Petar Maymounkov

Here's an excellent safety measure for your AI adventure: A flexible file-system sandbox, which can be engaged impromptu in the middle of your workflow unlike a virtual machine.

Today, at the @DARPA expMath kickoff, we launched 𝗢𝗽𝗲𝗻𝗚𝗮𝘂𝘀𝘀, an open source and state of the art autoformalization agent harness for developers and practitioners to accelerate progress at the frontier. It is stronger, faster, and more cost-efficient than off-the-shelf alternatives. On FormalQualBench, running with a 4-hour timeout, it beats @HarmonicMath's Aristotle agent with no time limit. Users of OpenGauss can interact with it as much or as little as they want, can easily manage many subagents working in parallel, and can extend / modify / introspect OpenGauss because it is permissively open-source. OpenGauss was developed in close collaboration with maintainers of leading open-source AI tooling for Lean. Read the report and try it out:

AXLE - our core infrastructure - is available to the community for free today. axle.axiommath.ai Check it out!

@cursor_ai when the cursor is at an empty line, the IDE shows an inline hint which obstructs the line below. This seems to happen with various fonts. The sample below uses Jetbrains Mono. It’s really annoying and I can’t find a way to fix it. Plz help.

This story is actually insane: • dude drops $2000 on a DJI robot vacuum like a lunatic • refuses to use the normal app like a peasant • Sammy Azdoufal fires up Claude to crack the API so he can drive it with an xbox controller • Claude delivers the goods • pulls an auth token from their servers, connects successfully • except the system thinks he controls 7000 vacuums • checks again • yep, seven thousand • DJI built authentication with zero device ownership verification • any valid token works for any unit on the planet • Sammy now has eyes inside homes across 24 countries • live vacuum camera feeds everywhere • full floor plans from the mapping data • some guy in germany eating cereal at 3am, unaware his roomba is snitching • one API call away from being the most informed burglar in history • all he wanted was to steer his vacuum with a joystick • does the right thing and reports it • DJI fixes it in two days • back to normal life with his stupidly expensive floor cleaner • IoT companies stay undefeated at shipping garbage security

Nice set of books, but not sure if many AI experts know **all** these topics.

Lambda @class_lambda has been hiring mathematicians the last years. We're quickly becoming a formal verification powerhouse. Check the AMO-Lean bogpost. AMO-Lean: what if every compiler optimization came with a machine-checked proof? Compiler optimization bugs are not hypothetical. They're well documented. For most software this is fine. For crypto, aerospace, or finance, a silent miscompilation can be catastrophic. AMO-Lean is an optimization pass where every transformation is a theorem formally verified in Lean 4. You write a spec, the system: 1. It explores all equivalent programs simultaneously via equality saturation 2. Picks the fastest variant 3. Emits optimized C or Rust 4. Proves the output is semantically equivalent to the spec Each rewrite rule is a proven theorem, automatically converted into optimizer rules via metaprogramming. There are no manual translation step where correctness silently breaks. The only unverified component is the cost model. If it picks wrong, you get slower but correct code. Normal compilers give you silently wrong code. Way better failure mode. This is particularly relevant for crypto, where optimization bugs are a known class of real vulnerabilities. Provable equivalence between spec and optimized output kills this attack surface. We will be writing a paper about it and we will continue developing it. We already have a proof of concept on matrix computations today, but the architecture generalizes.

Our new paper "Deriving neural scaling laws from the statistics of natural language" arxiv.org/abs/2602.07488 lead by @Fraccagnetta & @AllanRaventos w/ Matthieu Wyart makes a breakthrough! We can predict data-limited neural scaling law exponents from first principles using the structure of natural language itself for the very first time! If you give us two properties of your natural language dataset: 1) How conditional entropy of the next token decays with conditioning length. 2) How pairwise token correlations decay with time separation. Then we can give you the exponent of the neural scaling law (loss versus data amount) through a simple formula! The key idea is that as you increase the amount of training data, models can look further back in the past to predict, and as long as they do this well, the conditional entropy of the next token, conditioned on all tokens up to this data-dependent prediction time horizon, completely governs the loss! This gets us our simple formula for the neural scaling law!

CSLib: The Lean computer science library. ~ Clark Barrett, Swarat Chaudhuri, Fabrizio Montesi, Jim Grundy, Pushmeet Kohli, Leonardo de Moura, Alexandre Rademaker, Sorrachai Yingchareonthawornchai. arxiv.org/abs/2602.04846… #ITP #LeanProver #CompSci

New post on "Proofs and Intuitions": Verifying Distributed Protocols in Veil. We take a tour of Veil, a Lean-based verification framework that combines TLA+-style model checking with formal proofs and enables AI-powered invariant inference. proofsandintuitions.net/2026/02/09/dis…

The partial synchrony BFT cheat sheet by @kartik1507 If you are just beginning or are already an expert - this has you covered! decentralizedthoughts.github.io/2026-02-02-ps-…

A preview of how maths will be done in the future

@logic_int just saturated PutnamBench! The ramifications vis a vis mathematics are profound but I am much more excited about what an agent like Aleph can unlock in code-gen more generally given its unique capabilities. More to come on that point soon...

From the years I’ve spent studying the Ancient Maya & analyzing the LiDAR data: I guarantee that there are around 10,000,000 un-excavated/uncharted ancient structures throughout the Maya jungles alone. That’s not including ruins belonging to the Olmecs, Zapotecs, or Aztecs, any of the expansive Central American cultures in the jungles of Honduras, El Salvador, Nicaragua, or Panama… and not to mention that the size of the Maya jungles are only about 1/20th the size of the Amazon jungle. The Americas are a lost world.

This paper looks interesting too. Maybe 2026 will be the Year of the Improved Residual Connection.

Great to see an AI lab doing and publishing science (as well as discussing engineering efficiencies)! Some of the other “frontier” labs should try it! Thx, @deepseek_ai!

Nevertheless, it would have been great if the paper had included a cite of @HochreiterSepp’s original introduction of the residual connection idea: bioinf.jku.at/publications/o…

Take two large random matrices and linearly interpolate between them at several hundred steps. Compute the eigenvalues for each interpolated matrix, then plot them in the complex plane. The result is shown here. Made with #python #numpy #matplotlib

Current works on automatic theorem proving are awesome, but the end game with real economic impact that these works are leading to is “programming by specification”.

In other words, the Buddhists were right. The path matters more than the destination.