Comments URL: https://news.ycombinator.com/item?id=49912469

Points: 3

# Comments: 1

- we get novel, emergent properties and capabilities of these models that were not trained

- they have very clear generalization to out of domain problems

The point is people conflate the end product: a model that can clearly do very novel, useful and interesting things, with the vehicle for getting there which is a series of optimization steps involving next token prediction loss.

You mention limitations; we all clearly know the practical limitations of these models today, but if you look at scaling laws and empirical performance trends (epoch capability index for example) as well as the trajectory over the last couple of years (very stable), the claim that there is some sort of fundamental limitation is now surprisingly the claim that has the burden of proof.

You can claim it may be e.g. finite context. That is fundamentally bad for certain classes of tasks. This was the hypothesis of a lot of lab leadership of urgently trying to anticipate how to get around this bottleneck (still of course lots of work on this) but the surprising thing is it does not appear to be at this point a blocker.

A better example is maybe Maxwell's laws. Maxwell wasn't inventing a theory to try and explain confusing results, he was unifying a chaotic, empirical laws from existing experiments. That may be a cleaner example. That knowledge compression into satisfying theoretical framework is likely what is attractive.

You can potentially ask the same thing about like you say -- general relativity / quantum gravity but also likely plenty of other areas that may be like this today. Again going outside my particular area of expertise: standard model physics is in a large important sense empirical; lots of values and numbers that are simply unmotivated by theory or where we don't have a good way to make a principled theoretical choice. That could be a place where these models are able to help.

But right now: I doubt it. This is what everyone is working furiously on right now. How do you close a "science" verification loop? In principle this should be easy right: you have ideation (exploration, sampling with ~high temperature maybe as an analogue) and you have verification (which of these ideas are good) which amounts to rejection sampling in idea space. You have to have a sampler that is good at picking _good_ ideas for efficiency sake and you need a relatively fast and reliable verification step of "is this idea good and worth continuing to explore". But I may oversimplify

Things were slower and harder before but the baseline of frustration / rage was never this high, even as the models have gotten objectively better in many or most respects.

- Does the epoch capability index progress show signs of plateauing? I consider this a good aggregate measure of diverse benchmarks into a single capability index. If we see things slowing down here thats a pretty direct and convincing piece of evidence for a stall. - Do we see any signs that scaling laws are beginning to fail? That would be by far the most alarming to me, since I would interpret that to mean that the entire premise of this unprecedented capital allocation tsunami is broken.

Neither of these are true (for now). Progress is marching the same as it has for 4+ years now. It's still the same time to get a generation leap (I think like ~16-18 mo? Epoch has it) like GPT4->5. My theory is that people interpret plateauing because the releases are far more frequent now than they were in the past.

Or not providing inflation adjusted values when referring to historical financial reference points.

I have to go find a happy place now

> Private, refreshed test sets attack the mechanism itself, and in my view they are the only intervention that does. If the questions have never touched the public Web, they can’t be in the training data; if they rotate, memorizing this year’s set doesn’t help next year.

That’s what we have. A fresh public benchmark is also good, and teams do make efforts to decontaminate training data but there’s likely just no great way around leakage.

Btw, lots more issues in benchmarks than the ones discussed; for instance you can leak answers from the questions themselves or in the case of e.g. multiple choice formats in the actual answers. You just pass the MCQ choices themselves to the model and it may be able to guess way above chance. Coding agent benchmarks sometimes forget to delete .git. They mention e.g. a 6.9% error rate in one of the benchmark items, this seems pretty typical and I would actually be fine shipping that.

Benchmarks are very ugly, but if they didn’t exist we would need to invent them. All of the problems above and more do not explain the progress we see. There are probably 50,000 benchmarks in the literature and new ones get created frequently with varying levels of quality and usefulness.