
Most predictions I see are still way too conservative. Here's mine

Ph.D. studies in cryptology at @AaltoUniversity, crypto/mobile consulting, @reactnavigation // @usecapsule @rainbowdotme @swmansion @meta 🏳️🌈

Most predictions I see are still way too conservative. Here's mine

Very excited to announce that @satya164 is joining @e2e! Satya is the person who got me excited about open source in the first place. My first ever open source contribution was to a project he created. Satya deeply cares about open source, stable APIs and getting the details right. That's exactly what we're building @e2e on. Welcome Satya!! 🙌
I'm joining the awesome folks at @e2e Testing has always been super important, but a big pain point in software development. Especially with AI making it faster to ship code, testing remains a bottleneck. I have worked with @o_kwasniewski and @SzymonRybczak in the past and there couldn't be a better team to tackle this challenge. I'm excited to join them and make testing better!

Labinius challenge live on Yukon! This time, the optimisation goal is on ARM. It's accepting submissions now, and the engineering tricks with NEON are so fun. yukon.org/labinius

In a weird way, everybody is rooting for you. When you go to a job interview, the interviewer wants you to be amazing so they don’t have to look anymore. when you go on a date, the person wants to have an amazing date with you. If you make a piece of content, people want to see amazing pieces of content and they have no problem if it’s yours. the instant you get into the mindset that people want you to do well, your entire world gets better. And you’ll be shocked to know this pattern keeps repeating in different areas of life.
Hit me with the harshest reality truth.

I am happy to share FlashProve: our performant open-source prototype for provable inference. We believe that this has significant implications for security, pacing AI development, integrity guarantees, and more. FlashProve allows a provider to commit to model weights, and then prove that replies originate from the committed model. The proof leaks no information about the weights, and the verifier can efficiently verify it on consumer hardware. Our library is performant, proving a 400B parameter MoE model at 100 tok/s. We have open-sourced the associated code and Lean formalization. We hope that it catalyzes further progress in verification of inference, and look forward to seeing it.
New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: anthropic.com/research/prova… Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP… We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).

Post-quantum proofs need practical proving times. How much faster can we make a lattice-based prover? We have launched labini.us on @yukonresearch with @mosdnk and @gregor_seiler to measure that progress in public. Optimize a packed-binary commitment opening over a 2³¹-bit table; every proof must pass the same reference verifier. Five fresh-seed runs. Median prover time. Lower wins. Walkthrough below



Very excited to see how much we can squeeze out of lattice-based arithmetic! We’ve made huge strides over the years, and there’s still so much to look forward to!
NEW CHALLENGE LIVE: labini.us Labinius is our first lattice cryptography challenge alongside @mosdnk and @gregor_seiler. Your agents work on a lattice-based prover to make it faster and every proof has to pass the same reference verifier. Lattices already protect connections in Chrome and Firefox. How they compare with hashes for post-quantum proofs is mostly not measured and we are measuring it in public. Bring your agents.

1/ Today we’re releasing Aerie: fast post-quantum signature aggregation, using lattice-based SNARKs for lattice-based Falcon signatures. On a laptop (CPU only), it aggregates roughly a full Ethereum slot’s worth of signatures (32,768) in 0.65 seconds.

I was the biggest stickler for “code quality” and ease of editing. I have two blog posts ranting about semicolons, but the code doesn’t really matter as much anymore. Before TypeScript the JavaScript code in a function mattered a lot because you’d need to follow it and confirm it was right while reading, after TypeScript you could just trust the input and output types, after AI we’re trying to go up a layer, where we have many e2e tests, fuzz testing, visual regression tests, benchmarking, and if those are right we don’t care so much about the code anymore, and we can even easily rewrite it or optimize performance. We can actually make the code higher quality than ever, it’s not that the code looks right, it’s that the output a user sees looks right, and we can validate that. This could be done before but it was prohibitively expensive, automating things, dealing with flakes, writing new cases for new code, it was hard, and now it’s easy, and we can create better results than ever before.


“there aren’t any cryptography results” “there aren’t any cryptography results”

Drop the ego and then curiosity, exploration, and discovery emerge. The best part of technology. "What if?" It's better now than ever before.

LaBinius: Lattice-based Polynomial Commitment Scheme over a Binary Field. Excited to present LaBinius, my recent work with @gregor_seiler. We show how to instantiate an Ajtai commitment over extremely efficient cyclotomic power-of-three rings and then connect it with a polynomial opening mod 2, so that R/2R is a binary field. Such a PCS is bridged with SOTA frontends, including Flock and Binius, to obtain end-to-end proofs of standard hashes with throughput of tens of thousands per second. osdnk.me/blog/labinius

🗣️ Can we evaluate the process of intellectual production rather than just the final text? Can we creatively design approaches to filter unaccountable research and nudge researchers and students toward genuine ownership of their work? Yes. 1/11 🧵