First screen, captured 3 days ago. Open the video.

AlphaProof Code

What HN thinks

Six million lines of Lean later, everyone agrees the proof is fine and argues about whether the theorem was.

Proofs only as good as specs
Many argue formal verification just moves bugs into the specification, and this work proved narrow security properties of one binary, not bug-free software.
Impressive, useful, refreshingly unhyped
Others praise Google for a concrete, expert-in-the-loop use of AI, including reintroducing known bugs to check the process catches them.
Verified-everything future debated
Some predict unverified software will soon seem reckless, while skeptics note the tiny scope of what can be specified, hardware flaws, and the cost of re-proving changing code.

Consensus: Most agree the result is valuable but limited: it verifies a specific binary against formalized security goals under assumptions, and cannot guarantee the specs themselves are right or sufficient.

Recommended in the comments

Verus (tool): Maps Rust code to proofs directly, keeping proofs closer to real code.
Lean (tool): Commenters use AI-generated Lean as a logic gatekeeper and even to build whole verified languages.
lean-rs ↗ (library): Suggested as a way to bridge Lean and Rust code.
seL4 (product): Cited as precedent for verifying real systems via an intermediate model.

Summarised by Claude from the comment thread 3 days ago.

Points: 658
Comments: 187
Best rank: #1
Posted: 2026-11-10 by ilreb
Category: AI & ML

formal-verification alphaproof lean signal-protocol deepmind ai-agents cryptography

All captures:

Full page

Captured 3 days ago. Click for the whole screenshot.

Looks similar

Screenshot of deepmind.google: How AlphaChip transformed computer chip design

How AlphaChip transformed computer chip design (deepmind.google)

281 points by isof4ult 2 years ago | 162 comments

GPT-5-Codex (openai.com)

367 points by meetpateltech 1 year ago | 115 comments

OpenAI announcement of upgrades to Codex, including GPT-5-Codex, making it faster and more reliable across terminal, IDE, web, and mobile.

openai codex gpt-5-codex coding-agents

Screenshot of cloud.google.com: 'Reptar', a new CPU vulnerability

'Reptar', a new CPU vulnerability (cloud.google.com)

114 points by el_duderino 3 years ago | 62 comments

Screenshot of deepmind.google: CodeMender: an AI agent for code security

CodeMender: an AI agent for code security (deepmind.google)

180 points by ravenical 1 year ago | 26 comments

Screenshot of cerebras.ai: Cerebras Code

Cerebras Code (cerebras.ai)

423 points by d3vr 1 year ago | 165 comments

Cerebras blog post announcing Cerebras Code Pro ($50/month) and Code Max ($200/month) plans offering fast Qwen3-Coder inference at up to 2,000 tokens per second.

cerebras qwen3-coder ai-coding llm-inference