations |
Modern human language |
-
"Intelligence without Reason",
Rodney Brooks,
Proceedings of the 12th International Joint Conference on Artificial Intelligence (IJCAI-91),
1991.
A lengthy history of AI from the biologically-inspired angle.
-
"Today the earwig, tomorrow man?",
David Kirsh,
Artificial Intelligence 47 (1991) 161-184.
A reply to Brooks.
- "Why not the whole iguana?",
Daniel C. Dennett,
Behavioral and Brain Sciences 1:103-104,
1978.
A classic call to build whole creatures.
- "The animat path to AI",
Stewart Wilson,
Proceedings of the First International Conference on Simulation of Adaptive Behavior (SAB-90),
1990.
Dennett's call taken up by the bio-AI movement.
-
Out of Control: The New Biology of Machines,
Kevin Kelly, 1994.
- Online.
A mind-bending description of the type of machines we're
trying to build.
-
"AI is possible .. but AI won't happen: The future of Artificial Intelligence",
Mark Humphrys,
1997.
A short talk I gave of the history of AI,
emphasising the biologically-inspired trends.
Analysis v. Synthesis
We said AI is part of
engineering, but can also contribute to
science.
Many scientists who study human minds and animal minds are
interested in AI.
Consider this:
- Science is about analysis.
In neuroscience, we examine an existing brain in nature and try to
"reverse engineer" it, with no manual.
This is very difficult.
We are like Stone Age man looking at an iPhone and trying to see how it works,
with no manual and no science books.
- Engineering is about synthesis.
We try out models of mind by building them!
We know what the component parts are.
We can label them.
It is hopefully much easier to study a system we built.
So
AIs may not be exact models of natural brains, but at least they are comprehensible.
Or are they?
In fact, the more you allow machines learn and self-adapt and evolve,
the more you run into analysis problems of the final result.
Just like nature again!
Here is the result of a program that
evolves program code.
Synthesised - and yet hard to understand.
From
"Have Your Spaghetti and Eat it Too:
Evolutionary Algorithmics and Post-Evolutionary Analysis",
Kfir Wolfson et al.,
Genetic Programming and Evolvable Machines,
Volume 12 Issue 2, June 2011.
Discussion: What is the hardest part of AI?
What do
you think the hardest part of AI is?
Why do
you think AI has not been solved?
Discuss in comments.
Biologically-inspired AI was harder than expected
The
"biologically-inspired" movement
itself made
slower than expected progress.
It fell into some disillusion around 2000.
Why is AI so hard?
A lot of people talked about scaling up,
and "Not Enough Stuff" arguments.
Deep learning
Some parts of the "Not Enough Stuff" arguments
began to come true after 2010 or so.
Notably
not enough data, and not big enough networks.
It turned out that scaling up massive neural networks with massive quantities of data
really did make a difference.
Since 2010 or so, there has been
a new period of excitement and progress,
with the rise of Deep learning
and Generative AI.
The revival is
due to a number of causes,
including new theoretical breakthroughs,
more powerful small parallel machines,
and companies willing to spend on massive numbers of parallel machines for large-scale training experiments.
Click to run World:
Chat with GPT model at
Ancient Brain.
This is a demo of using JavaScript to call the
OpenAI
API to
talk to
GPT-3.5 (2022).
This is the model that
ChatGPT uses.
View the code to see how it works.
Clone and edit the World to modify it.