Tyyps

I think you are missing the points of my argument, my argument don't stand on AI isn't capable of discovering new techniques, as I don't believe in new techniques from the sky. My argument is about any targeted goal based AI (which to the best of my knowledge is the case for LLMs as used now). My point is, if there exists a computationally bounded path from existing work that led to solving a conjecture and if the goal of the AI is to solve this conjecture, then alternative path that would have led to new discovery will be dismissed on the way (or lost in the computational trace if you prefer), leading to the conjecture being solved but maybe closing forever/for a long time new paths. I don't see how you could have as a goal to explore alternative path without a good metric of what is a good alternative path (like rating a chess position), which to me, seems unlikely to exist. If you don't have such metric then you would have a clear exponential blowup. More like a percolation problem if you prefer, a neglected approach might have introduced a concept that would make further discoveries accessible. Missing that concept could therefore leave a whole region unexplored, not just one branch of one proof.

I guess you can see this as an exploration problem, in pure maths, while the goal is to solve a conjecture, the limitation of humans on pure computational power led to the exploration of alternative paths. Sometimes, these paths weren't leading to solving the initial conjecture but opened new idea and new direction. Sometimes a less direct but more humanly natural path was taken to solve the conjecture which also led to new and humanly understandable questions. In some ways solving the question wasn't the most important part of the work, as this doesn't have direct impact on our life (as I saw people comparing this with drug discovery), but the path leading to the solution raised new conjectures and techniques that further developed the field.

I have a really hard time reading AI proof so this might be a biased statement, but most of them feels like having a superpowerfull machine, that would have bruteforce all the possible words of finite length in your logical syntax. You have the path to the solution, using tools that where already known and even direction that where abandoned because they seemed to fail for our human brain. But at the end, as a mathematician, you don't learn anything that is really new.

To me this is the main risk with AI and in general the one most mathematican try to explain but fail, we might miss a lot of alternative path that would have raised more interesting questions (I think this is already more or less what is happening). On top of that, we will run out of mathematicians as no one wants to pursue a career in the field anymore.

Full title: Meta Paid Hundreds of Contractors to Pretend to Be Teenagers While Barraging Its Competitors’ AI With Disturbing Content

The model is constantly switching to Opus for me, this is kinda unusable sadly.

There is no clear evidence that the risk of "a practical post quantum computer would arrive in the next 5 years" is greater than "post quantum scheme X is broken" for any scheme X. The only way to go is hybridation and it is quite hard from an engineering point apparently.

I think it is pretty direct from my comment that if you use a hybrid approach (done correctly) you can rely on the hardness of dlog based assumption and therefore my comment on potential weakness of PQ assumptions can be ruled out. In this way we disagree that rushing PQ is the appropriate choice if it rules out dlog based security.

> He's also pointing out that the only scenario in which hybrid is of benefit is one in which crypto related QC remains either relatively ineffective or extremely expensive in the medium term. Since that assumption is looking increasingly suspect it calls into question the point of hybrid to begin with. In the face of cheap QC hybrid adds zero value.

This is exactly what I'm pointing out as extremely dangerous. My take was that the risk of seeing a quantum computer breaking dlog in a near future isn't stronger than breaking PQ assumptions in a near future.

He is assessing that the risk of seeing a quantum computer break dlog cryptography is stronger than the risk of having post quantum assumptions broken, in particular for lattices.

One can always debate but we have seen more post quantum assumptions break during the last 15 years than we have seen concrete progress in practical quantum factorisation (I'm not talking about the theory).

I agree, but the blog post was specifically ruling out hybrid approach.