Post

mr. TIM
@timkellogg.me

the Jacobian Conjecture prompt is 2 pages also. Over half of it is “search and coordination requirements” aaronlou.com/jacobian_cou...

mr. TIM@timkellogg.me

if you read the prompts for these big math problems, they’re non-trivial i like to laugh at how prompt engineering is dead, but goddamn, you still need a working mental model for how LLMs think, and that’s not easy

18 reposts156 likes

⊣
@lefttack.bsky.social

The thing is, we don't know whether "solve this wikipedia page" wouldn't work equally well. Is a frontier model really incapable of googling what its peers have done and what's best practice for that kind of problem of its own initiative?

Pedro Vezza
@pedro.vza.net

That’s crazy. The Anthropic guy copy/pasted the OAI cycle double cover prompt including whatever multiagent v2 means verbatim and it worked. Nothing wrong with the approach, it’s just remarkable that an undergrad could have done something similar.

Will
@thealpacalypse.bsky.social

this also stood out to me, I was gonna say the "Do not allow one approach to dominate merely because it gives elegant reductions." snippet was verbatim from that paper

M
@0xcd16.bsky.social

For the Erdős problem back in May (apparently one of Erdős' favourite problems and pretty well known), the prompt was fairly simple. But my guess would be that the more sophisticated prompts are capable of solving more problems? bsky.app/profile/0xcd...

M@0xcd16.bsky.social

Yeah, but did you look at the prompt? The hint he gave ChatGPT wasn't used and the rest was "here's the problem, don't look online, try something new and clever", and then it spat out the solution in its first reply. Great job human, impressive prompting skillz. chatgpt.com/share/69dd1c...

chatgpt.comChatGPT - 1196-1Shared via ChatGPT
mr. TIM
@timkellogg.me

implement a hash table in LLM prompts: “Maintain an explicit registry of approach families. Group agents by the mathematical idea they are using, not by superficial wording. Redirect agents when too many converge to the same incomplete route.”

mr. TIM
@timkellogg.me

seems like prompt engineering has only gotten harder and more advanced it used to just be woo belief in arbitrary phrases, but after subagents, idk, this looks like actual legit engineering maybe prompt engineering is *more* relevant now??

Rémy
@xowap.dev

It's literally magic for machines: use the proper misdirection and "optical" illusions to manipulate it into doing the right thing at the right time despite its worst instincts The more parameters available, the more residual cognitive capabilities you can keep after your prompt