astro1234

I agree with this and think this is the only argument I’m aware of for real decisions to use of Fortran, which I suppose means its future is doomed with human turnover + ease of porting in these AI days…

The answer was “my adviser insisted because that’s the language he used”. I had to fight to use Fortran 95 and not 77. So I think like you say value was building on an existing community’s codebase in a language they were familiar with and comfortable with.

Having written an open source Fortran library (for x ray scattering and absorption in a particular regime so this is fairly “niche” shall we say), genuinely curious: what if any advantages or strengths does Fortran have these days over other language choices? Are there circumstances where it’s arguably the “right” choice?

It gives me 404 error

I used Astra to help me build an immersive VR experience where you can watch the universe evolve from ~50M years old until the present day. Uses the TNG100-2 simulation. Works in browser too!

Well I think in the absence of convincing pieces of evidence to the contrary you might be right. You’re making an empirical statement but we have already answered it today:

- 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.

I think you make a fair point, but also remember: new paradigms don't necessarily require confusing / contradictory observations. You could have the simple idea of "what if gravity is an inertial force?" at any period in time and work out the mathematics of this. It would make theoretical predictions which could then be falsified, but then again, who would take it seriously enough to test it if it was maybe say 1850 and not 1899.

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

Do you mean after? People do this!! But I think it’s a bit different. It won’t be apples to apples because the data volume I think is just so much different. Maybe there are good experiments for something like this.

I think they mean improve our understanding of physics with new theoretical results or paradigms. Like if it’s 1899, would Astra develop General and Special relativity on its own?

My job has gone from coding, plotting, writing to solving the riddle of what Opus 5 is saying and figuring out what is bullshit, what is valuable, what is a total divergence from what I asked it to do. Then eventually at token like 300k starting to shout at it in all caps with obscenities.

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.

I don't think it's a stall, two ways I would believe there is a stall:

- 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.

I’ve noticed this too but it hasn’t been obvious to me that this style of search is not a learned behavior. Tool calling is very much part of the post training phase, I would expect that these style searches just naturally emerge during training. This is just my prior though.

The Opus-human patois is unbearable

Im curious if you find this to be a parody in a bad way or simply a “the state of the art in math research right now is telling a machine to believe in itself”. I am in the latter camp…

Oh my god do not trigger me about the news reporting “points” for the DOW or the S&P, I have seriously considered calling news stations about this.

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

I have to go find a happy place now

Yes! I am aware of this and they are really great. Problem is nobody will ever read that, but they WILL watch Fox News, CNBC, etc. and those outlets never even point out that there are errorbars/uncertainty. Then I hear things like outrage when BLS revises numbers up or down (facepalm).

Oh I’m well aware the BLS knows what they’re doing and publish error bars, they are great. The problem is news articles, no one ever reads BLS reports but they do watch Fox News, CNBC, etc. and those outlets never ever get into uncertainty.

As a data scientist it is very frustrating to see a consistent lack of error bars on these numbers or any discussion about the magnitude of uncertainty around them by all news reports every time these numbers get released

I agree and that’s why we need and indeed have an ever evolving landscape of benchmarks

> 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.

I don’t think that’s fair to say, there are two principle sources of data that paint a fairly consistent picture

- one is scaling laws, where we found years ago that pretraining validation loss scales in an almost miraculously predictable way with data volume and compute. There are apparently theoretical bases for this that I don’t quite understand but this property alone is holding at every scale we’ve ever tested. There is not just “one” scaling law but the point is there are scaling laws and they continue to faithfully predict the performance gains we see

- one is benchmarks, which I always point to epoch capability index as a good summary of them in aggregate which makes it nice to plot on one curve the capability improvement over time

To me either one without the other is substantially weaker, the fact that theory and empirical measurements give you a very good scaling law on a more unintuitive quantity (pretraining validation loss) that’s only indirectly related to the downstream performance you care about, benchmarks (in aggregate) are more direct measures of downstream performance but are harder to nail down clean and well motivated “laws” from theory (as far as I can tell). Nevertheless we do in fact see a clear trend that is not slowing.

That doesn’t mean there aren’t a whole host of benchmark problems that don’t impact the numbers involved here (leakage from training data, fundamental flaws in the design, benchmaxxing) but they don’t change the larger story. These problems don’t plausibly explain the clean trends we see.

Yea this may explain part of it or all of it, it’s likely a case by case kind of thing.

Also to respond to the parent comment: benchmarks have a variety of difficulty levels. Humanity’s Last Exam, though now hitting the beginning of a saturation phase with Fable, was long unsaturated while other benchmarks saturated awhile ago. So that’s what I meant by Epoch capability index: using IRT models this effect so that you gather robust signals from variety of benchmark difficulties and can track progress over time as model capabilities have evolved (and so benchmarks have had to evolve to keep up).

But yes like I was saying: all benchmarks are problematic, some are useful. Benchmark quality problems abound, so 90% being the true ceiling is not surprising. There may be other factors at play here too, I haven’t studied this problem that deeply to have a good thorough answer to this. But keep in mind there are probably 50,000 benchmarks in the literature and that is not a joke number. A crapload of noise in that signal but it’s not all noise.

I work in AI evaluation, lots of problems and leakage is an issue as is ecological validity, but they definitely do not explain the progress we see.

I think Epoch has the best analysis I’ve seen on evaluation trends; they use IRT to basically model a variety of benchmark difficulties, and then model a capability parameter for each model. This is as robust a sort of “meta-study” of evaluations as I’ve seen and the trend in capabilities show no sign of slowing down.

So I think people’s feelings clash with reality, and that’s because releases are more frequent and the jumps between releases are smaller, but the growth in capabilities _over time_ has not changed for the better or worse over a very very long period of time.

I find them useful bellweathers of genuinely out of domain performance and capability, regardless of their theoretical importance. What I see is that performance trends are remarkably stable both upstream (miraculous scaling laws of pretraining on validation loss) and downstream performance (epoch capability index). We get the equivalent of a GPT4->GPT5 performance leap every ~16-18 months, and we are not hitting ceilings nor do we see any deceleration.

Today we can solve nontrivial open problems. What will we be able to do next year or the year after? 18months ago no one was using a coding agent seriously. Now for a large segment of the population you cannot do your job without them.

I think the question and definition game is interesting only inasmuch as it helps us understand ourselves (what actually explains some of the mysterious properties of our perceived consciousness) or helps guide us towards improving performance and reliability of AI models.

Yes, I was hoping they would actually offer some kind of partnership with the sports leagues to stick a few of their cameras in stadiums and offer live sports games, but I don't think there's nearly enough users to support it (at $4k its an insane thing to buy), but honestly I don't even regret it, best gadget I've ever bought. About 80% of my time is spent using Mac Virtual Display, this is so much better than monitors -- I can work literally laying down or in any position, so much more screen real estate, plus it just feels like living in the matrix. But YMMV of course...

But -- you should DEFINITELY book a free demo at an Apple Store, just to see what its like and decide for yourself

It is absolutely unreal the experience watching the ~10 minute sports games on the Vision Pro. The thing is 4 grand but that’s how much the first 4K TVs were. Hope the price can drop where this becomes a more widespread experience, I am a virtual daily user of my AVP

Of course but I think labs is a good term. Like a lot of terms like this it comes from history. The people working on AI at these places come predominantly from academia and predominantly do research. Of course now these companies are far more than just a “lab” and I can imagine their headcount may be more non-research folks at this point but most of the results are indeed coming from the “lab”-y areas

I don't have personal experience with this -- but if I were in the situation where my job was jeopardized by better price efficiencies, I don't know that I personally would have a problem with that. If Ukrainians can do my job just as well or better than I can and they can do it cheaper, I would be worried if my company didn't take advantage of that.

That's my own personal opinion. I can understand how e.g. local and/or temporary changes to wages would be far more concerning and devastating to jobs that are lower paying, but since the question is on tech jobs which are typically much higher paying, I think it's fair game for companies to make strategic decisions in the company's best interest (as long as it doesn't run afoul of laws or human rights). I would be happy for the folks that were able to participate in an economy.

To me it would be unethical to create artificial barriers to lock humans out of opportunities just so I could keep an effectively inflated salary (since this assumes that the true value of the labor is actually lower than what I'm being paid).

Immigration having downward pressure on salaries is a common and demonstrably false misconception. Immigration adds more members to the labor pool while also increasing demand. The two, in most major studies, cancel each other out.

"Immigrant ethnic cronyism" -- would you care to explain a bit more?