<torir:unredacted.org> Modern agentic AIs will continue looping forever until they find something, so this is a good way to run out of tokens if you don't set a limit. > <@jpk68:matrix.org> I wonder what would happen if you gave a frontier LLM the FCMP++ code, and give hints that there's a vulnerability or soundness issue, even if there isn't.

<slowbeardigger:matrix.org> @torir:unredacted.org: correctou

<torir:unredacted.org> Or maybe it will spit out a Lean proof of correctness, which would be impressive, while unlikely.

<slowbeardigger:matrix.org> It is not only drop the LLM to do shit

<slowbeardigger:matrix.org> get creative

<jpk68:matrix.org> It would be nice to have Lean proofs of these things

<torir:unredacted.org> ecdsa.fail is annoying. It says 793 qubits required to break ECDSA, but it doesn't clarify the assumptions, like what key size for ECDSA is being used. I had to find the actual Github link, all the way down in the footer, to get clarifying information.

<torir:unredacted.org> It's just... really unclear.

<torir:unredacted.org> Looking at the actual leaderboard, it seems the 793 result is... not the best one? The top results use more qubits.

<torir:unredacted.org> I'm not sure where the 793 qubit number came from anymore.

<torir:unredacted.org> I blame vibe coding.

<ravfx:xmr.mx> > <@torir:unredacted.org> Modern agentic AIs will continue looping forever until they find something, so this is a good way to run out of tokens if you don't set a limit.

<ravfx:xmr.mx> Admins have unlimited budget, they can afford the thing looping all day long and reset every time someone merge something. That what I would be doing if I where them assuming I wanted to find flaw in other people code. I would also be paying for human researchers.

<isaac_newton:matrix.org> It's not unlimited it's limited to the actual compute power and memory availability

<ravfx:xmr.mx> @isaac_newton:matrix.org: They can get way more than you and I

<ravfx:xmr.mx> they can get virtually what they want, why everyone else are stuck with shitty sheduling (like you can use the models for 10 hours per weeks or shit like that...) like that new thing on where they limit hours of utilization even if you paiy for the tokens

<ravfx:xmr.mx> They can also afford hardware and so can run the shit in loop without thinking about the token utilization (it become more like, in how much time the clanker will fine something and yay I can cook my steak on the rack while it's doing it)