kzrdude

I don't think it's about that. It's not about the step by step plan for the murder. It's about - how does "the AI" do it? Do we give it access to our world, or do we give it a body, so that it can take its own physical actions?

I like you, I think very similarly. Humans are "like rats": we can live almost anywhere, we'll find a way to survive.

But that just means we won't all be wiped out. We need to understand when discussing global issues, such as this or like climate change that it's about prosperity and quality of life. We're trying to plan for a good life (for all people?).

That's an interesting point and I wonder how far it has been proven. I guess it is validated by the present. Scandinavian cultures had laws and communities with courts all without writing, they had a storytelling and law-reciting culture instead.

Speaking is a culture that has served us well for 300 000 years or so. To note, it has been used for both good and bad..

Fortunately, it's very simple to navigate between pdf, html, and abstract, just put "abs" in the URL where it says pdf and so on.

The topographic map says, the Bay Area is just around the first bay, there's a whole second bay opening up behind it (Sacramento area valley) which could flood with enough sea level rise.

What's more valuable than a good benchmark? IMO a benchmark that has been run against very many competitors and versions. Collecting data has something going for it, and it's up to the readers to interpret and make the best use out of it.

Looks like Q is used as an abbreviation for CH

Typical example to show that people who want to be angry will be angry, even for pointless things.

"a history lesson" and "an historic event" have different stress in the history area (or at least they can have, depending on the phrase context), and I think that explains it.

Saying "a historic event" requires you to pause and stress the start of historic (a + glottal stop + "his" + ...). It's a natural phrasing pattern to have the initial part of historic unstressed, and then the h is effectively dropped even if you're not speaking an h-dropping dialect; this way "an historic event" follows as correct.

We simplify a lot of sounds in unstressed parts of words, but because we know and "see" the spelling, our brains can keep the illusion of the whole word going for us.

Well you leave the C++ realm (execution model), as you should with UB and it depends on implementation. The implementation of the compiler was such that the two functions are placed after each other in the machine code; and if the first function doesn't return, then you continue executing into the code for the next function.

This is tricky, because we really want language-independent training of skills. We know that self-play type of reinforcement learning is incredibly effective when possible. But at the same time, they are our tools - so we need supervised language training for this reason? It's possible that training just needs to be rebalanced so that RL with rewards is balanced with rounds of language adjustment. And to really make that happen, benchmarks need to score the models on that.

We need to recognize this as a failure in training. It did some useful stuff but it can be much better. A training signal is likely missing.

Not only in America

But if I remember correctly, he gained recognition for his achievement rather quickly after posting.

If we read the link, it has a section called Gold Standard: comparator and external checkers, and comparator is how OpenAI has gone about checking their lean proofs.

I wonder what Demis Hassabis thinks about this. I thought he cared a lot about mathematics.

You should read his comment from wednesday here: https://terrytao.wordpress.com/2026/09/07/finite-time-blowup...

And yes, at one point it gives you right about them taking advantage of him. He sat down for an interview and found later that they just used the "best clips" out of it for an OpenAI ad.

Well OpenAI took one step on the back foot at least, withdrawing from sponsoring this math hackathon event https://xcancel.com/danintheory/status/2098125701782372640

My university has an agreement with Microsoft copilot. We can log into copilot in many ways, and it's only if you log in the correct way that you get the "Enterprise Data Protection" copilot version, with a green shield symbol. There are many ways to go wrong here!

It seems like we're not supposed to care about the code quality then? I guess that's the compilers argument. But I'm not ready to give up the code just yet.. These LLMs don't even have a stable interface, they change every few months in how they interpret our prompts and tasks.

They have plausible deniability on that one: not making any profit.

Is that something we have credible evidence for? Do they serve a better model when artificialanalysis (the benchmark site) is making the requests, and so on?

Interesting, and if you don't mind, where do we put humans (and superhumans like Magnus Carlsen)? I think they have a heavy evaluation function and do shallower search.

There's this whole discussion going on about agents being more independent now. They don't follow instructions so well, they continue until the problem is done (sometimes too long), they don't ask the user for feedback.

That is a kind of benchmaxing: they are made to complete benchmarks tasks and one-offs well, and no longer work well in tandem with the user.

Regardless what you call it, it's a divergence between what the power user wants and what the model developers want, I think.

But the very fact that you go to "chatgpt.com" and write to them; "Dear Diary, today I thought.."; there is no reason they would not receive and process your data, unless explicitly promising not to (which also requires us to trust them).

The fundamental rule in this case is that if we offload our data to a cloud provider we can assume they read it, if they can, unless they promised very clearly they will not.

And just a day later we have the tech report available for V4.1: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...

Isn't one problem that it's hard to determine what equivalent compute is, for CPU search vs a neural net based engine like AZ or Leela?

V4 Flash was one of the big events of this year, and its already retired and replaced by V4.1 Flash.

The main claim, that "they do not seem to have any mathematicians capable of understanding what they put out", was also corroborated by Sebastien (OpenAI) who explained they don't have any experts on Navier-Stokes.