Remix.run Logo
▲ vbezhenar 6 hours ago

What I found surprising. How many absolutely groundbreaking breakthroughs with AI are in Mathematics. And at the same time I can't remember anything similar in other fields. For example there's nothing groundbreaking in Software. Yes, Claude can write some PoC code faster than me. Great. What about creating some software which wasn't possible before? What about creating operating system without a single bug? What about inventing some video compression which is miles ahead of competitors? What about creating Rust compiler which runs 10x faster than rustc?

I think the closest thing is formal verification of Signal protocol, I guess. Which is insanely cool, but still not as cool as millenium problem.

▲zesterer 5 hours ago | parent | next [-]

The real innovation here, IMO, is theorem provers. Without theorem provers, the LLM brute-force-by-bouncing-off-the-guardrails approach just doesn't work to begin with and the whole thing spins off into hallucination and dodgy reasoning, just as we see in non-trivial AI software. Once you've got some magic oracle that gives you a binary yes/no answer to every attempt at a proof, your search strategy can be enormously simplified.

▲j2kun 4 hours ago | parent | next [-]

I also thought that, but I learned that most of OpenAI's results in their proof dump did not use theorem provers in the loop. They translated the papers to lean afterward.

▲zone411 3 hours ago | parent [-]

Right, that's how all these proofs (including a few of mine) are done. Given the large-scale AI efforts already underway to formalize proofs in Lean, there could be some very interesting gains once proofs are Lean-first. For now, though, formalization happens afterward.

▲famouswaffles 2 hours ago | parent [-]

I don't expect proofs are going to be lean first. There's a reason they already aren't.

▲jojomodding 3 hours ago | parent | prev [-]

It's not really the provers that are the innovation. They existed before AI. The innovation is that the computer can write the proofs.

▲rcpt 5 hours ago | parent | prev | next [-]

Math is pure logic. A game like chess or go. It's not bottlenecked by the physical world so AI can iterate at maximum speed.

▲oedemis 6 hours ago | parent | prev | next [-]

these tasks you mention are very long-horizon in context... i think if latent long-term memory or continuous learning approaches are getting better, we also see this

▲rcbdev 6 hours ago | parent | prev | next [-]

Its reverse engineering prowess is very impressive. Building a completely open source version of Microsoft Office or Adobe PS was definitely not possible before.

▲vbezhenar 6 hours ago | parent [-]

Microsoft Office and Adobe Photoshop were literally built before. Recreating their copies is not groundbreaking. That's what I'm talking about. AI in software right now feels like mostly about recreating things we already can do, but cheaper, with questionable quality and unclear maintainability.

▲hyperbovine 6 hours ago | parent | prev | next [-]

The recipe seems to be, pick a problem whose solution is highly verifiable, then throw the equivalent of ten billion hours of expert-level thought at it. If you can narrow the search space a lot, all the better. P != NP is sure to follow soon enough.

▲jonahx 2 hours ago | parent | next [-]

> P != NP is sure to follow soon enough

Not an expert, but from comments I've heard by experts P != NP is much further out of reach than Riemann. Eg, Scott Aaronson:

> First, almost no one in theoretical computer science—not counting the cranks, whose missives fill my inbox every week!—spends their time directly trying to prove P≠NP. Why not? For the same reason why you wouldn’t embark on a manned mission to another galaxy, if you hadn’t even set foot on Mars yet.

▲hyperbovine an hour ago | parent [-]

I defer to Scott. But, I mean, look at the trajectory in the past six months...

▲jonahx 35 minutes ago | parent [-]

Right. Just saying the fall of these "step function" type problems aren't a forgone conclusion yet. It could easily be the case that some problems (who knows which?) are in an entirely different dimension of complexity, and would be out of reach of even much, much more powerful AI.

▲conradkay 5 hours ago | parent | prev [-]

Their last two examples are extremely verifiable.

▲Rover222 3 hours ago | parent | prev [-]

Maybe you should start directing Astra to make one of these breakthroughs you're waiting for, see what happens