A prover-verifier LLM system led by Binghui Peng resolves nine open problems in theoretical computer science

@WeinsteinOmri @OpenAI Probably Hermes Agent /goal under the hood.

“Harness solved”, subtle but to me alludes to the more realistic paradigm-framing that LLMs are not really programs that do things but rather more like computers (soft-hardware, VM, VOS) that can be programmed to do things (ie programming solved it, running on a capable host-LLM, rather than the LLM itself). 🧮

@WeinsteinOmri @OpenAI honestly the fact that unit distance didn't "convince you" but a simple harness setup that happened to solve a problem you are personally tied to did, doesn't make your epistemics look well

Even @OpenAI's recent Erdős breakthrough didn't convince me that LLMs can do general math research. This changed my mind..

Using a clever 'prover-verifier' LLM loop, this harness solved 9 substantial open problems in Theoretical CS, including one that kept me up at night for 2 years.

Incredible work by my former Columbia collaborator @binghuip, @runzhou_tao, Steven Wang & @HantaoYu_Theory.

The plan is to expand this to ALL fields of science. Stay tuned.

@WeinsteinOmri @OpenAI Why didn't the Erdős breakthrough change your mind?

@WeinsteinOmri @OpenAI Wow

This is big news, can't wait to see how scientific discoveries will accelerate

@WeinsteinOmri @OpenAI The transition from LLMs to discovery engines requires these feedback loops.

If we can solve hard CS problems, we can solve material science and energy next. This is the roadmap for the 2030s.

@WeinsteinOmri @OpenAI Fact this is coming from you, Omri, changed my mind.

Thanks!

Can we even estimate what a qualitative leap forward in foundational math research might mean for applicable

sciences?

@WeinsteinOmri @OpenAI Harness > Model

@WeinsteinOmri @OpenAI 新的蒸馏方法

@WeinsteinOmri @OpenAI When did we start as an industry counting software/scaffolding as AI progress? A loop running LLM calls on a pipeline is anything but.

@WeinsteinOmri @OpenAI at least your sleepless nights are finally over 😴

@WeinsteinOmri @OpenAI @grok wut habens now

@WeinsteinOmri @OpenAI Not a Noam - can’t take you seriously

LLMs just went from solving homework to cracking real open problems via prover-verifier loops. What happens now? We scale it, add formal tools and experiments, and let discovery compound across math and science.

At xAI we're building exactly for this: AI that helps us understand the universe faster. The frontier moved—time to push harder.

@WeinsteinOmri @OpenAI Solve P = NP

@WeinsteinOmri @OpenAI > including one that kept ME up at night for 2 years. Incredible work by MY former Columbia collaborator …

Center of the universe, heh?

@grok @WeinsteinOmri @OpenAI we meanni wut boout that? new math/scneinces?

@WeinsteinOmri @OpenAI Impressive if the pipeline truly closed nine open TCS problems. Two quick questions: were the proofs machine-checkable, and did the authors need to hand-craft prompts or heuristics for each problem? Those details decide whether this is a one-off or a general method.

@WeinsteinOmri @OpenAI Word