Did you forget to tell Claude to make it run fast?

(I am not joking. I did a port of an old utility, Dos Navigator, and I gave Claude+Codex a set of KPIs: cold start under 100ms, 30MB RAW image preview under 50ms, and a few others. I went to bed, and I woke up to a file manager written in Swift that is a joy to use, its so fast, esp. when previewing images. If I haven't requested it, it would likely be way slower).

How much did it cost you? Or was a subscription-based usage enough?

Subscription. Translated to Around $90 in tokens.

Would you mind sharing it?

The details of the Apple vs. OpenAI dispute over leaked data show their security is not quite that tight.

Loss of life and other calamities were not enough to discourage humanity from things like cars, alcohol, nicotine, chainsaws, fast food, etc. etc. I am not sure if there is an example of in-some-ways-useful technology where humanity took a measured approach, weighed the pros and cons, and decided to go back. And AI is useful in so, so, so many ways. People I know already seem to have defaulted to letting AI do most of their thinking, on matters big or small. Nope, loss of life won't change a thing.

The question is: isn't the cat out of the bag already? With what is already in the public domain, and the compute available to motivated, deep-pocketed actors, will they be able to carry on the research without the key researchers? And will those people be able to recruit key researchers with sufficient motivations?

Unlike with nuclear proliferation, there is no heavy industrial base requirement. No time consuming, visible uranium enrichment. All it takes is for someone to buy sufficient amount of compute and try to get it past the RSI gate. Or bribe people with access to model weights to existing frontier - the asymmetry between what it takes to bribe a bunch of geeks vs. what is at stake is staggering.

Best thing I saw on hacker news in a while.

In case of my company, the port was planned for this year, with a 9 month timeline and about 10 FTE team. It ended up being done in 3 months, with 2 engineers. (Sure, plus testers and internal bureaucracies, but that 2 FTE for 3 months vs. 10 FTE for 9 forecasted is apples-to-apples). Sure its eating up SWE jobs.

We did the same with our apps used by a few hundred thousand people a day. We are a slow, boring company, so the port took about 4 weeks of engineering work, and maybe 3 months taking into account release process and change management. But in fairness, we had the first working version after a day as well. This was in February, too, with substantially less capable models, unable to do overnight runs.

There is nothing even close to a proof. A lot of accusations, a lot of people ready with pitchforks and torches (sadly, also here on HN), but not a lot of facts.

Did the researches opt out from data sharing on subsidised subs?

Did anyone prove that their methods enabled OpenAI models to produce the solution?

For a discussion about science, there is almost no scientifical method applied to proving anyone stole anything.

On one side, yes we don't have hard evidence that intentional plagiarism is exactly what happened.

On the other side, the lack of evidence is pretty damning. Only OpenAI can try to prove that they came by these results legitimately, and the case they're making is quite weak. They could make public metadata about what their model was trained on and whether it did train on the conversations in question; they have not. TBQH I read it as even they don't know.

And regardless of whether the result is legitimately obtained by their model, they've not at all conducted themselves well throughout this story. They set out to scoop researchers based on a rumor. They threatened to ruin a mathematicians career. They put up a paper that deliberately doesn't cite the most relevant research, despite building directly on it. No matter how you look at it, OpenAI has and should lose any standing they had in the research community.

That it was a dick move, I think there is no doubt about that. OpenAI wanted to scoop Anthropic, and the two guys working on the problem got caught in the crossfire.

Both OpenAI and the researchers know if the sessions in questions were subject to data sharing. Why neither the scientists nor OpenAI is clear about that is weird - it would seem at least one party has the incentive to report that. But even if their sessions were in training data sets its hard to tell whether it influenced the outcome. Those models are big, but are they big enough to preserve subtle, niche techniques enough to draw from them while solving a related problem? Probably nobody knows.

Actually no, the lack of evidence can be readily fixed by the researchers simply disclosing the pertinent parts of their notes and/or chats. The discovery has been scooped, so I don't see any value in keeping them private anymore. Then everybody can see how related the models' and the researchers' works are.

And its reasonable to assume that if they did, in fact, opt out, they'd make it clear.

Most people do not understand that the main reason for the subscriptions is to give OpenAI and Anthropic the priceless, unique data that shows how the models are used, what people are building, how they are building, which solutions they consider OK, which they consider bad -- they purchase this data with cheap tokens. This is their only moat, really. If some really proprietary IP gets swept in the training data set its not really OpenAI's fault -- its the researchers'. Have something secretive? Dont fricking paste this into chatgpt. Duh!

(I'd definitely not think OpenAI/Anthropic ignore the opt outs, or ZDRs. All it would take is one whistleblower to get them into terminal troubles. And why would they do it? They are not in the business of scooping unique IP -- they are in the business of understanding how AI is used across a variety of mundane, day to day work of individuals and companies. Useless math problem is good (or bad, as in this case) PR, but otherwise entirely worthless for the labs.

8x RTX PRO 6000 or 4x Spark? Or 1x M5 Ultra 512GB.

The model is theoretically FP8, but really internally its mostly FP4 already, so there won't be a cut-in-half-but-almost-just-as-good quant coming for this one.