Tech News
← Home  ·  All topics

Formal Verification

24 GoKawiil briefs on this topic · RSS for formal verification

Lean creator Leo de Moura discusses AI-driven formal verification on Gradient Flow podcast

Ben Lorica interviewed Leo de Moura, creator of the Lean theorem prover and VP/Distinguished Scientist at AWS, about how AI is making formal verification more accessible. The conversation covered machine-checked proofs, trust in AI-generated code and autonomous agents, and the concept of a fully verified software stack.

AI startup &Truth launches formal-verification tool, cuts engineering time from months to hours

Stanford researcher Chuyue Sun's AI system completed an engineering verification task in under half a day for less than $500, a job that previously took three experts five months. The work, used to verify part of Signal's encryption code, underpins the public launch of startup &Truth, which Sun joined in August and which has raised $70 million. Rival Harmonic, backed by $295 million and cofounded by Robinhood CEO Vlad Tenev, this week said its AI agent Aristotle can now verify software in C and hardware designs in Verilog, and is testing with firms including Tensordyne and Nvidia's venture arm.

Lean Formalization Completes Classification of Finite Simple Groups

A research team has produced a complete machine-checked formalization in the Lean proof assistant of the classification of finite simple groups, following the original first-generation proof. The formalization comprises over 43 million lines of Lean code and 2.05 million declarations, drawing on 715 papers and books. The team also reports finding an issue in a previously published argument for the Unbalanced Group Theorem and describes the alternative proof route they used instead.

Anthropic and Theorem use Claude to formally verify over 100 open-source libraries

Anthropic, working with Theorem, reports using its Claude models to generate formal mathematical proofs of security properties for more than 100 real-world open-source software libraries at the RISC-V assembly level. The company says this is the first large-scale attempt to apply AI-driven formal verification to production software at that level, and that Claude uncovered previously undetected bugs in the process.

Nvidia releases OpenShell 0.1, a sandboxed runtime for AI agents

Nvidia has launched OpenShell 0.1.x, a runtime designed to let autonomous AI agents access files, APIs, and credentials under enforced policy limits. The release adds a stable version cadence, new isolation primitives, an expanded extension surface, and new APIs, with kernel-level sandboxing and formal verification used to check policy changes before they take effect.

OpenAI releases 722 math manuscripts, mathematicians question validity and value

OpenAI published 722 manuscripts and proofs addressing longstanding mathematical problems, generated by an unnamed internal frontier AI model, with only 300 results formalized for computer verification. Researchers who reviewed the work told The Verge they found nonsensical explanations, malformed formulas, and results that duplicated existing research, while OpenAI withheld the model's identity and the prompts used to generate the findings.

Eurydice project compiles Rust code into readable C output

Eurydice is a new tool that converts Rust programs into C code while preserving the original code's structure, rather than optimizing or obscuring it. Started in 2023, it is part of the Aeneas formal-verification project maintained by researchers at Inria and Microsoft, and has already been used to translate post-quantum-cryptography routines from Rust into C.

OpenAI updates GitHub math repository, withdraws three Lean formalizations

OpenAI revised its public GitHub repository of formalized mathematical results, adding six new Lean formalizations and modifying 19 existing ones. It also withdrew three results previously listed as formalized. The company said roughly 42% of its top-line results are now formalized and that it will keep updating the repo as it finds errors.

Developer formalizes DFA textbook proof using Lean and Mathlib

A software engineer walked through a theory-of-computation textbook exercise asking for a proof, via finite automata, of a language property, then reformalized that informal constructive proof in the Lean proof assistant. The write-up builds a deterministic finite automaton (DFA) and leans on existing theorems in Lean's Mathlib library to formally verify it recognizes the target language.

TLA+ advocate pushes back on claims formal methods will fix AI coding

Following Claude Code creator Boris Cherny's comment that the Opus model used TLA+ to detect race conditions, online discussion has surged around formal verification as a fix for agentic software bugs. A longtime TLA+ educator and advocate argues this enthusiasm is overblown, noting that formal verification only checks properties that engineers explicitly define, and many important properties can't even be expressed in TLA+.

Explainer breaks down TLA+ formal specification language for distributed systems

An interactive tutorial walks through TLA+, a formal specification language used to describe transition systems and verify their properties. It illustrates core concepts like states, actions, safety and liveness properties using a leader-election example among three computers, showing how a model checker can explore all possible states and surface counterexamples when rules are violated.

Critique: Bend 2's AI-verification language demands hundreds of lines for simple proofs

A blog post examines Bend 2, a programming language designed for an AI-assisted workflow where humans write 'laws' and AI writes implementations plus formal proofs that a compiler checks. Using Bend's own homepage demo as an example, the author notes that stating a simple game rule takes 58 lines of code, while the AI-generated proof of that rule balloons to 442 lines. The piece argues this ratio illustrates a broader vibe-coding pitfall rather than being unique to Bend.