Yes, Claude can do nine loops(anthropic.com) |
Yes, Claude can do nine loops(anthropic.com) |
I think it’s a compelling idea, that even if AI never achieves superintelligence, there might still be enormous value in their ability to relentlessly pursue a solution for much “longer” (relatively speaking) than a human. Even if they never let us pick the highest hanging fruit, if they instead let us pick every single low hanging fruit, that’s still a massive win.
Here's my issue with this post:
> True to the spirit of the challenge, they didn’t use millions of dollars in computer power. They used Fable 5.1, working within Claude Science, a platform scientists can pay to use.
Okay, billions of dollars have been poured into these agentic LMs, right? Each training run to get the next increment is costing millions of dollars?
This feels like an obvious jab at Navier-Stokes, but where we get to shift the numbers around to hide where the compute actually is being spent ... compute is being spent. It's either being spent in amortization to make the search smarter ahead of time, during training, or its being spent after.
Also love: scientists get to pay Anthropic to work within their special science harness to do science. That's exactly what I dreamed of doing when I pursued physics in undergrad, one or two companies holding the keys to "progress" for a monthly subscription price.
I get this anxiety, and am largely an AI skeptic, but at one point there were only a handful of computers in the world too (same for batteries, or engines, or crucibles, or stills -- it goes way back), and the organizations that had them had a stranglehold on progress in the field, as did the small number of companies who knew how to make them. It got better as they got cheaper and more plentiful.
I guess my point is that there are more important anxieties to feed when it comes to LLMs and the current state of the world.
I don't see how that makes the state of affairs any better.
I think that argument is recursive? These posts aren't very complicated for either of us, but they're written on devices that are fabricated with billions of dollars of semiconductor equipment. At what point do we just acknowledge that we stand on the shoulders of giants?
To me, the distinguishing factor is that the expense not special-purpose but upfront. The model here is trained without foreknowledge of what problems it will solve. Solutions like nine loops are genuine expressions of a pre-existing model capability, even if that capability has not pre-existed for very long.
If I'm wrong here, I'd love reference links. I think of these companies as trying to inspire the idea that Claude (or GPT) are these special alien entities, in a sense?
The moat is very limited. Harnesses aren't crazy hard to engineer. Open models are quite capable.
Or embrace the future and realize that you couldn't imagine everything that would unfold, when you pursued your undergrad.
How you gonna get mad about all this?
>While it’s possible that this is just a much more AI-friendly problem, I don’t think it’s just that: I think the technology has genuinely gotten better.
It has gotten better imo. The blog author mentions 2 cases at the beginning -- users who think that AI will be capped and those who think it will be uncapped. From my perspective, both are technically right -- AI is capped or technically has usually reached some sort of cap, until human innovation improves it. AI doesn't really improve itself on a grand scale so much as humans improve it.
In other words, AI can and does iteratively improve, but every single ceiling we've spotted and broken through so far came from human ingenuity or effort. It will likely continue to require it, regardless of how much it can do on it's own. In that regard, it seems as though all of this will inevitably be "uncapped, until it reaches a cap, and then likely it will eventually be uncapped by humans (again)". Because of this, AI will never perfectly fit neatly into an 'uncapped' or 'capped' bucket, as long as time continues moving and we continue solving issues as they crop up.
In fact I will even admit it, yes AI has gotten better... but also how could it NOT? It has literally all the resources it can has, namely attention from everyone, everybody talks about it, a lot of workers, from state of the art researchers to annotators, all the material resources, from dedicated chips to networking to water to electricity, the entire available dataset of Human written and said thoughts, literally everything published and thus categorized.
AI literally has everything humanity can provide, it better be "better".
There are thousands of people figuring out how to make AI better for their particular thing. Way more than that walking the AI through taking things from problem to solution.
I also liked the article and the author before this post. But let's be honest that this was paid work as part of Anthropic's PR campaign, not just a random blog post.
What bothers me the the marketing angle which invites skepticism and criticism
> Anthropic invited Matt von Hippel to write this post and compensated him for his time.
If you are paying some one, tailoring the discussions then the end result is always going to be biased one showing yourself as the winner. I understand its somewhat organic, still the ratio of marketing and science needs to be balanced. Marketing has to be correct and the results/outcomes should be reproducible
> As it turned out, the result wasn’t all that far away for humans either. A few days after I heard from Anthropic, we heard from Song He, an amplitudeologist at the Chinese Academy of Sciences in Beijing. Song’s group had already gotten the majority of the result. They’d used some AI assistance, based on GPT-6, but not the kind of one-shot almost human-less approach Anthropic used.
This gets me wondering why ai labs aren't proposing their own millenium prize type challenges.
No human-only was close to Navier-Stokes. The team that was close was also using AI.
And how many runs did it take before this one? They say "in one shot" with nothing more than "keep going." But we only see the successful run, reported by the people who ran it.
Every time I mention this someone assures me that it's possible, and they point to simple board games that computers excel at, or real time strategy games where proper use of APM and clicking accurately go a long way. But I haven't seen AI succeed at any decision-focused game where describing the rules requires more than one minute.
I want to see what AI can do in, not simple, and not complex, but complicated toy environments, where decisions are all that matter.
I'm a bit skeptical of their "2 seeds in a row!" boast. Last time I investigated a claim like that I found the seeds were cherry picked. This was way back in the OpenAI Gym days though (remember when OpenAI was open and just doing goofy research like OpenAI Gym?), their leader boards had some amazing claims about certain RL solutions, but when I ran them myself on new seeds they were far worse than claimed.
A lot of influencers are trying it. You can set it loose on games like Slay the Spire without any training and it can win: https://www.youtube.com/watch?v=9bDG0uuHM2w
That's a general purpose LLM. Actually training a model for a game like that would be old news.
I do have to laugh at how the goalposts keep moving to higher and higher levels like "I need to see it win 20X in a row at the highest difficulty! Why has nobody shown this?"
Thanks for the video link. I tried to find one like this recently, but the video you linked wasn't in the results I looked at.
This is framed as an "or", as if they're contradictory.
IMO, Both of these statements are true.
but the author explicitly distinguished between "AI" and "LLM-based AI", and they work for an AI company, where making these distinctions are really important
It's interesting to note that the expensive part of this experiment was the Claude operations expense. For me I find that Claude is a small fraction of my cost with most of the bill attributed to computers to run simulations instead of the AI to monitor and tweak the simulations.
> These scattering amplitude formulas are hard to compute, so hard that physicists almost always use approximations. They do partial calculations, cut off at a specific number of “loops,” a measure of how complicated interactions between particles are allowed to get. The more “loops” they include in their calculations, the closer they get to the real answer, and the harder, computationally, the calculation is to do.
I don’t even understand what type of solution we’re describing here, is it a formula? A program? A Lean proof?
Often dumb models from other plans & subscriptions. I like to choose whatever model is on sale on opencode-go.
YMMV drastically based on what the actual work is. Dumb models can't handle everything. And of course I have no idea how much of a backlog of work you're feeding it or anything. Some people have enough work to exceed any plan, and do the work inefficiently to boot.
Also, you can just buy as much usage as you want at API rates.
I don't know how this can be verified, but I'm not convinced by the evidence in the post. It sounds like the final solution can be ran in a week on 96 cores. But how much compute was used attempting the problem? How much inference? 'Trust me bro?'
I think these solutions are worth millions to these labs, and I think that's what they're spending on them. Far more resources than have been directed at mathematicians and physicists to solve the same problems.
This is the crux of the matter here. Physicists are physicists. They are not software engineers. I read physics between 2001 and 2005, and the programming language they had us use for all of our assignments was FORTRAN 77. They were still trying to decide if it was ok to move students on to FORTRAN 95.
FORTRAN is pretty performant - if you know what you’re doing. Very, very few people did - and they, me, went on to have careers in software, not physics. I remember some FITS (astronomical image format) processing software someone was using to calculate ephemera - and it took DAYS to run over a few thousand images and produce an output. I sat down with it, screwed around for an afternoon, and it produced a result in under a minute - so much faster that I honestly thought I had broken it - but I hadn’t. It was just terribly written by someone who was excellent in their domain and terrible at writing code.
I see there being an enormous opportunity here, in AI providing scientists with software that isn’t diabolical.
So I took a look for about 30 mins figuring out the code (I don't know SPSS or R), and it turns out she was doing some calculations that were identical for every single line, regardless of that line's fields. So I pre-calculated a few things, prior to starting the per-record-loop, and then just used the pre-calculated values every time. After that, every run took less than 5 minutes. She's amazing at math and economics, but optimizing low-level code is not something she has done in the past.
The main fix was not making it hideously IO bound, as not only was it loading every image multiple times, it was then grinding away in swap as we’re talking tens of gigabytes of raw image data, in an era when 1gb of ram was a lot - and then making it spiral out from the last known location of the object of interest rather than brute-forcing it. My solution wasn’t even optimal, as it was literally just an afternoon of tooling around as a favour.