Anthropic's 3-SAT claim comes with a Lean proof

Anthropic says Claude found a subexponential algorithm for 3-SAT, contradicting the Exponential Time Hypothesis. Its repository includes an AI-written Lean proof, but no machine check confirms that the formal statement matches the paper. The result says nothing about practical SAT-solving performance.

Axiom releases its formalization of finite simple groups

Axiom Math

Axiom Math says it has completed a computer-checked Lean proof of the Classification of Finite Simple Groups, translating a mathematical result spread across roughly 15,000 pages.

Gemini 4 Carbon leads the Artificial Analysis index

Google

Google announces gains in coding, agentic tasks, and cybersecurity. Artificial Analysis's current table confirms Carbon's top ranking on its Intelligence Index; that aggregate ranking doesn't establish performance on a particular workload.

Anthropic and Sonar on verifying generated code

AI Engineer · Addy Osmani and Manish Kapoor

Addy Osmani reports 30,000 daily code merges at Anthropic and describes first-pass reviews by Claude, end-to-end tests, and browser evidence for UI changes.

Temporal puts state management ahead of model performance

AI Engineer · Temporal

In Temporal's survey, 35.7% of respondents cite state management as a productivity blocker. A payout-workflow demo shows durable execution preserving an agent's completed work when the worker process is killed.

Codo gives independent reviewers repository context

AI Engineer · Codo

Codo maps repository relationships and links ticket acceptance criteria to independent review agents. The talk also describes review checkpoints during code generation, before the pull request is ready.

Sonar demonstrates an agent for logic-level security flaws

AI Engineer · Taylor Luttrell

Sonar's Hunter Agent targets access-control and business-logic vulnerabilities that pattern-based static checks can miss. Its sample Python application shows 46 open issues, with findings presented in the usual SonarQube issue list.

Revolut reports a coding AI bill one-tenth as large

AI Engineer · Revolut

Following this week's open-weight cost discussion, Revolut describes a production deployment of a self-hosted GLM model. The bank reports cutting its provider bill to one-tenth of previous spending, with sandboxes and human approval restricting autonomous actions; the figures are self-reported.

A teardown examines Claude Code's background transcripts

flamerix-ai · GitHub Gist

The teardown claims Claude Code repeatedly sends background agents' transcripts back to the model, affecting token use and the data transmitted during a session.

Supabase discusses its database stack for agents

Latent Space

The interview covers Supabase Compute, OrioleDB, MultiGres, and the Turso acquisition, with a focus on databases used by agents.

Epoch maps AI chip arrivals by country

Epoch AI

The AI Chip Imports explorer estimates compute arrivals by country since 2022, complementing Epoch's chip-sales and data-center datasets.

WSJ reports Nvidia will invest $3 billion in Meta's offering

The Wall Street Journal · via TLDR AI

The reported investment accompanies an expanded AI partnership and Meta's planned development of Nvidia infrastructure. The article is paywalled; TLDR's summary provides limited detail.

xAI home-buyout offers reportedly require a lawsuit waiver

The Verge

The Verge, citing The Guardian, reports that an xAI subsidiary is offering to buy Mississippi homes on condition that residents don't join a lawsuit over a nearby data center and power plant.