DARTMOUTH AND BEYOND John McCarthy, Stanford University The - - PDF document

DARTMOUTH AND BEYOND John McCarthy, Stanford University

• The symbolic role of the Dartmouth Summer Wo

Artificial Intelligence in establishing AI as a field of res more important than the specific results obtained at th I didn’t expect this.

• The most important results presented at the mee

those of Newell, Simon, and Shaw. Their work, which done previously, included the list processing system IP use to program their Logic Theory Machine which cor to protocols of subjects given the task of proving logica as pure symbolic manipulation, i.e. with no explanatio

• Alex Bernstein of IBM reported on his design for a

chess playing program. My discovery of the alpha-bet for chess.

• The results obtained at the meeting included Mins

for a geometry theorem prover that only tried to prove true in a diagram, Solomonoff’s work on algorithmic c My work on logical AI only started two years later.

PREHISTORY OF THE DARTMOUTH WORKS

• The four organizers of the 1956 Dartmouth Worksho

ficial intelligence were John McCarthy, Marvin Minsky, Rochester, and Claude Shannon.

• My own interest in AI was triggered by attending th

ber 1948 Hixon Symposium held at Caltech. At this sy the computer and the brain were compared, the comp ing rather theoretical, since there wern’t any stored pro computers yet. The idea of intelligent computer prog in the proceedings, though maybe it was discussed. ready had the idea in 1947. I developed some ideas ab ligent finite automata but found them unsatisfactory

• publish. My Princeton PhD thesis was about differen

tions.

• Marvin Minsky was independently interested in AI an

while a senior at Harvard built, along with Dean Edmu a simple neural net learning machine. At Princeto pursued his interest in AI, and his 1954 PhD thesis e the criterion for a neuron to be a universal computing

• Claude Shannon proposed a chess program in 1950,

small relay machines exhibiting some features of intellig being a mouse searching a maze to find a goal.

• In the summer of 1952 Shannon supported Minsk

at Bell Telephone Laboratories. The result of my eff

paper on the inversion of functions defined by Turing I was unsatisfied with this approach to AI also, becaus permit the direct expression of facts about the world.

• Also in 1952 Shannon and I invited a number of r

to contribute to a volume entitled Automata Studies t came out in 1956.

• In the summer of 1952 Shannon supported Minsk

at Bell Telephone Laboratories. The result of my eff paper on the inversion of functions defined by Turing I was unsatisfied with this approach to AI also.

• Also in 1952 Shannon and I invited a number of r

to contribute to a volume entitled Automata Studies t came out in 1956.

• I came to Dartmouth College in 1954 and was

Nathaniel Rochester of IBM to spend the summer of 1 Information Research Department in Poughkeepsie, NY had been the designer of the IBM 701 computer. Roc came interested in AI and his department sponsored work until IBM had a fit of stupidity in 1959.

• That summer Minsky, Rochester, Shannon, and I pro

Dartmouth workshop. The proposal to the Rockefelle tion was written in August 1955, and is the source o artificial intelligence. The term was chosen to nail t the mast, because I (at least) was disappointed at h the papers in Automata Studies dealt with making behave intelligently. We wanted to focus the attent participants.

• The original idea of the proposal was that the p

would spend two months at Dartmouth working colle AI, and we hoped would make substantial advances.

• It didn’t work that way for three reasons. First the R

Foundation only gave us half the money we asked for the participants all had their own research agendas an much deflected from them. Therefore, the participant Dartmouth at varied times and for varying lengths of main reason is that AI presents many more difficulties known in 1956.

WHAT HAPPENED AT DARTMOUTH?

• Newell and Simon were the stars—with list processin

logic theory machine.

• Minsky’s diagram based geometry theorem proving id

got Herbert Gelernter to do it, but IBM had a fit of s 1959 and lost its advantage in AI.

• Solomonoff’s start on algorithmic complexity.
• Alex Bernstein’s chess program.

My alpha-beta he chess-like games.

• My own ideas on logical AI came two years later.

A SAMPLE OF WHAT AI HAS ACCOMPLISH

• As in any scientific field, we need to distinguish basi

from applications. At present too large a fraction of t going into applications.

• Basic AI research accomplishments include succes

as a drosophila for AI (and by way of contrast, non- go, formalisms for action including situation calculus calculus, non-monotonic programming systems like M ner and Prolog, theory of non-monotonic reasoning, circumscription and logic of defaults. The surprising propositional satisfiability programs and their widespr

• cations. The Causal Calculator.

• Applied results include many classification application

computer vision, and driving a vehicle.

• There quite a few accomplishments I would include if

I almost forgot propositional satisfiability. Apologies to looked.

WHEN WILL WE HAVE HUMAN-LEVEL AI?—Ku blunder.

• This is the wrong question.

We can’t yet extrapo present AI to human-level in 2029 or any other fixed

• The right question. We will reach human-level AI w
• ne solves some basic problems. Maybe five years—m
• years. The genetic code came 100 years after Mende
• What problems?
• How to do nonmonotonic reasoning in general.

F entities that don’t have if-and-only-if definitions. S

self-aware systems. I know more, but most likely there lems no-one knows about yet.

• Three classical problems of AI are the frame proble

to avoid specifying what doesn’t change when an eve the qualification problem of avoiding specifying ever qualification for an action to be successful, and the ra problem of avoiding specifying all the side effects of an three have been solved in important contexts and for applications, but I think none have been solved at t leve of intelligence.

• These ideas require extensions to logic, but much

has been achieved by restrictions. We need logical sys can reason about their own methods.

THREE OF THE PROBLEMS BETWEEN US HUMAN-LEVEL AI?

• Non-monotonic reasoning. If I hire you to build me a

you must presume the bird can fly. But if you then lea is a penguin, you can no longer infer that. Non-mono has formalized these, e.g. with circumscription, and ap to cases like birds, but a general way of doing non-mo come.

• Logical treatment of partly defined objects.

Exam snow and rock that make up Mount Everest, the wa U.S. Idea: An object ill-defined in general, may have

• nly-if definition in a particular context.

Start from

• contexts. Example: mother to a small child.

• Self awareness.

Example: How do you know you whether George Bush is sitting or standing at this mom should a robot know?

• I know several more problems AI has to solve, e.g.
• contexts. There are probably several important proble

knows about yet. How long will it be before they fied and solved—or turn out to be automatic consequ general method?

• These ideas are elaborated in articles reprinted on m

http://www-formal.stanford.edu/jmc/.

HOW LONG TO HUMAN LEVEL?

• Maybe five years.

Maybe 500 years. If your int dates, begin by speculating about when general non-m reasoning will be understood.

• Human-level AI is more likely to result from indivi

research, theoretical and experimental, than from res grammes recommended by commmitees. None of my

• n a topic suggested by a committee.

A MODEST PROPOSAL

• This proposal is based on the proposition that new

are probably necessary for human-level AI, and these are more likely to come from individuals rather than pr proposed by committees.

• The proposal is for a programme of individual fellow
• Each fellowship supports the individual for five years
• The programme is aimed at graduate students or new
• The recipient can move among institutions.

• Except in unusual circumstances no intermediate pr

ports are required.

• The programme could be run by NSF or DARPA

with the $7 million that AAAI has left.