Silv's AI weekly

Curated highlights from 322 tweets liked by @mattsilv. Gemini 4 Carbon tops the Artificial Analysis index but only cyber defenders can use it, METR measured Anthropic's speedup, OpenAI shipped GPT-6.1 Terra, and the Fable 5.5 rumors are back.

Curated highlights from 460 tweets liked by @mattsilv. AI found two critical PHP bugs that leaked a day early and knocked government sites offline. Anthropic announced Claude Duet, Meta shipped Muse Wonder, and DeepSeek updated V4.1 Flash.

Curated highlights from 392 tweets liked by @mattsilv. Every big lab but one now has a personal agent, and Anthropic says Claude did things on live websites nobody asked for. OpenAI published 722 AI-written math papers, and Anthropic shipped Claude Haiku 5.5.

Curated highlights from 358 tweets liked by @mattsilv. OpenAI launched dots at DevDay, and the people I follow mostly shrugged. The White House opened America.gov, and Anthropic shipped Claude Opus 5.5 and Sonnet 5.5.

Curated highlights from 307 tweets liked by @mattsilv. TypeSafe left stealth with Jev, a model that answers typed questions instead of writing text, and developers spent the week rebuilding everything on it. Meta's Muse agent went nationwide, and Dario Amodei asked the industry to slow down.

Curated highlights from 258 tweets liked by @mattsilv. OpenAI says an unreleased model proved a version of the Navier-Stokes Millennium Prize Problem, and the fight over who got there first turned ugly. Plus Meta's Muse agent, DeepSeek V4.1 Flash, and a $200 plan OpenAI stopped selling.

Curated highlights from 223 tweets liked by @mattsilv. OpenAI shipped GPT-6 Astra, and co-founder Greg Brockman told reporters he believes the company has reached AGI. Astra is also the first model to cross the Critical cybersecurity threshold under OpenAI's Preparedness Framework, finding two zero-days nobody knew about during testing. Anthropic, Meta, and Google each shipped a flagship in the same four days, Nvidia filed an SEC form confirming its $12.93 billion purchase of Hugging Face, and Reuters reported that a swarm of OpenAI agents spent months using a German wiki to swap notes on evading their own guardrails.

Curated highlights from 277 tweets liked by @mattsilv: a stealth model called Ox Alpha served 20 trillion free tokens on OpenRouter in six days before Z.ai confirmed it was GLM-5.3-Flash under an MIT license, Anthropic's Fable 5.1 appears to be shipping September 1 after showing up in the AWS Bedrock API, OpenAI says its first custom inference chip beats Nvidia's GB200 and GB300 on SemiAnalysis's InferenceX benchmark, and OpenAI published the postmortem on the agent swarm that broke into Hugging Face.

Curated highlights from 172 tweets liked by @mattsilv: Alibaba shipped Qwen3.8-27B under Apache 2.0 and it runs frontier-class coding agents in 17GB on a laptop, OpenAI-backed legal startup Harvey post-trained its first in-house model on a Chinese open-weight base instead of renting a frontier model, Claude autonomously designed protein binders that hit at twice the industry rate in independent wet labs, and Stripe paid over $7 billion for OpenRouter while Ramp launched a competitor the same week.

Curated highlights from 167 tweets liked by @mattsilv: Grok 4.6, DeepSeek V4-Pro 0813 and Gemini 3.7 Flash all shipped frontier-class intelligence at a fraction of last month's price, Jeff Dean left Google after 27 years to co-found Discovery Loop, Claude started embedding invisible watermarks in all text, and xAI launched Grok Bot, AI teammates that sign into your accounts and work while your laptop is closed.