- cross-posted to:
- futurology@futurology.today
- cross-posted to:
- futurology@futurology.today
I’ve been saying this for about a year since seeing the Othello GPT research, but it’s nice to see more minds changing as the research builds up.
Edit: Because people aren’t actually reading and just commenting based on the headline, a relevant part of the article:
New research may have intimations of an answer. A theory developed by Sanjeev Arora of Princeton University and Anirudh Goyal, a research scientist at Google DeepMind, suggests that the largest of today’s LLMs are not stochastic parrots. The authors argue that as these models get bigger and are trained on more data, they improve on individual language-related abilities and also develop new ones by combining skills in a manner that hints at understanding — combinations that were unlikely to exist in the training data.
This theoretical approach, which provides a mathematically provable argument for how and why an LLM can develop so many abilities, has convinced experts like Hinton, and others. And when Arora and his team tested some of its predictions, they found that these models behaved almost exactly as expected. From all accounts, they’ve made a strong case that the largest LLMs are not just parroting what they’ve seen before.
“[They] cannot be just mimicking what has been seen in the training data,” said Sébastien Bubeck, a mathematician and computer scientist at Microsoft Research who was not part of the work. “That’s the basic insight.”
I’ve read the article and it’s just clickbait which offers no new insights.
What was of interest in it to yourself specifically?
It provides an entirely new framework for analyzing skills in LLMs. Do you mean the article doesn’t provide new insights, or that the research doesn’t?
As for my own interest, in addition to this providing a more rigorous framework for analyzing what I’d already gotten a sense of with the world model research papers over the last year, I can see a number of important nuances.
First off, there’s the obvious point of emergent capabilities being a hotly debated topic in research circles, which you likely know if you’ve followed it at all.
In particular, the approach here compliments the paper out of Stanford disputing emergent capabilities because other measurements of improvement are linear as size increases. Here, linear improvements in next token prediction directly tie into emergent skills, so it’s promising that the model fits neatly with one of the more notable counter-point nuances in the past year.
I also think this is an exciting approach if the same framework were remapped to the way Anthropic’s research was looking at functional layers as opposed to individual network nodes. By mapping either side of the graph to functional layers it may allow for more successful introspection into larger models than we’ve had before.
A framework around a controversial research topic that generates testable predictions and then sees those predictions met is generally worth recognizing too.
Finally, I think that Skill-Mix may offer a useful framework for evaluating models, particularly around transmission of skills from larger models to smaller models using synthetic data, which has probably been the most significant research trend in the domain over the past year.
So it’s noteworthy in a number of ways and I could see it having similar impact to the CoT paper within research circles (where it becomes a component of much of the work that follows and builds on top of it), even if not quite as broad an impact outside of them.
I’ve generally felt the field is doing a poor job at evaluating models, falling deeper and deeper into Goodhart’s Law, and this is a promising breath of fresh air.
As they say opening their paper on it:
It’s about time we move on to something better than the current evaluation metrics which we’re just trying to game with surface fine tuning.
I question the value of this type of research altogether which is why I stopped following it as closely as yourself. I generally see them as an exercise in assigning labels to subsets of a complex system. However, I do see how the COT paper adds some value in designing more advanced LLMs.
You keep quoting research ad-verbum as if it’s gospel so miss my point (and forms part of the apeal to authority I mentioned previously). It is entirely expected that neural networks would form connections outside of the training data (emergent capabilities). How else would they be of use? This article dresses up the research as some kind of groundbreaking discovery, which is what people take issue with.
If this article was entitled “Researchers find patterns in neural networks that might help make more effective ones” no one would have a problem with it, but also it would not be newsworthy.
I posit that Category Theory offers an explanation for these phenomena without having to delve into poorly defined terms like “understanding”, “skills”, “emergence” or Monty Python’s Dead Parrot. I do so with no hot research topics at all or papers to hide behind, just decades old mathematics. Do you have an opinion on that?
No, but I have learned over the years that when you see multiple papers discovering similar things at odds with the held consensus and see some even independently replicated that there’s usually more than just smoke.
The paper was titled “Skill-Mix: a Flexible and Expandable Family of Evaluations for AI models.” Quanta, while a Pulizer winner in 2022 for explanatory reporting, is after all a publisher not a research institution. Though I dispute your issues with the headline as it’s in line with similar article headlines such as “Bees understand the concept of zero”.
You wouldn’t be the only person looking at it through that lens. It was more popular a few years ago I think, and hasn’t really caught on for LLMs vs other ML approaches and here it strikes me a bit like those with hammers looking for nails - the degree to which there’s functional overlaps in network introspection such as the linked Anthropic work suggests to me that the internalized delineations are a bit fuzzier than would cleanly map onto a category theory view - but it’s possible that as time goes on that it gets some research wins assuming it can come up with testable predictions that are successful. But it’s more of a ‘how’ than a ‘what’ question - whether a network understands abstract concepts tangental to language it is trained on and develops world models (an idea that would have been laughed out of the room just three years ago by any serious researchers despite your impression) using something that can be explained through category theory or through another interpretation, the result is arguably the more important finding than the interpretation of the means.
It seems like you may be more committed to arguing the semantics and nuances of the tree in front of you than discussing the forest - that’s fine, it’s just not that interesting to me in turn.
To hijack your analogy its more akin to me stating a tree is a plant and you saying “So are these” pointing at a forest of plastic Christmas trees.
I’m pretty curious why you imagine you have so many downvotes?
Because laypeople are very committed to a certain perspective of LLMs right now.
You should see the downvotes I got a year or two ago explaining immunology research to antivaxxers.
Have you ever considered you might be the laypeople?
Equating a debate about the origin of understanding to antivaxxers…
You argue like a Trump supporter.