4 ms·
Draft Paper Discovered in Which Joseph Weizenbaum Envisions ELIZA's Applications
- adamgordonbell 3y agoI interviewed Jeff Shrager. He is one of the people behind this site and the effort to investigate Eliza. They do a lot of interesting work and the history of Eliza is more complicated than you would guess. (The interview was before chatGPT was a big thing.)If you don't mind the plug: https://corecursive.com/eliza-with-jeff-shrager/ https://corecursive.com/eliza-with-jeff-shrager/
- MR4D 3y agoI’d highly recommend that podcast as well. I’ve listened to that episode twice. It was really fascinating. EDIT - meant to add that the summary style of your interviews is great. Keep up the good work!
- alexvoda 3y agoSame, another recommendation for the podcast.
- adamgordonbell 3y agoThanks!
- zilti 3y agoI really enjoy your podcast! :)
- adamgordonbell 3y agoThanks!
- leejoramo 3y agoWow this is a great podcast episode. I remember using a version the Creative Computing code that I reworked for my TSR-80. Fun to hear the back story
- whyenot 3y agoIt's amazing to me that the "chat bot" interface for ELIZA, developed in the mid 1960s really isn't very different from that of ChatGPT 4, 60 years later.
- hprotagonist 3y agoI recommend also “computer power and human reason”, his 1976 treatise. it really presages all of the last few years of AIspew.
- vincent-manis 3y agoTo Weizenbaum's point, back in the 80s I used to cart a Teletype and an acoustic coupler to high schools to talk about computer science. The big demo was Eliza. Even after I explained that it was a simplistic program (getting it to say “Perhaps we fragisticulate each other in your dreams”), and showed the students the scripts it was using, I found students would want to have serious conversations with it, of the “please don't look at it just now” variety. Seeing that the students and teachers consistently missed the point, I stopped using it.
- TMWNN 3y ago> Seeing that the students and teachers consistently missed the point, I stopped using it. In retrospect you missed the point. But so did everyone else. As primitive as ELIZA is, it is actually closer to the LLM approach than, well, everything else in AI over the past 60 years. People don't realize how simple LLMs' code is. We're talking a few hundred lines! That's comparable to that of ELIZA. All the magic is in the data those lines process. That's what takes up gigabytes of storage and requires many gigabytes of memory and GPUs/Apple Silicon to run. (Quite possibly representative of how seven pounds of meat and 20w of power can still outperform said gigabytes and silicon.) Had such compute power been available, there is no reason to think that Weizenbaum or others at the AI Lab could not have built a true language model in 1967. I suspect that, if anything, ELIZA's simplicity caused researchers to not pursue it further, as it was obvious that true AI would of course require millions of intricate lines of code. Thus we got 50 wasted years building ever more-complex expert systems (i.e., fancy versions of Twenty Questions, or Akinator), or attempts to replicate the human brain in hardware (Danny Hillis named his company Thinking Machines for a reason). A future history of AI may describe the 50 years between ELIZA and "Attention is All You Need" in a chapter called "The Route Not Taken".
- chipotle_coyote 3y agoAnother possibility is that, in fact, vincent-manis and everyone else did not miss the point, and the folks who believe we can achieve “true AI” if we just give Eliza more CPU power and disk space are barking up the wrong tree. To paraphrase Ted Chiang: a learning thermostat has a goal, but it has no preferences or subjective experience. You’re training it to maintain the temperature of your house, but you’re still interacting with a thermostat. The way we’re approaching AI currently supposes that if you just cobble enough thermostats together, you get a person.