AI News 2026-08-02

$ cat top-story.md

OpenAI's Next Model "Astra" Solves 10 Decade-Old Math and CS Problems for ~$2,000 Each

An internal, unreleased version of OpenAI's next model family, code-named Astra, produced ten new results across group theory, cryptography, coding theory and complexity theory — including the first explicit construction of a non-sofic group, a question open since 1999. Each result ships with a machine-checked Lean 4 proof, and OpenAI pegs the compute cost at roughly $2,000 per solved problem.

Why it matters: It's the first real signal of what OpenAI's next-gen model class can do before release — though Simon Willison points out OpenAI hasn't disclosed how many problems failed at that same $2,000, so the "cheap genius" framing deserves a skeptical read.

// source: OpenAI · Hacker News · Simon Willison

## 🛠️ Tools & Repos

▸ ego-lite Is an Actual Browser Built for AI Agents to Drive Alongside You

An MIT-licensed macOS browser that gives Claude Code, Codex or Cursor an isolated "Space" inside your real, already-logged-in Chrome profile, so agents inherit your cookies and sessions without you handing over passwords — 7.4k stars, 4k+ added in the past week alone. Why it matters: Sidesteps the usual choice between screenshot-heavy browser automation and unsafe credential sharing for agent workflows. // GitHub

▸ Hugging Face's speech-to-speech Puts a Fully Local Voice Agent on Top of Your Own LLM

A modular VAD→STT→LLM→TTS pipeline exposed over an OpenAI-Realtime-compatible WebSocket API, where every stage — including the LLM — is swappable for a self-hosted model; it's the same backend running Hugging Face's Reachy Mini robots in production. Why it matters: Point the LLM slot at a local vLLM endpoint and you get a fully offline voice interface on top of an existing self-hosted stack, no cloud API required. // GitHub

## 🤖 Agents & MCP

▸ SkillGate Screens Agent Skill Files for Malware Before They Ever Run

A new paper proposes a lightweight MCP proxy that screens third-party SKILL.md files with a 530-rule regex prefilter plus a snippet-only LLM judge, hitting F1=0.817 with a 1.13% false-positive rate at ~818ms latency — 5-6x better than existing scanners on threshold-independent metrics. Why it matters: A practical, pluggable runtime defense for anyone installing third-party Skills, rather than another "please audit your skills" warning. // arXiv

▸ Datasette-Agent Lets Agent Tools Run Code Directly in Your Browser

Simon Willison's datasette-agent 0.4a0 ships context.browser_task(), letting an agent tool execute custom JavaScript in the user's actual browser session; he immediately used it to build an invisible-iframe smoke-test loop in the follow-up datasette-apps 0.2a0. Why it matters: A concrete, working pattern for closing the loop between an agent and a live app for self-verification, useful beyond the Datasette ecosystem. // simonwillison.net

▸ Anthropic Ships an Official Claude API Skill, Bundled Into Claude Code

The open-sourced claude-api skill gives Claude live, progressive-disclosure reference docs for the Messages API and Managed Agents across eight languages, auto-detects your project's language, and can even run full model migrations via /claude-api migrate. Why it matters: A template for how Anthropic wants first-party Skills built — scoped, auto-triggering, low-context — worth studying before writing your own. // Anthropic

▸ Every.to Open-Sources Its Cross-Tool "Compound Engineering" Workflow

A 32-skill plugin implementing a brainstorm → plan → work → review → compound loop that persists what an agent learns from each task into docs/solutions/ for reuse next time, installable across Claude Code, Cursor, Codex and more. Why it matters: A portable, vendor-agnostic pattern for making coding agents actually improve on a codebase over time instead of restarting from zero each session. // GitHub

## 🏠 Self-Hosting & RAG

▸ Notion Acquires ZeroEntropy — Its SOTA Rerankers and Embedding Model Go Fully Open-Source

Notion acquired reranking startup ZeroEntropy and, as part of the deal, released its entire lineup — rerankers zerank-1, zerank-1-small, zerank-2 and embedding model zembed-1 — under Apache 2.0 on Hugging Face; zerank-2 had been powering Notion AI's unified search in production with an 85% latency improvement. Why it matters: These were commercial API-only models until now — worth benchmarking as a drop-in reranking step ahead of Qdrant retrieval in a self-hosted RAG pipeline, no per-query API costs. // ZeroEntropy

## 🎬 Worth Watching

▸ Boris Cherny: Building Claude Code (YC Startup School 2026)

Claude Code's creator sits down with Y Combinator's Diana Hu to talk through how the tool came together, what Opus 5 changes about building on fast-moving model capability, and his advice to build for "the model six months from now." Why it matters: Rare direct insight from the person actually building Claude Code, straight from the source rather than a third-party review. // Y Combinator

▸ I Tested Abacus's New SUPERCOMPUTER... (INSANE)

Wes Roth spins up Abacus AI's always-on cloud Linux box, wired to persistent agents (Hermes, Claw), and live-builds several projects on it — including a self-hosted AI chat backed by Qwen. Why it matters: A concrete look at the "persistent agent environment" pattern, relevant even without adopting the product, for anyone thinking about giving their own agents a durable place to work. // Wes Roth

## 📰 Worth Reading

▸ 1,324 AI Lab Employees Sign a Letter Asking Governments to Build a "Pace the Frontier" Mechanism

Simon Willison maps three dueling open letters from the last week: Microsoft's 235-signatory letter defending open-weight models, Anthropic's Dario Amodei countering with concerns about authoritarian misuse, and "Pacing the Frontier," signed by 1,324 frontier-lab employees (including Amodei and OpenAI's Pachocki) asking for mechanisms to slow AI development if it starts self-accelerating. Why it matters: A rare condensed map of where the actual policy fault lines sit between labs right now, not just vibes. // Simon Willison

▸ Software for One: What Six Months of Building With Claude Code Actually Looks Like

Adam Waxman describes building five bespoke personal apps — a sleep tracker, a marathon planner, a medical-records manager — almost entirely via Claude Code over six months, and argues that "9 times out of 10 a feedback screenshot shared with Claude gets the job done." Why it matters: A concrete, lived account of what a Claude Code-heavy personal workflow looks like day to day, rather than another benchmark post. // ajwaxman.com

▸ The Development Pipeline Is a Production System

Jerry Orr argues CI/CD and dev tooling deserve the same incident-response urgency as customer-facing outages, since a silently broken pipeline quietly stalls everything downstream of it. Why it matters: Especially relevant once agents are driving your pipeline — a stalled CI run doesn't just annoy you, it silently blocks every automated coding agent behind it. // jerryorr.com

AI News — August 2, 2026