Latinum Frontier Mathematics Research Lab

Excited to share a new podcast, this time with Fabrizio Montesi, Professor of Computer Science at SDU and lead maintainer of CSLib (Lean library for CS) We kept landing on a question most AI code talk skips: we can prove software correct, but correct against what? Thread👇

🏅 @LatinumAI (Honourable Mention) Smart contracts are still mostly secured through informal reviews and expensive audits, leaving critical bugs to slip through. Latinum solves this with a formal verification toolchain that compiles Lean 4 to Solana SBF, proving contract properties at build time before code can ship. Their demo shows a verified vault contract deployed onchain, with the compiler refusing to build when a balance leak is introduced. In short: they’re making it possible for @solana developers to write contracts that can mathematically prove they work before they go onchain.

Ireland takes home TWO winners and one honourable mention from @colosseum! 🏆 Grand University Award: @iochaindev - a decentralized AI inference marketplace (@Ericxin123) 🏆 @MentionedMarket - a social prediction market to trade words that get said across any form of media (@taylorferran & @Murraymiint) 🏅 Honourable mention: @LatinumAI - using LEAN programming language to write smart contracts on @solana 🏅 Shoutout to Superteam IE member @YellowFishZ on @craftsdev's W also. Small country, big at heart 💚 @solana @superteam

@JasonRute @leanprover Just the compiler. I added cross-target support to the LLVM backend and wrote a small runtime to replace the host lean.h/libc. The interpreter is untouched, so eval and tactics still run on the host. An external compiler bench would be great, agreed.

@LatinumAI Did you rewrite the interpreter too or just the compiler? I guess now @leanprover needs an external compiler bench (alongside their external kernel bench).

Why are formal languages not used as programming languages? Most were built only for mathematics. Lean lets you write real code, but its compiler assumes a garbage collector, threads, and an operating system underneath. So I rewrote it.

In Lean, you only write the theorems. The AI writes the code and the proofs. Full autonomy, no risk. Code becomes like assembly. No human review needed.

Why this matters for AI: Today, people prompt Claude Code. It needs supervision, fakes tests to pass, bugs slip through.

Full episode: 🎧 Spotify: open.spotify.com/episode/6sI79r… 📺 YouTube: youtube.com/watch?v=Czir_2… 🌐 latinum.ai

And the closing line that’s making me dust off my old maths books :) "We're all going to be math engineers. The math is correct, but what is the math saying? That's going to be the crux of the problem."

Excited to share a new podcast, this time with Fabrizio Montesi, Professor of Computer Science at SDU and lead maintainer of CSLib (Lean library for CS) We kept landing on a question most AI code talk skips: we can prove software correct, but correct against what? Thread👇

@dharmatrade Thanks @dharmatrade. Updated all videos now

@LatinumAI Hello 🙋‍♂️ YouTube won't allow people to save the video to a playlist because the video was marked as "for kids" when it was published. You have to mark it as "not meant for kids" to allow users to add it to playlists. (I know, it sounds weird...)

Just released a podcast with @AlexKontorovich, Distinguished Professor of Maths at Rutgers. He has also been a forward looking thinker on how mathematics will evolve alongside AI. He has shared his forklift analogy for AI & learning here before, but it’s too good to not reshare!

And on what's coming: "There's plenty I can say AI can't do well now. But something they'll never be able to do, I always come up short. I'm not able to articulate what I do that these systems won't ever be able to do."

On the maths community and AI startups: "We're very non-confrontational as a culture. If someone is claiming things and you don't agree with them, you just don't invite them to your next conference. What we as mathematicians need to do is not just play defence."