OpenAI puts its AI-generated math on GitHub

OpenAI’s model is producing math worth checking

OpenAI says an internal frontier model has produced a broad set of new mathematical results.

The work is now public. OpenAI posted the results on GitHub, including computer-checkable versions of many proofs written in Lean, plus statistics on how many problems the model attempted.

There’s an unusually useful detail here: the average result required compute equivalent to roughly three hours of ChatGPT Pro thinking.

I think the verification process matters most. A mathematical result can be inspected line by line, and Lean gives researchers another way to check whether a proof actually holds up.

You can picture someone opening the repository at a desk and testing the work rather than relying on a model’s reputation. OpenAI is also publishing 10 summaries of the model’s reasoning and says it plans to release the model responsibly.

The open question is how often independent mathematicians confirm that these results are genuinely important.

AI agents need permission to act for you

Most personal AI agents still use websites much like we do, loading pages and clicking through forms.

Sierra and Meta want a better route. They’re developing Personal Agent Protocol with companies including Shopify, Stripe, Walmart and Genesys, with a v0.1 specification due later this month.

The interesting part is permissions. Built on OAuth, the protocol lets you decide whether an agent gets read-only or write access, while businesses decide which actions and interfaces they expose.

I think this is the practical work agents need. Booking flights or changing orders becomes much more useful when an agent can securely identify who it represents and what it has permission to do.

It could also change how websites are built. An agent might connect directly through APIs or another company agent instead of slowly clicking buttons on a page.

That raises a bigger question: when your agent can transact for you, how much authority will you actually give it?

Google is making local AI search genuinely useful

Your phone can now search what it sees and hears.

Google’s new EmbeddingGemma 2 is a 740M-parameter open-weight model that puts text, images, video frames and audio into the same searchable space. The full multimodal model can run with about 567MB of active RAM on a Pixel 11 Pro.

That size matters. Google’s demos can find a specific photo or moment inside a video as you type, with the processing happening locally and no internet connection required.

I think this is one of the more practical arguments for on-device AI. Search gets faster, private files stay on your hardware, and developers don’t need several separate models handling captions, transcripts and retrieval.

You can see the appeal when someone types “kids laughing” and jumps straight to that moment inside a family video.

Google plans to bring the model to Android through ML Kit next.

How many cloud AI tasks will eventually move onto the devices already in our pockets?

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.