zarzavat

That's for running your own code. Not for running other people's code.

Type stripping cannot ignore tsconfig.

TypeScript has different semantics depending on settings in tsconfig e.g. useDefineForClassFields

The "bubble" is being driven by blockbuster earnings. The market is irrational but it's not brain dead.

Surely you must reset something? At the very least box-sizing

OpenAI is trying to build a business to justify their valuation, while they pursue ASI at the same time.

Anthropic is just trying to get to ASI and relying on the blind faith of investors that there will be a return at the end.

What matters is which strategy is more effective for attracting investor capital, because neither can survive without it.

I'm saying that the gains were not a result of 1 week of post-training 6.0 Sol.

It seems that 6.1 Sol is closely related to Astra. As for whatever 6.0 Sol was, it's anyone's guess. I'd speculate that 6.0 was just 5.6 with post training. It's possible however that it is the same model but with severe performance issues resolved so they bumped the version number.

Every writer has a style and if their style is strong enough you can identify the writer by reading enough of their works. AI style is very easy to spot.

There may have been some human involvement with this website but they certainly didn’t edit it for style. Unfortunately if you don’t edit for style then people will default to the assumption that it’s entirely vibecoded.

It's more like they released GPT 6 Sol too early because they were under pressure and now they are releasing the real version. You cannot do anything more than minor post-training in a week.

It’s not that planning is dead, but rather that planning has outgrown the simple “Plan Mode” feature as models have become capable of taking longer turns.

Why commit copyright infringement when you could commit criminal fraud instead?!

Presumably it's so it can be used as a subdependency for setting up an attack in a popular, legitimate package, e.g. via a pull request. The code in the legitimate package would not arouse suspicion at all.

Hindi uses the sound /ʋ/ which is a sound with elements of both v and w. This means that to many Indians, v and w are allophonic: they are heard as the same sound.

An example of allophony in English is the aspirated /pʰ/ in pin and the unaspirated /p/ in spin, which to a speaker of Hindi and many other languages would be considered two distinct sounds.

Would be remiss to not mention W's biopic: https://www.youtube.com/watch?v=sg2j7mZ9-2Y

Yes and the critical issue was tree shaking. Nowadays we have tree shaking so it doesn't matter as much but in the past people preferred small single function packages because they had less impact on the download size.

More sensible, yes... but less ideologically pure.

The hand wringing is premature. Thus far AI has only shown a superhuman aptitude for brute forcing proof of existence:

- Disproof of the Jacobian conjecture by example

- Construction of a non-sofic group

- Existence of singularity in Navier-Stokes

Mathematical conjectures tend to be universally quantified, especially those conjectures that are used as building blocks (e.g. RH). If anything, AI models are currently performing a useful service by disproving false conjectures, a kind of mathematical weeding.

The good news from the last couple of years of coding agents is that while models have become more persistent and knowledgable, their creativity (defined as being able to escape their training distribution and synthesize completely novel ideas) is improving at a much slower rate.

AI will only become a threat to mathematics if/when it develops the capability for creative big-picture problem solving. If that happens, the impact on mathematics will be a footnote compared to the impacts on society at large, since creativity unlocks a host of new economic capabilities.

Because the complete power of Python also includes the power to fuck things up.

Even if you judge OpenAI solely on their public communications it still sounds really bad.

That they heard a rumour that a major open problem had been solved, so they decided to try and scoop the other mathematicians while they were writing up their preprint is extremely unsporting.

Then they decided to exclude an author because of his employer, even though he had used their own products to write the proof!

They haven't necessarily breached any formal ethical rules but their behaviour will lead to them and their products being shut out from the mathematical community.

It's more fundamental than that. AlphaZero is a shallower search with a heavier evaluation function. Stockfish is a deeper search with a lighter evaluation function.

In chess, depth usually wins because of how narrow the search tree is compared e.g. to Go.

The marginal improvements add up over time. Most people upgrade every few years, amortized over a 3 year upgrade cycle it's not really that expensive.

Many mathematicians would be willing to end their careers for $1m. What makes this so sad is that OpenAI spent more than that for this empty PR stunt.

Many people don't like the GPL for exactly that reason. Free software (i.e. copyleft) benefits from copyright at the expense of the wider open source community.

Were it not for copyright then BSDs could take code from Linux and perhaps there'd be less of a monoculture, for example.

AI just solved a millennium prize problem. In a matter of days. Because of a rumor that someone else solved the same problem with AI.

What exactly would AI have to do in order to not be called a bubble?

It also fragments power, which makes it worth it.

I agree and I was being facetious since I know the answer is nothing much. The point is that IBM write all this breathless marketing copy that seems detached from the actual utility of the device.

Although I believe there's a difference with the Manhattan project. The original atom bomb was "useful" even at its small scale. Whereas there's likely to be a lot of time between a quantum computer that can factor 69 and a quantum computer that can factor RSA 2048.

What's the largest semiprime it can factor?

As the recent "proof" of the Collatz conjecture shows, that's not enough in an adversarial context. Human mathematicians don't submit proofs that take advantage of soundness bugs in Lean. AIs do.

Are you talking about Japan? Because when I think of "Asia" strong job protections are just about at the bottom of the list.

Humans are evolved to survive in the wild. We are not evolved for circuit design. Yet we can design circuits because evolution found it easier to develop a general problem solving nervous system than a nervous system which is adapted for every single specific problem a human might encounter.

> Humanity's Last Exam (w/ tools)

This is one of the only benchmarks that actually matters for testing the frontier however. Other benchmarks can be gamed by simply being more persistent, but HLE is a diverse set of open-ended research-level questions. It tests domain knowledge and problem solving skills. Burning more reasoning tokens may help somewhat but not as much as e.g. coding benchmarks.