I worked at OpenAI previously, but don't know any of the people involved in this.
My guess was it was probably this was more a nerd snipe than any action from OpenAI that was a "massive team" being put on it. Literally someone looking at this and asking "I wonder if our models are good enough yet".
It's easy to assume that having access to massive compute amounts means significant coordination, but this assumes that you're looking at the costs of this sort of thing from an external lens. Internally, tokens are often treated as free and infinite.
They said that a customer would have paid around 15 million for the required compute. I can't imagine that this was not a significant internal spending even with "free" tokens.
At Astra API prices that's 300B tokens (I saw 130B output tokens claimed elsewhere), large but not unheard of if you consider it across a few people doing random experiments with best-of-n type things. On my personal account, I've done a billion+ token days just on a normal pro 20x subscription. I know many others that wildly outpaced that by orders of magnitude. This was apparently 130B over 89 hours, so about 30x that rate. When things are free and you're expected to token max 30x seems fairly reasonable to me.
If you consider this as a cost to be compared against the question: "What does it take to be able to prove that you have a model that can solve the hardest problems that humans know about?", then spending a some amount of thousands/millions to know the boundaries of that seems not too important in comparison.
You've also got to consider this as compute that's allocated to pushing the frontier of what models can do, so while it's using GPUs that have been paid for etc., it's not like it's a cost that's supposed to be use less of this so that others can have capacity. If you made researchers afraid to use capacity like this, a lot of the things that improve would tend to do so significantly slower. (some may say that's a good thing ;)
A good way to think about this is when tokens are free, you get to choose whether you're optimizing for latency or intelligence rather than having to consider price.
If you're using coding agents for this, it may be worth splitting this up into multiple well arranged modules that tell a coherent story and make it easy to browse, and add explanatory docs based on the various things the LLM has found about each function / type.
Your decompilation threads have all the necessary info in them for this, anyone coming after lacks that foundation and effectively is doing a second inference over the hidden state, assumptions, etc. that your sessions have in them. A simulacrum of a simulacrum in essence is likely to be not particularly good.
To some extent, and obviously grossly over-simplified here, copyright is the act of taking some already public good and putting private rights on it. My reading of the amicus is that the government is arguing that particular rights that NY Times are asserting that they have are not ones that promote the aims of the original reasons for copyright to exist and so aren't necessarily ones which need to govern the behavior of any other people (in this case OpenAI, but this likely applies to any LLM training no matter the size).
> A real capacitor makes the task more interesting. A ceramic part may provide much less than its advertised capacitance once it has voltage across it. Parts have tolerances. Adding more capacitance costs more, takes up space and makes the rail slower to recharge when the power returns. A design that works with nominal values can fail with the parts that arrive.
It sounds like from a reasonable reading of the benchmark post that there's some things that are being tested that are assumed to be criteria that you expect the models to intuitively find those things to be important (i.e. the stuff about working on parts that have tolerances etc.). If that's so, then this really feels like mostly an exploration of whether an LLM has a good understanding of unstated constraints and has an appropriate in distribution set of priors that would be able to form models where it's reasonable to design on those lines.
It's hard to tell whether this is a problem though as the methodology is imprecise.
If you're spending time on evals against your own product, I'd be super curious to see how far you can get to by using a top tier model to produce generalized instructions for lower tier models. E.g. in a loop: "This eval missed X. what's the simplest single instruction that would have helped this session consider that as necessary that can benefit all future runs. Stick that in AGENTS.md and retest."
A lot of this feels like it comes down to the training of the agents to produce code that satisfies the various benchmarks combined with reactions to things which were previously maladaptive. I.e. things which were explicitly trained out of the model in post training. I think there's a lot of missing long term software engineering principles that don't seem to be baked into the way the models tend to write code by default.
I have some speculation that maybe the people doing the model post-training tend to be younger researchers that haven't worked on large complex software systems, so their taste isn't as developed in this regard about what things are important here.
But this is an area that can be steered with appropriate early instructions ("When choosing tradeoffs of implementation, build for long term maintainability and understandability of code over implementing just the exact code necessary to solve the issues. etc. chain of thought often includes information that would have to be repeated in a future agent session, make sure to persist it to code or external docs so that future sessions and user understanding is respected.")
It can also be done as a post-change step with similar effects. And you can use your agents to build this layer into your general modus operandi for dealing with the crimes of generated code. But one of the things that all AI labs should be doing is looking at AGENTS.md on real project as being hard expressions of what failure modes real projects have noticed in models generally. Don't wait fo the bugs to be raised on these things, use express preferences that show that there's a problem. Go trawl github for these in bulk to use for future post-training.
reply