Transcribers: Chih-yu Chao & Varun Aggarwal
* The history part
The symbolic role of the Dartmouth summer workshop on AI in establishing AI as a field of research was more important than the specific results obtained at the meeting. I didn't expect this.
The most important results presented at the meeting were those of Newell, Simon, and Shaw...
Alex Bernstein of IBM reported on his design for a full game chess playing program. My discovery of the alpha-beta heuristic for chess -- there are aspects of our own behavior that we fail to observe
The results obtained at the meeting included Minsky's ideas for a geometry theorem prover that only tried to prove sentences true in a diagram. Solomonoff's work on algorithmic complexity. My work on logical AI only started two years later.
* Prehistory of the Dartmouth workshop
The four organizers of the 1956 Dartmouth workshop on AI were John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon
My own interest in AI was triggered by attending the September 1948 Hixon Symposium held at Caltech. At this symposium, the computer and the brain were compared...
Marvin Minsky was independently interested in AI and in 1950 while a senior at Harvard, along with Dean Edmonds, built a simple neural net learning machine...
Also in 1952 Shannon and I invited a number of researchers to contribute to a volume entitle Automata Studies that finally came out in 1956
I came to Dartmouth in 1954 and Rochester became interested in AI and his department sponsored important work until IBM had a fit of stupidity in 1959.
Minsky, Rochester, Shannon, and I proposed the Dartmouth workshop. The proposal to the Rockefeller Foundation was written in August 1955 and is the source of the term AI.
The original idea of the proposal was that the participants would spend two months at Dartmouth working collectively on AI, and we hoped would make substantial advances.
It didn't work that way. The RF only gave us half the money we asked for. The participants all had their own research agendas and were not much deflected from them.
* What happened at Dartmouth?
Newell and Simon were the starts with list processing and the logic theory machine
Minsky's diagram based geometry theorem proving idea...
* A sample of what AI has accomplished
We need to distinguish basic research from applications. AI present too large a fraction of the work that is going into applications.
It gave a lot of people the wrong idea that if you have a lot of junk, the knowledge will organize by itself
Flush the word "agent"
Difference between "agent" and "agency" (assembled agents can do jobs), switch to the word "resource"
Disappointed in the course of AI after 1980
Daniel Bobrow's gadget solved high school algebra word problems -> Dan knows which word to type
Gerry Sussman' program, haven't see anything like that
Semantics disappear under Chomsky. If we know about grammar, we don't have to figure out how they learned it
Interesting thing: sequence of tense to understand children's story, give a new story it will have to start from scratch Eugene Charniak analyzed the story -- "she wonders if he will like a kite"
World spinning around getting good results for the reporters instead of building actual robots
Good for high school students to build robots, but not reflection
Get more people working on reasoning, commonsense knowledge and neural networks
Why do students work on things that are popular?
Scientists should go into the popular level
Lucky to meet John and Marvin in the Dartmouth workshop 50 years ago
Inspired Shannon with information theory
1948, at Berkeley -- a machine that would be able to think
Support from MIT
Bob Kahn's support
Checkers machines, chess at IBM
Orthogonality -- not so much as being orthogonal.
Efforts in learning not well published
It is not just to checkmate to the other person
We want to do many things in mind
Try to change
Software specifying what various modules should try to do
How blind people make models of the world
Looking at the nature of mankind
We are doing basic research needed
Even if we different widely, we try to achieve the goals of AI
Quote from Shakespeare (Saint Crispin's Day Speech)
The pattern of my tie is a core memory
Celebrate a history of innovation, scientific excellence
Should have started our anniversary 10 years ago, the decade of cybernetic systems science
From physical system to social systems
The concept of bits as a unit of information
Cybernetic and math model of psychology were picked
Minsky built the first neural network machine
There are precursors of the Dartmouth workshop
"The world will know little about what we say here." We now know that Lincoln is wrong, so please pay attention.
Interdisciplinary nature of this revolution -- psychology, EE, linguistics, statistics
Two very important people - Newell and Simon
The power of interdisciplinary knowledge
Their contributions
Bored with thermodynamics -- "thinking machine"
One of the most important social sciences
MIT Lincoln Lab played a big role the history California: iSystems Laboratory
Fortran
IPL
general theory: generate subgoals in the search, means analysis
Where has NL gone?
What's exciting 50 years ago; what's exciting now?
Three events:
What you might be able to do with these machines
You believe it wouldn't work unless it has common sense
Late 60's, NLP, especially machine translation, a lot of non-AI part
Direct attack on the derivation of meaning
Doing things in NL is the key thing, way to think about KR in general
Can't get far without solid syntax and dictionaries
"Computational dictionary"
Have to know about the specific meaning of the word
AI, NLP, and their relationship
It helps to reduce the ambiguity that each produces
How to chose the alternatives? relative frequency
Web - source
NLP: combine statistical processing and symbolic processing
Have a deep, rich world model
Simulate the details of what humans do
In using web as a source, we are capturing the outcome of many people's collective behavior -- call it intelligence?
We are doing things beyond the scope of an individual
Then: model what individual might to based on collective behavior
Now: it's not necessary beyond the scope of an individual
Basic AI research accomplishments include success in chess as a drosophila for AI
* When will we have human-level AI? Kurzweil's blunder
Three classical problems in AI: frame problem, qualification problem, ramification problem
Non-monotonic reasoning
Logical treatment of partly defined objects
Human-level AI is more likely to result from individual basic research theoretical and experimental, than from research programs recommended by committees. None of my work was on a topic suggest by a committee.
Concrete proposal: some agency sponsor a program of fellowships given to young researchers/students to the individual, 1 out of 20 will break through.
a1: (to John) inviting alpha-beta in 1956, most advance in game playing. Why no one thought about writing a paper until 1963?
John McCarthy: Newell & Simon had this in their publication. Theorem of alpha beta - gives the same result as min-max, but takes less time. My version has optimistic and pessimistic evaluation.
a2: (Thanked founders, reflected life at MIT...)
a3: All criticisms of AI -- any advise on what criticisms we should take seriously?
Oliver Selfridge: Ignoring the criticisms is the decent way.
Karen Sparck-Jones: A lot of pressure from national security (US), industrial relevant (UK)
what's the motivation for it
Nils Nilsson: Motivation is important. The fundamental assumption is incompatible
a4: Proposal offered as criticism? Too many researchers work in a small field. We should understand the general science of information processing. I suspect if this is true.
Oliver Selfridge: The intelligence of people is important in the society. We are social creatures and use it that way. Take into account: we adapt that in continuing way. AI is the same but under different basis. We learn how to talk to the system, the system learns how to talk to us.
a5: Grand vision of AI, anxiety of evolution... for many people in the scientific world, the grand idea of AI is deeply threatening. Nothing mechanical can possibly dignify as a soul that can be loved and respected. Moral fears about AI...
Nils Nilsson: There are continuous worries in the past. Each obstacle will be overcome. Progress will continue, try to help people understand it. It's difficult to help
Karen Sparck-Jones: People fear because the machines have the capabilities we have. Machines might be different from us but can still be very powerful. Unless we have observations that machines have the capabilities we have...
a6: These are separate fears. People should be worried, apprehensive.
John McCarthy: Should think about what people really worry about more precisely. The tendency is to spend a lot of worries on hypothetical things. Many of them is to postpone any action. 1970 world leaders should get together to decide AI. It's still premature... The worriers should be encouraged to worry about something else.
a7: Decades before 1956, discontinuity in symbolic processing... Whether it was overemphasized at that time. To build an intelligent agent in a dynamic world, we need all the sciences together.
Edward Feigenbaum: I don't sense there was a discountinuity (when you live through it you don't sense it). We did field differentiation a little bit, so we didn't collaborate as much, but you can't make a strong case about symbolic processing deliberately creating such a situation.
John McCarthy: Things were diverse before remain diverse afterwards. Avoid using the word "cybernetics".
Oliver Selfridge: They ignore that we or the problems never existed, which is a terrible tragedy.
Karen Sparck-Jones: 1956 was the Dartmouth conference, as well as the second conference on machine translation (the first one was in 1952). If you look at some of the papers that contributed to these conferences, they were more heterogenous. When people actually got to do things on a real computer? It was very difficult to write paper description for this process.
Edward Feigenbaum: What you call a thing matters. Newell & Simon call their work CIP (complex info processing), cut across human and artificial intelligence. The AI community spin off and do it whenever they want to call themselves something different.
a8: (follow up) NL was the center of AI. Should we try to bring things back again? Yes, but any ideas?
Oliver Selfridge: We should. Some people should look at the central issue of the new challenges that arise.
Edward Feigenbaum: It's a good idea to head back in the direction. We haven't learned enough about intelligence. AI people as a whole are better at discovery things and inventing them. We have achieved a great deal areas of cognition, but in other areas we are not even close to human capability yet. If the insight comes from looking at the brain, good, but we need to go back to study that machine.
Nils Nilsson: What age do you start with?
Edward Feigenbaum: College sophomore
a9: Turing test as a public source of worries. Division of the fields... is it that the splintered field should be in the inventory or should be left out?
Karen Sparck-Jones: Perhaps the question should be: wouldn't it be surprising if wasn't splintered? If you look at every other subject it's splintered. It's the worry that whether we splintered before we got enough common knowledge.
John McCarthy: I wish someone would say something technical.
a10: Post a challenge -- engineering field. We have to simulate somehow. It's a compositional problem -- many facets. In order to progress, we have to build up an infrastructure to support investigation in research. AI haven't created such a infrastructure. Are we going to progress without the infrastructure?
Edward Feigenbaum: A large KB. You are stuck because you have to build things first. If we have a big KB, it will help he newbies going fast.
Oliver Selfridge: It's not the infrastructure. Not until we get the system to handle central topic... when it "wants" to do something. Children don't have infrastructure, but a set of desire of hope, motivations, etc.
Edward Feigenbaum: We recruit people of highly math background. Before the experiment, we need knowledge of instrumentation; to provide them with the knowledge base is a good thing.
John McCarthy: It might be a fundamental mistake. There are tools we can use. I'm not sure I want to pursue knowledge base for life. It's something you want to interact with.
a11: Whether what we study consciously will solve the AI problems? Given the things we might consider at any given moment, how do we narrow the world down to a small number we can track?
John McCarthy: I don't find it coherent.
a11: How do we narrow down things to a tractable number? We typically narrow them down to 1.
John McCarthy: For a particular problem, only a small number of concepts are involved. I don't think about what's going on at the beach. Humans have no problem with that. Computer programs using logical sentences need not have special problems in keeping less to the point... computer is a device for communicating with humans. Logic is reasonably well adapted to this. Use sufficient set of predicates and functions.
a12: When the field involves invention, what would make substantial progress if with enough money?
Oliver Selfridge: Basic research... DARPA being restricted by that. 2 kinds of reference... how to use logic...
the central ones... teaching, learning, and education are the ones we don't want to get put as rewardable problems
a13: (the Turing paper referred in the McCarthy slides) To what degree the influence of any of those ideas in the paper went on during the Dartmouth conference?
John McCarthy: One of his earlier paper is more important. Programming computers is the key thing. Turing is the first person to say that...
Transcribers: Chih-yu Chao & Varun Aggarwal