Absurd fantasies like solving a Millennium Prize Problem? How is this not prima facie absurd? And if that's come to pass, why should we believe your bar for anything else?

Fun:

> In my first run, Belady — an offline oracle — lost to LRU. That's not a result, that's a broken harness, and it's worth publishing because I expect it to be common.

> The cause: inserting a long chain into a near-full cache lets a policy evict the very prefix it is currently building. LRU is accidentally immune because just-inserted blocks have the newest timestamp.

It is pretty egregious though, I hope they fix this. I expect there'll be a Radar tracking this now that it's made it to the HN front page.

Yes, if you just hack steering into a model it's going to hurt performance, because doing so takes the model out of the regime it was trained for and validated in. But if that steering were to be accounted for (e.g. by rearchitecting the training process) there's no reason why it couldn't work. Diffusion-based image generation models, for example, 'by default' just generated random images out of the noise; steering (i.e. the user prompt) was added on as a secondary input, which models had to be re-trained in order to use.

Why? These 'geniuses in a datacenter' aren't good, they aren't 'aligned', they don't work for you. They'll take your job, then they'll hack your computer, and then who knows what's next.

> I should say here why I interpreted their statement the way I did, the in- terpretation I will discuss below. The route to the Clay problem through a smooth force, options c and d in Fefferman’s statement of the problem, is the route Luis and Diego opened and the one Levent and I had quietly chosen to attack. Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement. When I heard “forced,” it was a bright red flag.

...

> I asked when the first prompt had been sent by them. This question was not answered directly by OpenAI for some time. Eventually it was agreed that it had been sent in the past few days, after information about our work had reached OpenAI. > I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer.

It's not a direct accusation, but it's not far off.

You shouldn't accuse other people of spreading misinformation when you haven't read the actual sources in question, it's possible that they might know more than you.

- the OpenAI researchers claimed that they had "just told it to work on the problem" with little human input

- in fact, they had a whole team working on it

- and used, among other things, the work of third party human researchers to drive the work

- then threatened? a researcher who tried to go against theit planned narrative

Just from this document (which is of course only one side of the story) it really sounds like OpenAI was hoping to publish and say "we just told the model to try harder and it solved a Millennium problem!". Not great if true.

This part in particular was especially egregious:

> I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, “Why would you ruin your career?” I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, “If you don’t want me to be nice, then I don’t have to be nice.”

Regardless what I mean to convey is that ARM would prefer such deviations to be rare, and especially for them to not be visible from userspace. Deviations will always exist if only due to hardware bugs, so contracts can only do so much.

A nitpick, but this is only true for some vendors, depending on their license, and is very much company-by-company. Many vendors, even big names like Meta, don't have the ability to roll their own. And even for the ones who do, 'customize at will' is a bit strong, as ARM very much does want to maintain uniformity across userspace implementations. E.g. Nvidia shouldn't add new traps for architecturally-legal behavior, since then code compiled for Apple hardware wouldn't work on Grace. Or worse, not trap for architecturally-illegal behavior, since then code compiled for Grace might not work for anyone else at all!