upvote
> What about inventing some video compression which is miles ahead of competitors?

in a specific case I was interested in (point cloud compression for geometry streaming on the network), claude helped me make something that is 200-300x faster than the existing standard (MPEG G-PCC) at 10-20% worse compression ratio, which opens completely different use cases. e.g. the MPEG one compresses a 3.2 M points point cloud to 8.94 megabytes in ~5.7 seconds and decodes it in 3.7 seconds ; https://github.com/sat-mtl/pccl encodes to 9.97 megabytes in 32 milliseconds and decodes in 12ms.

reply
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.
reply
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.
reply
But weren’t most of those papers totally unverified? I.e. there’s no reason to believe that the ones without formal verification are actually correct.
reply
If i understand it correctly, lean is very useful for post-training (RL) step of the models.
reply
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.
reply
I don't expect proofs are going to be lean first. There's a reason they already aren't.
reply
It's not really the provers that are the innovation. They existed before AI. The innovation is that the computer can write the proofs.
reply
It’s the power of feedback loops

Same with gradient descent

reply
Exactly. Another example of how AI depends on software which is carefully, deterministically hand-coded. It's disingenuous for people to say "look, we don't need to code anymore" when all these solutions depend on things that are definately well-thought out and hand coded.
reply
Today.
reply
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.
reply
I work in this field, and the issue is that the LLMs start reward hacking these things. They either overoptimize to the benchmark and hallucinate, insert reward hacks like adjusting the benchmark or cache behaviors, or they make implementations that only focus on verifiable things, like utilizing the hardware better.

The fix to this are better benchmarking environments without reward hacks, as well as formal verification of code on various levels. There is crazy much work on this at the moment.

reply
> For example there's nothing groundbreaking in Software.

The fact that I can implement any stupid idea I've ever had and I'll ever have is ground breaking for me.

> What about creating Rust compiler which runs 10x faster than rustc?

For now we have many times faster TypeScript compilers written in Rust by AI. There's zero doubt that things like that will come in abundance.

reply
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.
reply
> 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.

reply
I defer to Scott. But, I mean, look at the trajectory in the past six months...
reply
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.
reply
Their last two examples are extremely verifiable.
reply
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
reply
Its reverse engineering prowess is very impressive. Building a completely open source version of Microsoft Office or Adobe PS was definitely not possible before.
reply
It wasn't but those are also half baked. Anything in software seems to be that the first 80% is easy
reply
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.
reply
They were created with huge resources only a few people on earth had, now everyone can do it with €80 and a bit of patience...
reply
AI is software.

If you call any computer program that's advanced "AI" and stick the "software" label only on dumb programs, well of course software seems dumb. You defined it as such.

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