Hacker News Reader

I think you've fallen into the trap of "AIs don't generalize, they memorize." But they do in fact generalize. The reason ChatGPT is so valuable is precisely because it can help out with situations that have never been seen before, not because it merely unlocks old preexisting knowledge. The fella who saved their dog with ChatGPT comes to mind. https://nypost.com/2023/03/27/chatgpt-saved-my-dogs-life-aft...

> I think you've fallen into the trap of "AIs don't generalize, they memorize."

Binary classifiers don't generalize?

Just because my output is not generative does not mean we are cannot learn / generalize elsewhere. Think of it as a 2-stage process.

Did you read the article you posted?

-ChatGPT initially gives the same diagnosis the vets did, Babesiosis

-It notes that Babesiosis may have either been a misdiagnosis or there may be a secondary condition/infection causing the remaining symptoms after the Babesiosis treatment didn't resolve all of them

-It suggests such a hypothetical secondary condition could be IMHA, which the article notes is an extremely common complication of Babesiosis with this specific dog breed

-A quick Google search brings up a fair amount of literature about the association between Babesiosis and IMHA

So in fact this is the opposite of a never before seen situation, ChatGPT was just regurgitating common comorbidities of Babesiosis and the vets in question are terrible at their job.

Are you suggesting that it was commonplace for machine learning models to be able to extrapolate medical case reports into an actual diagnosis? Even being able to read and understand a case report is a minor miracle in extrapolation, and it’s interesting how far the goal posts move.

I have another example of ChatGPT not generalizing but just being a really good statistical model. I needed a solution to a problem that you can't find on Google, and a variation of a problem that also can't be found on Google. I attempted to obtain around 15 lines of code from ChatGPT that would solve the problem, but it consistently failed to produce the correct solution. I spent a few hours trying different prompts, indicating its errors and receiving apologies, only for it to generate another incorrect solution while acknowledging its mistake. Solving out of distribution problems correctly seems almost impossible for it.

GPT 3.5 or 4? Surprisingly it makes a huge difference. I think a lot of peoples’ impressions are with 3.5, but many startups couldn’t have been built on it, whereas with 4 they can.

If it was 4 I’d be curious about the specific problem if you’d be willing to link to the chat.

I had similar issue with gpt4, was looking for a library that given a grammar and a string, would produce a list of next valid symbols.

Gpt4 only ever suggested grammar validator to the point I had given up and was going to write a grammar generator, and so I started looking for the equivalent of antlr in python, and in three searches I find out nltk.grammar that actually solves the original problem

It's not a new library either, so I'm dumbfounded by why gpt4 couldn't make sense of my request.

If it hasn't been used much/talked about much for this purpose then you wouldn't expect it to, right?

true, but then it's neither interpolating between the public documentation nor extrapolating from the possible similarities between concepts

I tried this to see. It does suggest using validators, though in a way that mostly solves the problem. The failure mode is when the string is valid complete or extended. It does have a bit of a workaround for that.

https://chat.openai.com/share/54cb8876-96ff-4434-a479-4d2dde...

using the exception only works at whole symbol level, i.e. with "cat do" as imput

No terminal matches 'd' in the current parser context, at line 1 col 5

cat do ^ Expected one of: * CAT * SPACE * FISH * DOG

Next valid symbols: ['* CAT\n\t* SPACE\n\t* FISH\n\t* DOG']

I'm not sure I understand, what that reports is enough to solve it right? Was the problem you couldn't get it to solve it from that point?

No because if I add space to the original string I don't get a valid string

Or given the context space is a valid production from the last error, but then I would need to roll back the incomplete "do" terminal and lose that context.

> No because if I add space to the original string I don't get a valid string

Yeah, it's telling you what it was expecting when it hit an invalid token.

You take the invalid part and filter the list of suggestions with it. If you tell chatgpt about the problem it solves it.

(Edit - when I was running this, having any number of spaces was parsed as valid btw if that's the concern - space was a valid option)

I think you have fallen into the trap of mistaking interpolation for generalization.

Working with these models every day, it's clear that they can certainly interpolate between points in latent space and generate sensible answers to unseen questions, but it's pretty clear that they don't generalize. I've seen far to many examples of models failing to display any sense of generalization to believe otherwise.

That's not to say that interpolation in a rich latent space of text isn't very useful. But it's not the same level of abstraction that comes from true generalization in this space.

I think we’ve all fallen into the trap of mistaking pithiness for evidence

>it's pretty clear that they don't generalize. I've seen far to many examples of models failing to display any sense of generalization to believe otherwise

Failing to generalize on whatever you have in mind is not evidence that the models are incapable of generalization. If you really think so, just be prepared to write off a good chunk of humans as well lol.

This is generalization. https://general-pattern-machines.github.io/

Just seems to me like you've taken examples of generalization and invented an alternate meaning just so you can't admit it generalizes.

"Oh but you see this example of answering an unseen question is "insert meaningless distinction" so it doesn't really count"

With a sufficiently large dataset, every problem becomes interpolation.

Given a multi-dimensional latent space with enough dimensions it's hard to imagine cases of generalization that aren't interpolation between points in latent space (given enough data).

The one possible exception is logical inference, and this problem seems tractable with tool use or programming.

I object on geometric grounds. You can't interpolate outside the convex hull, if your have an outlier in any dimension then you're going to need to extrapolate, that seems to me like a reasonable way to ask for generalisation that isn't interpolation.

> With ChatGPT, you cannot realistically explain why you got some output in a way that anyone other than an AI/ML expert would find satisfying

IMHO OP is talking about "explainability" of the results, which is notoriously bad for current AI. For certain applications (idk if SQL would be one but mortgage application might be one) it is required to be able to explain how the computer got to the decision.

Yeah louie.ai query generation internals feel more like classic program synthesis, where's it's more about staged compilation, iterative refinement, etc, and we use gpt3/gpt4 for each stage rather than a traditional solver. "Which tables would help this query, which parts should go to text search vs vector search, ...", and we iteratively rewrite and pass along until we reconstitute.

At the same time, we're also moving to make a lot more of that process AI controlled, just across LLM calls (e.g., LLM-dictated ones), so it's a funny maturity process.

Curious to learn more about your use case. If fine-tuning is only ineffective for your most complex queries (and presumably those are less frequent as well, since you mentioned you have few examples), then couldn't you use fine-tuning to handle the simpler queries (presumably the lion's share) and thus free up excess man hours to focus on the more complex queries? Is there any benefit to AI being able to answer 90% of queries vs 0%?

These tools are already fantastic at our 80% average case, even without fine tuning. We are seeing some value, but the real pain is in that other 20%.

> Binary classification into deterministic query building.

I have no clue what this means.

I tried something similar to this so I assume it’s to classify a table as being part of the resulting query or not, and using the schemas of the resulting tables in the prompt for the question. I found it just as useful to give it the full schema and ask it for a list of relevant tables, certainly more cost effective.