Meta goes truly open + OpenAI ships a hacking model

Meta published Muse Glimmer, a 30 billion parameter agentic model, under Apache 2.0 — a standard permissive license with no revenue threshold, no geographic carve-out, and no attribution clause, which is a real break from the custom community licenses Meta attached to the Llama family. The model targets always-on local agents rather than frontier benchmarks, ships a 131,072-token context window, and is small enough to run on a single consumer graphics card. Meta reports 76 percent on SWE-bench Verified and positions it against Google's 31 billion parameter Gemma 4 and Alibaba's 27 billion parameter Qwen 3.6.

The Muse Glimmer release landed alongside The Future is for Everyone, a 6,500-word essay in which Mark Zuckerberg argues that superintelligence held by "a small number of individuals, businesses, government, or AI itself" produces worse outcomes than widely distributed capability — an explicit inversion of the safety case made by OpenAI and Anthropic. Meta committed a $1 billion fund for communities near its data centers, said its board will review whether releases meet approved safety guidelines, and opposed a proposed 30-day federal review window. The Register noted the essay pairs that openness pitch with $600 billion of planned spending through 2028.

OpenAI released GPT-5.6-Cyber to vetted defenders through a new Daybreak Red tier, built on GPT-5.6 Sol and tuned to comply with zero-day discovery, exploit-chain development, and authentication-bypass requests. It completed 95 percent of those requests against 1.5 percent for the guardrailed Sol model — a refusal-rate figure, not an accuracy one. The release came days after OpenAI paused its Astra model because early evaluations could not rule out a Critical cyber rating; GPT-5.6-Cyber was rated High. Access requires identity verification and legal attestations, with hardware security keys mandatory from September 1.

Anthropic asked a research build of Claude to attempt the Riemann hypothesis. It did not solve it, but it raised the known lower bound on the fraction of the zeta function's zeros satisfying the hypothesis from 41.6 percent to 67.2 percent, using about 31 million output tokens across two sessions and 54 arXiv papers. Two Anthropic mathematicians checked it, Claude produced a formal proof in Lean, a proof-checking language, and Brian Conrey and Dan Goldston reviewed the manuscript on short notice — Conrey set the 40 percent record in 1989. There is no published referee report yet, so treat it as a strong preprint claim rather than a settled result.

OpenAI bought back roughly $7 billion of employee shares at an $852 billion valuation, Bloomberg reported — the same price as its March 2026 round, so the sale creates liquidity without marking the company up. That matters mostly for what it says about timing: OpenAI filed confidentially with the US Securities and Exchange Commission in June, and a tender offer this size relieves the pressure that usually forces a company public.

Amazon launched the Trainium Frontier competition, which asks teams to train language models from scratch on Trainium2 hardware and report what changes when the accelerator is not a graphics processing unit. Entry is capped at the first 100 teams, prizes run to $25,000, and finalists present at the NeurIPS conference in Sydney in December. The framing is the interesting part: Amazon is betting that architectures co-designed with custom silicon beat architectures ported onto it, which is the argument every non-Nvidia chip program eventually has to win.

Discovered Materials, a Y Combinator company founded by Advaith Sridhar and Akash Ramdas, raised a $9 million seed round, led by Lightspeed India Partners with Peak XV Partners and angels including Paul Graham. It runs swarms of Anthropic-powered agents that propose candidate semiconductor materials, then screens them through physics-based simulation — thousands of guesses a day against a manual process measured in months. The target is heat: thermal limits, not transistor counts, increasingly cap what an AI chip can deliver. No customers are named yet.