Post AUPS2tX32xrf4Scgpk by billseitz@toolsforthought.rocks
(DIR) More posts by billseitz@toolsforthought.rocks
(DIR) Post #AUO9XUVS37iA92nB3Y by hasanahmsd@twit.social
2023-04-06T21:12:09Z
0 likes, 0 repeats
Here is a good test to show LLM’s don’t have any intelligence. Because of the fact that gpt4 has learnt the original quiz and has learnt to not leave a goat with cabbage or goat with lion alone, it falls flat when you take one small step outside its training . It cannot think for itself. It refuses to say “there is no solution” which even a 10 year old can see. @caseynewton @jeffjarvis @simon
(DIR) Post #AUO9XVVqJ9cPGXz0Iy by simon@fedi.simonwillison.net
2023-04-06T22:00:57Z
0 likes, 0 repeats
@hasanahmsd @caseynewton @jeffjarvis That shows that a language model can't solve a logic puzzle - because it's not a logic model, it's a language model
(DIR) Post #AUOCLfjglug1MPVX1c by simon@fedi.simonwillison.net
2023-04-06T22:35:12Z
0 likes, 1 repeats
@hasanahmsd @caseynewton @jeffjarvis Here's something fun: I have the ChatGPT alpha that can use Python, so I posed the problem to that (using different words to avoid confusion)> Based on the depth-first search algorithm, there is no solution to this puzzle. It's impossible for the man to take the zelpug, sphang, and cnurdle across the river while obeying all three rules. Therefore, the man cannot solve this puzzle and get all three things across the river.Transcript https://gist.github.com/simonw/79755552c534cf9044b8451e184dde73
(DIR) Post #AUOCaOy7OVhvhYHRlx by sxpert@mastodon.sxpert.org
2023-04-06T22:00:39.023256Z
0 likes, 0 repeats
@jeffjarvis @hasanahmsd @simon @caseynewton proof the term “artificial intelligence” is not appropriate. Next
(DIR) Post #AUOCaPfimOaNsmgGjg by simon@fedi.simonwillison.net
2023-04-06T22:36:27Z
0 likes, 0 repeats
@sxpert @caseynewton @jeffjarvis @hasanahmsd I think the term "artificial intelligence" is massively misleading with respect to large language models - sadly I don't think we can convince society to use a different name for them at this point https://fedi.simonwillison.net/@simon/110152928385185151
(DIR) Post #AUOE7FkIo02kbk2xE0 by mcv@nerdica.net
2023-04-06T22:46:29Z
0 likes, 0 repeats
@simon @sxpert @caseynewton @jeffjarvis @hasanahmsd Artificial Intelligence is a long established field, and these LLMs absolutely fit in. The fact that they have weaknesses doesn't make them any less AI than expert systems or chess computers.
(DIR) Post #AUOE7GZLkCsPA3vinA by simon@fedi.simonwillison.net
2023-04-06T22:53:30Z
0 likes, 0 repeats
@mcv @sxpert @caseynewton @jeffjarvis @hasanahmsd That's true... people started out calling fundamental computer science "Artificial Intelligence" back in the 50s, it's a little unfair to turn around and say "that was a terrible name for the field" 70 years later!
(DIR) Post #AUOEWzKGTdWzDA8hDk by simon@fedi.simonwillison.net
2023-04-06T22:55:47Z
0 likes, 0 repeats
@hasanahmsd @caseynewton @jeffjarvis Amusingly I had to change the names of the objects to sphang, cnurdle and zelpug because when I tried it with goat, lion and cabbage it wrote the code in a way that excluded the "lion and cabbage" rule because it was so thoroughly trained on the classic version of the puzzle!
(DIR) Post #AUOFK8DCEBLoxbGZjU by simon@fedi.simonwillison.net
2023-04-06T23:03:23Z
0 likes, 0 repeats
@hasanahmsd @caseynewton @jeffjarvis I gotta ask: given that GPT4 with access to a Python interpreter CAN solve the trick goat/lion/cabbage puzzle, does that combination pass your criteria for intelligence now?(I don't think it should personally, it's still just a language model with some extra tooling)
(DIR) Post #AUOFooK2E8xMFb5tSK by NireBryce@hachyderm.io
2023-04-06T23:05:55Z
0 likes, 0 repeats
@simon seems like this is going to lead to some very interesting bugs with how uncritical a lot of people are about output@hasanahmsd @caseynewton @jeffjarvis
(DIR) Post #AUOGNyHxAEbpIQt0s4 by hasanahmsd@twit.social
2023-04-06T23:11:22Z
0 likes, 0 repeats
@simon @caseynewton @jeffjarvis no because it is using like a child using a calculator to solve math while the child doesn’t know the rules of math . It built the calculator because it was taught to assemble it from training. I would say intelligent is when it doesn’t require help (tools or training) but it’s own thinking
(DIR) Post #AUOGfslmQE0hONNVT6 by simon@fedi.simonwillison.net
2023-04-06T23:23:51Z
0 likes, 0 repeats
@hasanahmsd @caseynewton @jeffjarvis I don't think ruling out tools makes sense: any form of AI more powerful than a LLM is inevitable going to involve other systems - logic models, world models, long-term memory etc. Each of those could be considered a "tool" in the same way the Python executor it used is a tool
(DIR) Post #AUOzrz6bBfzOSFGEgi by trindflo@hachyderm.io
2023-04-07T07:50:14Z
0 likes, 0 repeats
@simon @hasanahmsd @caseynewton @jeffjarvis I misread your post initially to say that ChatGPT has claimed there was no solution because you changed the names. After reading the problem over there is no solution and the chat engine is giving a completely appropriate response.It is cool that you can hook ChatGPT into python. That makes me wonder how much better it could be if it hooked into prolog.
(DIR) Post #AUPEzuPM2EIZVyGonw by bobthomson70@mastodon.social
2023-04-07T10:39:43Z
0 likes, 0 repeats
@simon @jeffjarvis @hasanahmsd @caseynewton it can’t even manage langage puzzles like Wordiply - stating the longest word that contains a 3 letter sequence.
(DIR) Post #AUPI70z6HWRdnj4PeS by calotriton@sauropods.win
2023-04-07T11:14:38Z
0 likes, 0 repeats
@simon @hasanahmsd @caseynewton @jeffjarvis if it doesn't understand what it says, and in fact, it's nothing like language acquisition, it's funny to call it a language model
(DIR) Post #AUPJ6URvLWvkqirkB6 by Holten@mastodon.cloud
2023-04-07T11:25:36Z
0 likes, 0 repeats
@simon I believe this is a sensible approach. Also, the term "understanding" may ultimately turn out to be unhelpful, a metaphorical cul-de-sac.I'm reminded of Sabine Hossenfelder's recent turn-around regarding the "understanding" part of "AI", pointing out that no human really does "understand" quantum physics, yet still we apply the theory "as is" in technology every day. For practical purposes, we do therefore "understand" it well enough.
(DIR) Post #AUPNVZEt4hbelDytyC by simon@fedi.simonwillison.net
2023-04-07T12:15:14Z
0 likes, 0 repeats
@bobthomson70 @jeffjarvis @hasanahmsd @caseynewton amusingly I tried that against GPT4 and it invented a word that doesn't appear in any dictionary as far as I can tell!GPT3.5 also invented a word, albeit a shorter one
(DIR) Post #AUPOUakE5lT15ujqWu by simon@fedi.simonwillison.net
2023-04-07T12:26:21Z
0 likes, 0 repeats
@calotriton @hasanahmsd @caseynewton @jeffjarvis the term "language model" makes sense to me in terms of how computer science refers to a "model" - something that detects patterns in data based on its training, without implying inherent comprehension
(DIR) Post #AUPS2tX32xrf4Scgpk by billseitz@toolsforthought.rocks
2023-04-07T13:05:47Z
0 likes, 0 repeats
@simon @hasanahmsd @caseynewton @jeffjarvis and language ain't logic
(DIR) Post #AUPSFDQOuaVEF5OAXg by billseitz@toolsforthought.rocks
2023-04-07T13:08:06Z
0 likes, 0 repeats
@simon @calotriton @hasanahmsd @caseynewton @jeffjarvis it's models all the way downhttp://webseitz.fluxent.com/wiki/IsA