Adel Ardalan
Melanie MitchellOne of my new years resolutions is to blog (from time to time) about interesting work in AI. I'm trying out Substack for this. My first post is a perspective on recent paper by Webb et al., "Emergent Analogical Reasoning in Large Language Models".
Luiz Pessoa@melaniemitchell
This is amazing!
This is amazing!
Ross Gayler@melaniemitchell blogging on a recent paper by Webb et al., "Emergent Analogical Reasoning in Large Language Models".
@cogsci #CogSci #CognitiveScience #analogy #AI #ArtificialIntelligence #LLM #LargeLanguageModels
Sal Randolph@melaniemitchell Subscribed!
A Martin@melaniemitchell this is great. I’m excited to read it!
Timothy Gowers@melaniemitchell A very nice and persuasive post.
I noticed what I think is a small typo that's worth correcting (if I'm right): in example 2 in your appendix, I think ababc should be aababc. (I also think that the abstraction ability I needed to spot that is way beyond GPT3's capabilities!) Also, in 9 I don't see how to justify rloyg -- I would go for rlyg. Much less interestingly, the arrow in 8 seems to have become an à.
Many thanks. I will correct the errors.
Martin Wattenberg@melaniemitchell Very interesting post! A question: what makes an error "nonhumanlike"? If I imagine reading these analogies out loud to a random friend, I can easily see them making the kinds of errors the essay calls "strange." Is my intuition off? Or does the definition of "humanlike" depend on whether the baseline is "quick casual oral conversation" versus "someone in a library reading, with a pencil and paper and time to check answers."
@wattenberg When humans make "errors" on these, the errors typically have certain kinds of logic that other people can understand. But many of GPT-3's errors don't seem to have that kind of logic.
@melaniemitchell It's funny, a lot of the GPT-3 errors seemed very much like mistakes I could make if I heard these spoken aloud and had to answer quickly.
E.g. the essay says: "abcdx —> abcde, pyrstu —> ? GPT-3 answered pyrst, which made no sense to me."
But in a spoken word context, with limited human short-term memory, I could easily imagine thinking the rule was "remove the last letter," because abcd and abcde are easy-to-confuse "chunks".
Ah well! Maybe I just failed the Turing test!
ecsd@melaniemitchell
You have a "signup" wall.
You have a "signup" wall.
What is a "signup wall"?
The site won't let me see anything until I have "signed up", which I consider an intrusion on my privacy. I can try again to verify so. If you yourself specifically elected the policy, I can waive my objections.
@ecsd If you click on this link it should let you see the article without signing up: https://aiguide.substack.com/p/on-analogy-making-in-large-language
I am not current with AI, but I am /dramatically/ interested in machine intelligence. I rely a great deal on my intuition, which has told me that GPT is (fatally) off-base, did you see my remark gone by here a day or two ago. One does not MODEL things as (I presume that) it does and expect anything useful from it - the clumsy results I've seen so far, I mocked it as monkeys trying to replicate Shakespeare.
It's interesting, but intuition says it's not the path to the goal.
1/
It may be /a tool/ on the way to the goal, but will never per se arrive on its own.
(Having said that:)
A first application for GPT would be to analyze search engine patterns to 'comprehend' distinctions people make as to what they're searching for, to narrow the results down properly.
Otherwise, as I've seen it so far, it's just a toy that is nowhere close to "almost but not quite." You have STUFF but no CONNECTIONS. "this defines as that" doesn't DO anything.
{laughs}
2/
now on to read your article.
3/3
See my pro-AI short story at https://shitnobricks.com/?p=105
I wrote it originally in response to sci-fi always assuming AI would be evil. I claim it could not be, if we've taught it enough.
meanwhile, a "paywall" requires pay and an "adwall" requires disabling adblockers on the site.
For fun see "Cogito Ergo Nego" at https://shitnobricks.com/?p=105
Matt Slocombe@melaniemitchell Super interesting post. Regarding four-term verbal problems, I wonder how useful these are as tests of generalisation in models like GPT-3. From my understanding, GPT-3’s architecture and training might make it more adept than humans at generating solutions via non-relational associative strategies. If GPT-3 is generating solutions via associative strategies, is that really generalisation / analogical reasoning?