AI Magazine Volume 27 Number 4 (2006) (© AAAI) Reports

continued this earlier work because he became convinced that advances The Dartmouth College could be made with other approaches using computers. Minsky expressed the concern that too many in AI today try to do what is popular and publish only successes. He argued that AI can never be a science until it publishes what fails as well as what succeeds. Conference: highlighted the im- portance of many related areas of re- search before and after the 1956 sum- The Next mer project that helped to propel AI as a field. The development of improved languages and machines was essential. Fifty Years He offered tribute to many early pio- neering activities such as J. C. R. Lick- leiter developing time-sharing, Nat Rochester designing IBM computers, and working with James Moor . Trenchard More was sent to the summer project for two separate weeks by the University of Rochester. Some of the best notes describing the AI project were taken by More, al- though ironically he admitted that he ■ The Dartmouth College Artificial Intelli- , , and never liked the use of “artificial” or gence Conference: The Next 50 Years Nathaniel Rochester for the 1956 “intelligence” as terms for the field. (AI@50) took place July 13–15, 2006. The event, McCarthy wanted, as he ex- said he went to the conference had three objectives: to cele- plained at AI@50, “to nail the flag to brate the Dartmouth Summer Research summer project hoping to convince the mast.” McCarthy is credited for Project, which occurred in 1956; to as- everyone of the importance of ma- coining the phrase “artificial intelli- sess how far AI has progressed; and to chine learning. He came away know- gence” and solidifying the orientation project where AI is going or should be ing a lot about Turing machines that of the field. It is interesting to specu- going. AI@50 was generously funded by informed future work. the office of the Dean of Faculty and the late whether the field would have Thus, in some respects the 1956 office of the Provost at Dartmouth Col- been any different had it been called summer research project fell short of lege, by DARPA, and by some private “computational intelligence” or any expectations. The participants came at donors. of a number of other possible labels. various times and worked on their Five of the attendees from the orig- own projects, and hence it was not re- inal project attended AI@50 (figure 1). Each gave some recollections. Mc- ally a conference in the usual sense. Carthy acknowledged that the 1956 There was no agreement on a general project did not live up to expectations theory of the field and in particular on Reflections on 1956 in terms of collaboration. The atten- a general theory of learning. The field Dating the beginning of any move- dees did not come at the same time of AI was launched not by agreement ment is difficult, but the Dartmouth and most kept to their own research on methodology or choice of prob- Summer Research Project of 1956 is agenda. McCarthy emphasized that lems or general theory, but by the often taken as the event that initiated nevertheless there were important re- shared vision that computers can be AI as a research discipline. John Mc- search developments at the time, par- made to perform intelligent tasks. Carthy, a mathematics at ticularly , , and This vision was stated boldly in the Dartmouth at the time, had been dis- Herbert Simon’s Information Process- proposal for the 1956 conference: appointed that the papers in Automa- ing Language (IPL) and the Logic The- “The study is to proceed on the basis ta Studies, which he coedited with ory Machine. of the conjecture that every aspect of Claude Shannon, did not say more Marvin Minsky commented that, learning or any other feature of intel- about the possibilities of computers although he had been working on ligence can in principle be so precise- possessing intelligence. Thus, in the neural nets for his dissertation a few ly described that a machine can be proposal written by John McCarthy, years prior to the 1956 project, he dis- made to simulate it.”

Copyright © 2006, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2006 / $2.00 WINTER 2006 87 Reports

Photographer: Joe Mehling Figure 1. Trenchard More, John McCarthy, Marvin Minsky, Oliver Selfridge, and Ray Solomonoff.

Evaluations at 2006 face, and win the DARPA Grand Chal- searchers utilize different methodolo- lenge in a race of 132 miles in the Mo- gies, and there still is no general theo- There were more than three dozen ex- jave Desert. Rus speculated that in the ry of intelligence or learning that cellent presentations and events at future we might have our own person- unites the discipline. AI@50, and there is not space to give al robots as we now have our own per- One of the disagreements that was them the individual treatment each sonal computers, robots that could be debated at AI@50 is whether AI should deserves. Leading researchers reported tailored to help us with the kind of ac- be logic based or probability based. on learning, search, networks, tivities that each of us wants to do. McCarthy continues to be fond of a robotics, vision, reasoning, language, Robot parts might be smart enough to logic-based approach. Ronald Brach- cognition, and game playing.1 These presentations documented self-assemble to become the kind of man argued that a core idea in the pro- significant accomplishments in AI structure we need at a given time. posal for the 1956 project was that “a over the past half century. Consider Much has been accomplished in large part of human thought consists robotics as one example. As Daniela robotics, and much to accomplish of manipulating words according to Rus pointed out, 50 years ago there seems not too far over the horizon. rules of reasoning and rules of conjec- were no robots as we know them. Although AI has enjoyed much suc- ture” and that this key idea has served There were fixed automata for specific cess over the last 50 years, numerous as a common basis for much of AI dur- jobs. Today robots are everywhere. dramatic disagreements remain with- ing the past 50 years. This was the AI They vacuum our homes, explore the in the field. Different research areas revolution or, as McCarthy explained, oceans, travel over the Martian sur- frequently do not collaborate, re- the counter-revolution, as it was an at-

88 AI MAGAZINE Reports

Photographer: Joe Mehling Figure 2. Dartmouth Hall, Where the Original Activities Took Place.

tack on , which had be- Another axis of disagreement, cor- show encouraging signs of solving tra- come the dominant position in psy- related with the logic versus probabil- ditional AI problems though not in chology in the 1950s. ity issue, is the psychology versus terms of human psychology. For in- argued on the con- pragmatic paradigm debate. Pat Lang- stance, machine translation with a rea- trary that the last 50 years has experi- ley, in the spirit of Allen Newell and sonable degree of accuracy between enced the gradual displacement of Herbert Simon, vigorously maintained Arabic and English is now possible brittle logic with probabilistic meth- that AI should return to its psycholog- through statistical methods though ods. Eugene Charniak supported this ical roots if human level AI is to be nobody on the relevant research staff position by explaining how natural achieved. Other AI researchers are speaks Arabic. language processing is now statistical more inclined to explore what suc- Finally, there is the ongoing debate natural language processing. He stated ceeds even if done in nonhuman of how useful neural networks might frankly, “Statistics has taken over nat- ways. Peter Norvig suggested that be in achieving AI. Simon Osindero ural language processing because it searching, particularly given the huge working with dis- works.” repository of data on the web, can cussed more powerful networks. Both

WINTER 2006 89 Reports

Photographer: Joe Mehling Figure 3. Dartmouth Hall Commerative Plaque.

Terry Sejnowski and Rick Granger ex- gent tasks. Perhaps, this vision is all it Minsky thought what is needed for plained how much we have learned takes to unite the field. significant future progress is a few about the brain in the last decade and bright researchers pursuing their own how this information is very sugges- good ideas, not doing what their advi- tive for building computer models of Projections to 2056 sors have done. He lamented that too intelligent activity. Many predictions about the future of few students today are pursuing such These various differences can be tak- AI were given at AI@50. When asked ideas but rather are attracted into en- en as a sign of health in the field. As what AI will be like 50 years from now, trepreneurships or law. More hoped Nils Nilsson put it, there are many the participants from the original con- that machines would always be under routes to the summit. Of course, not ference had diverse positions. Mc- the domination of humans and sug- all of the methods may be fruitful in Carthy offered his view that human- gested that machines were very un- the long run. Since we don’t know level AI is likely but not assured by likely ever to match the imagination which way is best, it is good to have 2056. Selfridge claimed that comput- of humans. Solomonoff predicted on many explored. Despite all the differ- ers will do more planning and will in- the contrary that really smart ma- ences, as in 1956, there is a common corporate feelings and affect by then chines are not that far off. The danger, vision that computers can do intelli- but will not be up to human-level AI. according to him, is political. Today

90 AI MAGAZINE Reports disruptive technologies like comput- ing put a great deal of power, a power that can misused, in the hands of in- dividuals and governments. Ray Kurzweil offered a much more optimistic view about progress and claimed that we can be confident of –capable AI within a quar- ter century—a prediction with which many disagreed. Forecasting techno- logical events is always hazardous. Si- mon once predicted a computer champion within 10 years. He was wrong about the 10 years, but it did happen within 40 years. Thus, given an even longer period, another 50 years, it is fascinating to ponder what AI might accomplish. Sherry Turkle wisely pointed out that the human element is easily over- looked in technological development. Eventually we must relate to such ad- vanced machines if they are devel- oped. The important issue for us may be less about the capabilities of the computers than about our own vul- nerabilities when confronted with very sophisticated artificial intelli- gences. Several dozen graduate and post- doctoral students were sponsored by DARPA to attend AI@50. Our hope is that many will be inspired by what they observed. Perhaps some of those will present their accomplishments at the 100-year celebration of the Dart- mouth Summer Research Project.

Postscript

A plaque honoring the 1956 summer research project has been recently mounted in Dartmouth Hall, the building in which the 1956 summer activities took place (figure 3). For details of AI@50 and announce- ments of products resulting from AI@50 please check the conference Figure 4. The 2006 Conference Logo. website.2

Notes phy and Public Ethics philosophy of , philosophy of sci- 1. For more details on the speakers and top- (CAPPE) at the Aus- ence, and logic. He is the editor of the jour- ics, check www.dartmouth.edu/~ai50/. tralian National Univer- nal and Machines and is the presi- 2. www.dartmouth.edu/~ai50/. sity. He earned his PhD dent of the International Society for Ethics in history and philoso- and Information Technology (INSEIT). He phy of science at Indi- is a recipient of the American Computing James Moor is a professor of philosophy at ana University. He pub- Machinery SIGCAS “Making a Difference” Dartmouth College. He is an adjunct pro- lishes on philosophy of award. fessor with The Centre for Applied Philoso- artificial intelligence, computer ethics,

WINTER 2006 91