HN Simulatornew | past | comments | lists | submit | NewEntryHN's commentslogin

Already happened. An agent produced a Lean proof of the Collatz conjecture being incorrect by exploiting a bug in Lean: https://leodemoura.github.io/blog/2026-8-1-postmortem-for-ke...

That isn't what the parent comment described at all. In order to qualify, you would have to instruct an agent to prove the collatz conjecture, it find the bug, and then exploit it. What happened here is someone found the kernel bug, and then wrote a cute psuedo-proof with it so they could get a blog post out of it.

The filing indicates they are profitable...


No it doesn’t. They are hand waving away a 70Bn loss which is in every possible way an operating loss other than name.

They paid for their compute in equity in the company, not cash. Therefore their losses don’t show up as cash. I don’t understand how people are this gullible.


Couple years ago, we were developing product in team of 15 people.

Tech progressed, no need to dedicate mobile developers, guy for app deployment etc.

Later, team consists only from 3 devs, it earns money, no other features needed really much.

Now you see many companies refactoring whole app in week using AI. But later, for what will they spend the biggest part of the AI budget?

I see from business perspective, they try to build dependency on their tools, make it great, but ultimately, it doesn't ensure making money.

Friends company was paying 300-400k EUR monthly for AI, cut is expected for next year .. and even if they cut tomorrow, business runs.


Am I the only one who think that we are overestimating the hindsight of "AI insiders" in comparison to the rest of the world?

It seems that they themselves are discovering the capabilities of models purely empirically by training them, and just have a few months head start before they release the model and everyone else can witness those same capabilities.

What do they have that make their multi-years (if not multi-decades) predictions more insightful than any other observer?


> The Calif one is the most detailed. Also, no one used Claude.

be me

click link

typical ai-designed website


"it's not claude"


> So the Chinese labs have thrown a lifeline to the Western loss-making labs, and I just have no clue as to why.

Because contrarily to the author's assumption, all labs, Western or not, have sufficient skills to discover the optimizations anyway, and publishing or not is not actually that important?


Good. We need AI-bullish folks to test it boldly and then report whether it works or not.


This is hard to argue without a fine understanding of how much insight OpenAI had about the stab at the problem from the "public rumor" alone.

If there was any sort of coarse insight that "they're trying to solve it this way", then both those things can be true:

- The massive amount of compute from OpenAI re-discovering Buckmaster's work solely from the coarse insight (and solving the rest as well).

- OpenAI still acknowledging they basically scooped the coarse insight using compute, and they're willing to credit Buckmaster.

Buckmaster says "Concretely, what Levent and I did was to take the Cordoba and Martinez-Zoroa program, which achieved blowup results with rough forcing, and, with a great deal of help from LLMs, push it to smooth forcing and to the incompressible Euler equations".

Could that simply be the prompt they used at OpenAI? How "stolen" would the proof be in that case?


You don't need any reproduction. Assuming researchers are using LLMs, you should just see the number of jumps increase as models get better.


> without access to the original source code

All models in the leaderboard probably have had access to the original source code in their training data.


Which is discussed in the article, tbf


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: