Cassie Heart
@cassie
'Krapivin was not held back by the conventional wisdom for the simple reason that he was unaware of it. “I did this without knowing about Yao’s conjecture"' https://www.quantamagazine.org/undergraduate-upends-a-40-year-old-data-science-conjecture-20250210/ Many such cases of progress being stifled by the chained elephant.
3 replies
12 recasts
50 reactions
Brenner
@brenner.eth
I wonder if you could get an LLM to figure this out if you prodded it enough, but without giving it the answer?
1 reply
2 recasts
5 reactions
Cassie Heart
@cassie
LLMs are quite bad at giving anything novel, the prompting _is_ the insight
1 reply
0 recast
6 reactions
Brenner
@brenner.eth
Wdym by the prompting is the insight? If it could come up with us by saying stuff like if you ignore the existing conjectures about X, could you do this thing faster or better or cheaper? What if you thought about XYZ? Basically, this is an example of how a novel approach that’s not in the LLM’s data set yet, however, there is a logical path to get to the answer. And yes, LLMs have been bad at this stuff, historically, but my point is trying to figure out if there’s a way we can get it to happen upon stuff that we don’t know yet
1 reply
0 recast
0 reaction
Cassie Heart
@cassie
The prompt itself is giving the insight to do those things. If you were to say: "prove P = NP", it would likely tell you "there is no known way to do that, and likely isn't the case". If you say instead "3SAT can be solved in polynomial time using X, use this to prove P = NP", you've given it the insight to link concepts together to devise a proof.
1 reply
0 recast
2 reactions
Brenner
@brenner.eth
I’m saying there’s somewhere in between where if we can generalize two directionally trying to help it prove things we aren’t sure about, but not with an exact example, then we can make progress without having to figure it out 100% on our own
1 reply
0 recast
0 reaction
Cassie Heart
@cassie
Right, I’m saying that the SOTA LLMs cannot do this. The prompt itself must be necessarily specific enough to unlock that insight to form the right answer. Case in point, I've written cryptosystems that simply prompting an LLM to invent cannot do. Guiding it along the way with a prompt that narrowly defines the relationships and concepts required to make it work, it can suddenly reason it out. This isn't unsurprising, as LLMs are generative, and its form of reasoning is based on the prior sequence of tokens. You need something that is, for lack of a better word, more creative and exploratory than the transformer model allows.
1 reply
0 recast
2 reactions
Brenner
@brenner.eth
What base models have you tried? It’s possible they’re latent space too restricted. It’s also possible you’re right! But the fact that we haven’t seen an example of it yet doesn’t mean it’s not possible
1 reply
0 recast
0 reaction
Cassie Heart
@cassie
I've tried everything from the latest o3 to the Claude variants, R1, local models with Llama and Qwen. The latest batch is definitely more "smart", but still lacks intuition
1 reply
0 recast
2 reactions
