Thank you for your reasonable reply. There is so much bad faith arguing on the Internet and some people are very motivated to prove that mathematicians are useless. Even if future AI becomes super-human at explanation, I hope we will always have true human experts

It is definitely the difficult part, because it is not amenable to RLVR. Steven Pinker: "the main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard". Your intuitions of what is easy and hard are worthless.

Okay I'll just wave my magic wand and make AI good at stuff it's bad at. Oh wait, why do I need to use MY magic wand? Why don't you use YOUR magic wand to do it?

Because if it could, OpenAI would output readable proofs. But they consistently fail to do so.

We *already have* AIs that are good at proving and bad at explaining. Why? a) generating a proof is a task highly subject to self-guided reinforcement learning and explaining it is not and b) humans prefer to learn from other humans. Nobody knows if these hold in the future

@tallmarkmusic "Eat It Anyway" has just one more show in Vancouver. Highly recommended! September 19 Just $20 He's been getting rave reviews across the country. vancouverfringe.com/events/t…

Move 37 was new knowledge. Indisputably. Also Gödel’s theorem has nothing to do with the creativity needed to come up with new knowledge. Gödel’s theorem in no way distinguishes between humans and AI and is therefore irrelevant to this discussion.

Serious question: how would you know if you had stumbled across an unprovable statement? Do you mean that you seldom stumble across statements that persistently defy proof?

What is happening in the world where Slate and @alex_kirshner can publish a piece about cross-continental travel for “day trips” and not even consider the environmental aspects? slate.com/technology/2025/07… When did we decide that the planet is going to be fine no matter what we do?

It is not true that DeepSeek-R1 cost ~$5M to train. That was DeepSeek-V3

"We will always find solutions to dangers" is as pure an expression of faith as "God will always protect us from evil."

So I heard that you can get banned from this bastion of free speech by pointing out that @elonmusk admires Nazis (Trump, in particular). Let's see if that's true.

@skdh What does this mean? arxiv.org/html/2409.03680v1 Negative amount of time?

It's a dynamically changing model called `chatgpt-4o-latest` and not recommended for production use because it can change with no notice.

It seems to be!

@MIT_CBMM Why did videos disappear? cbmm.mit.edu/video/cbmm10-pa…

Congratulations @jasontlantz, muselab.com/ looks great! It sounds like you've pulled off some real magic moving Ford/VIIZR to the Produce Development Model. Much more impressive than most case studies I have read.

Ugh...Twitter seems to have swallowed the "https:"

Chrome also visits that site happily. It's paulgraham.com that we're talking about. Try this: $ wget paulgraham.com ERROR: no certificate subject alternative name matches requested host name ‘paulgraham.com’.