Hacker News Reader

Comment by ldjkfkdsjnv | Hacker News Reader

My uneducated opinion is that this paper is bullocks. Maybe they are looking at deeper mathemtical results, instead of every day tasks.

But every single day I am using OpenAI GPT4 to handle novel tasks. I am working on a traditional saas vertical, except with a pure chatbot. The model works, is able to understand which function to call, to extract which parameters, and to know when the inputs will not work. Sure, if you ask it to do some extraneous task, it fails.

Google/Deep Mind need to start showing up with some working results.

Where. are. the. models. google.

View on HN

I use it frequently with my own custom UI framework. It’s never seen my framework before but it can output new, useable code with just a few examples. If that’s not generalization I don’t know what is.

I've given it descriptions of non-existent "franken-languages" composed by telling it to imagine taking programming language A and adding various features I want to explore to it, and then had it correctly symbolically reason about a program written in this hypothetical language that doesn't exist anywhere, so yeah, the notion it doesn't generalize to at least some degree is nonsense, but note this involved tests on a GPT2 scale model so it's not very surprising they had poor results.

That said, even GPT4 certainly has pretty significant limitations on what it manages to reason about. But without comparing their capabilities in other aspects, arguably so do most humans. We tend to force our way past those limitations by learning incrementally by doing over and over. Current models don't get that luxury without complicated fine-tuning steps, so if anything what should surprise us is how well they do with the limitation of only context to act as short-term memory.

The way I understand the paper, his is still a domain-specific tasks, and it is interpolating. What (according to the paper) transformer models are bad at, is out-of-domain extrapolation (which given that modern models are basically trained on all the internet is kinda uncommon).

If so, then I'd argue that most of humans' "out of domain extrapolation" is domain-specific interpolation as well, and that most humans would struggle to meet the bar.

No disagreement.

This is only relevant for the AI-bro fantasy of AI becoming exponentially smarter than humans.

I don't think it's clear whether or not extrapolation in a sense that'd clear that bar is of much relevance to becoming smart at all.

The AI-bro fantasy is of AI building a smarter AI, which then builds an even smarter one. If this isn't extrapolation, does it exist?

I'm not sure whether that requires extrapolation or not. It depends on what you consider extrapolation. I'd be pretty sceptical of the notion that it requires extrapolation on the basis of a use of the term as extreme as the one above.

(And btw., if that is what you consider the "AI-bro fantasy" then I'm firmly in that camp. There is no logical reason to assume that's not possible; short of identifying a non-deterministic, non-materialistic source of intelligence in human brains that violates known physics and that we can't emulate, it's just a question of when, not if, and the reason for not going full tilt on that right now is down to a variant of the Wait Calculation, not a need for intelligence to kick it off - that is, trying to kickstart that now is likely to be more expensive and not be any faster than trying to kickstart it next year, and so on, and the question is guessing at when that stops being true)

The issue is though the the line between in domain and out of domain is fuzzy. This sort of means that generalization is in a continum. Chatgpt has seen enough UI framework code that it can interpolate concepts. This is a form of generalization but people would be looking for a lot more. I guess a better way to check generalization capability is to train the model on just C++ and then see how much it can do stuff in python using only few shot examples.

Another important thing to keep in mind is one paper(wish I could remember which one it was) that showed even larger scale llms have trouble understanding that A=B is same as B=A if they have not seen A or B before

Going from C++ to python is a little unfair. How many examples would a human need to learn python coming from C++? Its probably more than you can fit in a LLM context window.

Not with GPT-4-Turbo 128k, at least if attention works as well at that context size as on smaller ones.

I know it's unfair compared to a human but I'm more interested in how much it can do. Like what level of leetcode problems can it solve and how well does it use concepts presented in the few shot applications. The whole point is to establish an upper limit on it's generalization power instead of comparing it to a human

For anyone else interested, the paper he's referring to is "The Reversal Curse": https://arxiv.org/abs/2309.12288.

"If you've seen one custom UI framework....."

Which part do you think is incorrect?

An aside, but do you find the quality improvements in GPT-4 are necessary for your daily use, or is 3.5 80% as good?

I upgraded because I wanted to see what it could do with a screen shot of a web page. I had it describe the page and create an html version of the page. it wasn't horrible.

Reply on news.ycombinator.com