\magnification=\magstep1
\parindent 0pt
\parskip 6truept

The Dartmouth Conference

The Dartmouth Summer Research Project on Artificial Intelligence is often
cited as the beginning of AI. I want to give my opinions on the extent to which
it was and wasn't. This opinion is based on my recollections and has not 
involved interviewing other people or looking at documents. A historian is 
required for that.

First some prehistory. Thoughts about machines thinking were old. Here are some
threads.

1. Religion. The gods made man according to religions, and this gave rise to ideas
of proposed man-made life, but there were no serious ideas about how to do it.

2. Science fiction. Robots made good stories, but don't require ideas about how 
to do it as contrasted with ideas about its effects.

3. Enlightment philosophy and empirical philosophy. We have {\it L'Homme Machine\/}
by La Mettrie asserting that humans are material systems operating according
to the interactions of their parts according to physical laws. Vague ideas about
associations and primary and secondary qualities arose. Leibniz aspired to a
mathematical logic applicable to solving non-mathematical human problems by
calculation, but logic took a long time to develop, and its applicability
to solving human problems is taking longer. Babbage and Lovelace had some
AI ideas but Bubbage's failure to get his computer built, inhibited
carrying them further.

The advent of electronics and control systems around World Was II inspired many
more people to think about intelligent machines. Many approaches were explored 
some of which are pursued today -- whether continued or repeatedly invented.

Using stored program computers for AI has first proposed in a paper by Turing in 
1950, but it seems to me that the first continuous work was began by Newell and
Simon in 1954 and that Newell merits being called the first computer-based
AI professional. Other people had written papers, but AI wasn't their full time
activity. I became interested in AI in 1948 while a Caltech graduate student
in mathematics, but I pursued it only part time and didn't think much about
using computers until 1955. Minsky wrote his 1954 Ph.D. thesis on AI but not
about using computers.

I came to Dartmouth in February 1955 as a mathematician and became converted
to AI as a full time activity based on computers during a summer job at
IBM's Poughkeepsie Laboratory. IBM wasn't interested in AI but tolerated it--
most prominently tolerating Arthur Samuel's work on checkers. (It became 
intolerant of AI in 1959 and stayed intolerant until about 1983). Nathaniel
Rochester was head of the Information Research Department there, after having
played an important role in designing the IBM 70 computer, IBM's first
stored program machine, and he was or became an AI enthusiast, which 
probably adversely affected his career in IBM after 1959.

The idea of summer study on AI was mine, and it was based on several premisses.

1. I knew of a number of scientists interested in intelligent machines from a
variety of points of view. These included Minsky, Claude Shannon, Oliver
Selfridge, Nathaniel Rochester, John von Newmann and Donald MacKay in
England. Minsky, Rochester, I don't recall if I had yet heard of Newell, and
Simon, Shannon and (I believe) Selfridge agreed to put their names with mine on 
a proposal. I think it was Shannon who suggested the Rockefeller Foundation.

2. I had heard of a summer study on some defense topic conducted by the Rand 
Corporation that was judged successful. My idea was that if several
people worked full time for a summer on AI, we might make significant progress.

3. Minsky and I had spent the summer of 1952 at Bell Laboratories in Shannon's
group and had worked on our various approaches to machine intelligence.
Subsequently, Shannon and I decided to solicit papers relevant to the subject
and publish them as an Annals of Mathematics Study which appeared in 1956 as
Automata Studies. Shannon wanted a modest name for the collection so as not to
seem arrogant (any phrase suggesting machine intelligence was out).

Many of the papers that arrived seemed pedestrian to me, so when the time came
to choose a name of the summer study I wanted to nail the flag to the mast.
After some thought I chose {\it artificial intelligence\/}. I vaguely recollect that
I'd seen or heard the name before, but no-one has found a reference.

We asked the Rockefeller Foundation for \$15,000, and they gave us \$7,500, but I
don't believe the \$7,500 was all spent. The problem was that they didn't give us 
the money till December (actually a remarkably fast response), and most people
already had rather firm summer plans. With hindsight we should have tried for
summer 1957.

In the event, some invitees didn't come at all and almost all came for short
periods, not all overlapping. There was no real joint work. Those people already
working on AI mainly continued their previous approaches, and those who
mainly worked on other problems didn't start on AI.

The only people who had results from computer programs were the Newell-Simon-
Shaw team reporting on the Logic Theory Machine and Samuel on learning in checkers.
Minsky and I had theories.

There was no report on the summer, though numerous documents were exchanged.
Here are the results that I can recall. Minsky proposed geometry as a domain
where looking at a figure could guide logical reasoning. Rochester decided to
take it up, and a new Ph.D. in physics, Herbert Gelernter, who started
work in September 1956 took up the project and produced a geometry theorem prover 
by 1959, IBM's best AI work so far.

Also I took the idea of list processing from Newell, Simon and Shaw and decided
to combine it with the Fortran idea of a programming language based on assignment
statements. As a consultant to Gelernter's project I proposed that they 
develop a Fortran based list processing language. They did this, and added at 
least one essential new idea -- that {\it cons\/} 
be a function. When I later invented
conditional expressions and then their recursive use to define functions, it
became clear that Fortran was inadequate and Lisp started in Fall 1958.

It would be interesting to know if any other direct scientific influences came
from the Dartmouth project.

In spite of this meager direct scientific result, though perhaps no more meager
than that of many conferences, I believe it played an important role in the
future of AI. First it named the subject and nailed its flag to the mast.
It focused attention of the engineering community at an IRE meeting in 
September 1956. When in Spring 1958, Minsky and I told Jerry Wiesner that we
wanted to start an Artificial Intelligence Laboratory of Electronics we were
received enthusiastically rather than with a blank stare. Newell and Simon
were already established and continued to use the term ``complex information
processing'' instead of AI for some years, but eventually the importance of 
expressing the long term goal in the name of the field became generally
accepted. Those organizations that held out the longest against the name also
didn't contribute much to the subject.

\end