As the world grapples with a flood of Large Language Models, some computer programmers have gone old school, reaching back to where it all began by reanimating the first chatbot using its original code. This comes sixty years after Joseph Weizenbaum at MIT created ELIZA, named after the character Eliza Doolittle from Pygmalion and generally regarded as the original operating chatbot.
The idea of a computer that could talk to you as if it was a person had been around before, for example with Alan Turing’s proposal of a machine that could fool a person into thinking it was human, now known as the Turing Test. The authors of a preprint – which has not yet completed peer review – note that a century earlier, Ada Lovelace had suggested Charles Babbage’s Analytical Engine’s capacities could be extended beyond numbers, but it was Weizenbaum who brought it to life.
Now ELIZA is back, and can run on your computer if you’re feeling nostalgic.
ELIZA has several scripts, but the most famous operates like the sort of psychotherapist who mostly reflects your own words back at you in the hope it will keep you talking until you make a breakthrough. Users would type a message and ELIZA would respond, mostly asking the user for more detail or why the topic concerned them. Updated versions of ELIZA became popular with early adopters of computers.
As Carl Sagan demonstrated in The Dragons of Eden, a somewhat updated ELIZA could perform the role well enough that, at least temporarily, people might think they were talking to a real therapist. Records of some conversations also demonstrated ELIZA’s capacity to reveal truths some people would prefer not to face about themselves and their relationships.
ELIZA’s descendants have certainly come a long way since then, although whether it is for the good is an open question. After all, because ELIZA asked the questions rather than answering them, it couldn’t lie. That seems rather refreshing these days when compared to the steady stream of misinformation modern counterparts cook up, extending into allegations of defamation.
The team makes the case that lessons learned from ELIZA have shaped the artificial intelligence of today. “It is embedded in the AI psyche,” they write.
Although ELIZA’s fame has remained, at least among those interested in the history of computing, the code with which it was written was lost as programming languages shifted. However, in 2021, Jeff Shrager (who had written a 1970s version of ELZA using BASIC) and MIT archivist Myles Crowley discovered a complete copy of ELIZA’s source code in Weizenbaum’s papers.
With permission from the Weizenbaum estate, Shrager, Lane and others have collaborated to “reanimate” ELIZA, a process they write; “Was not simple! It required numerous steps of code cleaning and completion, emulator stack installation and debugging, non-trivial debugging of the found code itself, and even writing some entirely new functions that were not found in the archives or in the available MAD and SLIP implementations.”
In some cases, what the team found in the archives were draft versions of ELIZA scripts, which were not necessarily those released on an amazed world. The resurrection team made some corrections and tested the result using the published version of an original ELIZA conversation to see if the new ELIZA gave the same responses. They report the reaction was almost identical.
One issue the authors discovered in the process of reanimating ELIZA is that the input of numbers by anyone conversing with it causes it to break, at least in the version that has currently been resurrected.
That’s a necessary warning, because ELIZA can be downloaded here to run on your own computer.
Now you too can have a frustrating conversation with a therapist who forces you to face things you are trying to avoid without having to pay for it. But at least it won’t tell you the JWST created the first exoplanet image or turn racist within days, and it may not even lower your critical thinking skills.
The preprint is (appropriately) available on ArXiv.org.
[H/T Tech Xplore]