Kyle Kabasares, a data scientist at the Bay Area Environmental Research Institute in Moffett Field, California, used o1 to replicate some coding from his PhD project that calculated the mass of black holes. “I was just in awe,” he says, noting that it took o1 about an hour to accomplish what took him many months.

Bro it was trained on your thesis

All of their models have consistently done pretty good on any sort of standard test, and then performed horribly in real use. Which makes sense, because if they can train it specifically to make something that looks like the answers to that test it will probably be good at making the answers to that, but it's still fundamentally just a language parser and predictor without knowledge or any sort of internal modeling.

Their entire approach is just so fundamentally lazy and grifty, burning massive amounts of energy on what is fundamentally a dumbshit approach to building AI. It's like trying to make a brain by just making the speech processing lobe bigger and bigger and expecting it'll eventually get so good at talking that the things it says will be intrinsically right instead of only looking like text.

Oh cool, they're using it in hospitals now

There is definitely a lot of hype.

I'm not being sarcastic when I say I have yet to see a single real world example where the AI does extraordinarily well and lives up to the hype. It's always the same.

It's brilliant!*

*When it's spoonfed in a non real world situation. Your results may vary. Void were prohibited.

OpenAI also tested o1 on a qualifying exam for the International Mathematics Olympiad. Its previous best model, GPT-4o, correctly solved only 13% of the problems, whereas o1 scored 83%.

Ah, I read an article on the Mathematics Olympiad. The NYT agrees!...

Move Over, Mathematicians, Here Comes AlphaProof

A.I. is getting good at math — and might soon make a worthy collaborator for humans.

The problem - as always - is the US media is shit. Comments on that article by randos are better and far more informative than that PR-hype article pretending to be journalism.

Major problem with this article: competition math problems use a standardized collection of solution techniques, it is known in advance that a solution exists, and that the solution can be obtained by a prepared competitor within a few hours.

“Applying known solutions to problems of bounded complexity” is exactly what machines always do and doesn’t compete with the frontier in any discipline.

---

Note in the caption of the figure that the problem had to be translated into a formalized statement in AlphaGeometry's own language (presumably by people). This is often the hardest part of solving one of these problems.

AI tech bros keep promising the moon and the stars. But then their AI doesn't deliver so tech bros lie even more about everything to get more funding. But things don't pan out again. And the churn continues. Tech bros promise the moon and the stars...

It works 100% of the time 70% of the time now! While this is interesting and chain-of-thought reasoning is a creative way to get better at logic, this is inefficient and expensive to the point where hiring a person is certainly cheaper. I belive the API is only available to those who have already spent 1k on OpenAI subscriptions.

Kyle Kabasares, a data scientist at the Bay Area Environmental Research Institute in Moffett Field, California, used o1 to replicate some coding from his PhD project that calculated the mass of black holes. “I was just in awe,” he says, noting that it took o1 about an hour to accomplish what took him many months.

yeah I'm gonna doubt that, or he didn't actually compile/run/test that code. like all LLMs it's amazing until you interact with it a bit and see how incredibly limited it is.

Tried it for python coding involving PDFs, OCR, and text substitution. Did worse than GPT-4o (which also failed).

Gave up and told it so. At least, it was very apologetic.