Back to News
Advertisement
Advertisement

⚡ Community Insights

Discussion Sentiment

47% Positive

Analyzed from 1318 words in the discussion.

Trending Topics

#relativity#models#benchmarks#theory#often#model#training#knowledge#something#benchmark

Discussion (44 Comments)Read Original on HackerNews

cb321•about 23 hours ago
Something Dan does not observe in his article (perhaps Jamie does elsewhere? edit: or even Dan elsewhere) is that the same problem which makes the memory latency benchmark unrealistic (or at least misleading) often impacts hash table lookup benchmarks as mentioned at https://github.com/c-blake/bu/blob/main/doc/memlat.md and probably many other benchmarks. Essentially, CPU work prediction/speculative execution has become so good that much care is often required to measure latency rather than reciprocal throughput. This all started in the 1990s (or probably earlier with Cray), but I guess there's been an ongoing educational failure/oversimplification tendency.

Of course, throughput at one "level of work" is often latency at a higher level (like command-to-command execution, for example). So, "what matters" might be throughput or "latency". It all depends. :-) People often fiddle with such semantics to market methods, products, ideas, ... and marketing is often at cross purposes with understanding.

dom96•about 24 hours ago
This is great. I've been building my own model benchmark lately and it has indeed been so easy to mess up the scoring. It's simply much harder to come up with an algorithm that combines all your individual scores into something that isn't broken in some special circumstances. That's why I think many just start capping the results.
jbellis•about 21 hours ago
While I'm generally sympathetic to the idea that the public coding benchmarks are inadequate, to the point that I've written my own in the past and will likely do so again, the complaint here is that "these tasks don't match what I do in my day job" which is ~always going to be the case. The hope with benchmarks is that you can capture properties that generalize, from examining performance against small set of tasks, and I do think that this is at least directionally true for well-designed evals.

(There's https://withspecific.com/benchmarks/real-swe but since the tasks are private we still don't really know what they're representative of.)

Zigurd•about 20 hours ago
I don't mean to make you write a dissertation but to say that AI benchmarks can "capture properties that generalize from examining performance against a small set of tasks" is a bald assertion. It's a hypothesis without a theory behind it.

I can profile some code and then tell you where the slow parts are. Unless there's some explanatory power to an AI benchmark, it's a bit like benchmarking pillows visually.

menaerus•about 20 hours ago
I started using gemini with caution given the "much worse" benchmarking points it has gotten and still does but in practice there's very little evidence I found in comparison to claude models. It performs really well on non trivial tasks.
stephantul•1 day ago
The (relevant) segue into tires elevates this post so much. I don’t understand why, but it does
vintagedave•1 day ago
Maybe that it's an example of the same pattern in a real-world, non-digital area? It makes it feel more grounded.
jdw64•1 day ago
I feel like AI development is hitting a wall now. Evaluations are driven by benchmarks, but I have no idea who is even evaluating the validity of these benchmarks. The idea that continuously training and scaling models will automatically yield broad capabilities across multiple domains hardly seems true anymore.

Benchmarks like Senior SWE-Bench apply discontinuous and arbitrary thresholds to evaluate results. If the generated code is semantically identical to the reference solution but even slightly exceeds a length cutoff, it fails? That seems like a genuinely flawed criterion, and the LLM-based graders themselves feel highly unstable.

Honestly, my impression is that LLM advancement is now completely dictated by benchmarks. But looking at recent trends, token prices are skyrocketing while coding capabilities haven't shown any massive improvements beyond a certain model generation. In fact, comparing GPT-6 Astra to 5.6 SOL, SOL often writes better code.

Considering all this, while training across diverse domains can pack a model with various pieces of knowledge, ultimately, it feels impossible for this approach to do something like derive the theory of relativity from medieval knowledge.

NitpickLawyer•about 23 hours ago
> AI development is hitting a wall now

People have been saying this for at least 2 years now.

> token prices are skyrocketing

Today's SotA (fable and astra @ 50$ /Mtok output) are cheaper than o1-preview (sept '24, 60$ /Mtok output).

And other models are workhorses, with much better capabilities, are at least 1 oom cheaper today than o1-preview. (I'm using this model, since it was the first "thinking" model)

> it feels impossible for this approach to do something like

The models have just provided lean proofs for FLT (a ~1M$ project that was expected to take a human expert ~5 years to complete) and a Millennium prize problem. These are current, relevant, and previously unsolved problems. The obsession people have with "proving relativity from stone-age data" is just moving the goalposts.

mjburgess•about 22 hours ago
No one spend ~10 mil USD using gpt 5.1 to solve a millenium prize problem with similar solutions in the training data. So we have no reference for whether 5.1, with that training data and that amount of moeny, could likewise produce such a solution.

For each generation of advancement the "AI psychosis" of the previous wave wears off. Those who believed 4.6's reasoning was an accurate account of its behaviour; those who believed it had goals and solved useful problems reliably; and so on -- now, attribute only these things to Fable. And no doubt when Fable 6 comes along, it will be only v. 6 that does that.

We have seen fairly marginal progress in LLM reliability and performance since the meaningful start of the high-inference/high-reasoning harness era. It just takes people a few model version bumps to break out of the addiction loop to realise this.

At some point progress will stall entirely, and a couple years after that the spell will break and people will stop treating LLMs "as AI" in the wide-eyed sense, and start treating them as unreliable tools that map Text->Text -- as they do now with earlier model versions.

usef-•about 21 hours ago
There's definitely hype, but the newer models can undoubtedly do things the older ones couldn't. I had multiple long-term issues that earlier models couldn't solve (after repeated attempts) that fable did in 1 shot. On small models too, the differences in what I can trust them with has dramatically changed compared to a few months ago. Unless you breathlessly never touched the limitatioms of earlier models, it's very clear that the wall of limitations has been moving outward.

Note also when a new benchmark is released, older models do worse on it than recent models, despite none of them being trained to the benchmark.

lh712•about 23 hours ago
Nitpicking: It is not possible to derive the theory of relativity from medieval knowledge. At least not in any way resembling how the development actually happened, that is, heavily influenced by observations (or rather an iteration of speculation and observation). [Unless one counts both Galilean relativity and some parts of the theory of electromagnetism as being contained in medieval knowledge.]

Perhaps it would be possible if the classically observed reality turned out to be possible (consistent under some reasonable conditions) only as a consequences of some deeper, sufficiently determined theory; but that would go far beyond our current level of knowledge, at least as far as I can say.

wongarsu•about 23 hours ago
The earliest you could reasonably derive special relativity is probably 1881 with the first version of the Michelson–Morley experiment, which showed that speed of light is the same no matter how you move through space. Which, on the scale of discoveries, is a pretty short timeline to the 1905 publication of special relativity.

Which doesn't really stop you from training an LLM on 1904 knowledge and having it derive special relativity. LLMs trained with knowledge cutoffs that far back is a fun exercise, and one I've also dabbled in in the past. But you don't have anywhere near enough training data to reach SotA levels of intelligence, even if you had the money for those training runs. And lobotomizing all modern information out of LLMs doesn't seem viable either. So I don't see how you could ever turn that into a viable benchmark

cb321•about 22 hours ago
This is something of an aside on your first paragraph and lh712's. The theory of special relativity (and field theory in general) can be "derived" (as done in e.g., Landau, Lifshitz _Classical Theory Of Fields_) without empirical observation using concepts known to Galileo/Newton (if they count as "Medieval") - just with different assumptions/ideas about inertial frames. If you assume there is any phenomenon with a fixed observed speed in all "inertial frames" then you get special relativity with Einstein's gestalt-switch. After that it is, like so much in physics, a matter of thinking of an experiment to distinguish what matches capital-N Nature best.

Thinking otherwise mistakes a fundamental error that the only way for something to happen is how it historically happened. Reconciling observations forced relativity historically, but it could have been sussed out without those. That there were very fast but finite speed phenomena (which could motivate relativity reducing to Newtonian models) was seen in 1676 by Rømer with the speed of light and eclipses of Io, a Galilean moon of Jupiter. It probably could have been done with Galileo's telescopes in 1610. (That is just one example of a phenomenon with a very fast yet finite speed to suggest other experiments to test that "hypothetical relativity".)

TBH, this all relates to teaching "physics without calculus" and that sort of thing. Is the most clear presentation/derivation that which mirrors the history of our muddled yet ever demuddling ideas or that which starts from our most thoroughly demuddled ideas, best notations, etc.? The momentum is certainly the historical approach, yet there are notable exceptions.

WarmWash•about 19 hours ago
Math produces a bunch of theories when you extrapolate a system forward (or sometimes backwards)

Experimentation is the hammer that smashes all the incorrect theories.

Without the ability to do experimentation, deriving new laws is virtually impossible. A new next step for AI would be coming up with novel experiments, because that is often the hardest part.