·
Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
PLUS: OpenAI starts watermarking text
Reflection is releasing Beam, its first open-weight model, aiming to be the DeepSeek of the West after raising $4.6 billion.
OpenAI will watermark eligible ChatGPT and Codex text in the EU to comply with the EU AI Act.
Reflection's Beam is a 501B-parameter open-weight model Alex Finn says rivals GLM 5.2, Qwen 3.8, and Opus 4.8.
Unlike Anthropic and Google, OpenAI is launching text watermarking only in the EU for now, not globally.
OpenAI made GPT-6 Astra and GPT-6.1 Sol about 50% faster by default across subscription products and partners.
a16z's seventh Consumer AI ranking adds a revenue leaderboard, with Claude climbing to #3 on web and ChatGPT topping 1B mobile users.
A16z's revenue data shows AI spend is concentrated, with the top 1% averaging $903 a month against a $25 median.
a16z's partners say consumer AI spending follows a power law dominated by developers and creators, and subscriptions may not reach everyone.
Hugging Face turned coding harnesses into RL environments, and the same model scored 62% in one harness and 33% in another.
Epoch AI's trend fit shows OpenAI researchers' coding-agent spending doubling about once a month from January to mid-August 2026.
Kevin Roose published a 2017 Dario Amodei document, the essay he says started the AI race, on scaling models.
Azeem Azhar says the leaked Anthropic S-1 is an older draft covering only 2025, with 2024 to 2025 revenue up 1,089%.
ContextQA launched Ship to reproduce and investigate bugs from Slack, Linear or Jira before an engineer hands the fix to an agent.
Pietro Schirano says his new ChatGPT plugin added $500K in ARR in a single week.
ChatGPT users no longer need developer mode to connect custom MCP servers, and can add them under plugins.
Claire Vo calls OpenAI's Sites updates an underhyped dev day launch, especially connected enterprise data and site-as-infra.
John Collison says personal agents are working because computer use fills out web forms without waiting for the DMV to add an MCP.
Hugging Face's Thom Wolf says an attack from a closed-source OpenAI model led the company to start its Open Alignment team.
A founder behind Underdog says local models on a MacBook now outperform frontier models OpenAI and Anthropic released six months ago.
n8n's CEO says the company grew more than 10x last year despite repeated claims that new AI tools would kill it.
Elvis built a personal agent with Pi Durable that keeps working when he's not watching, checkpointing every step.
Vercel's Guillermo Rauch released gdp-ts, a TypeScript library where the typechecker verifies authorization proofs for sensitive functions.
Replit's weekly changelog adds GPT-6.1 Sol and Claude Sonnet 5.5 as build models, plus enterprise workspace settings.
Grok Bot now has a public changelog with an RSS feed listing every release since launch.
Claire Vo says Opus 5.5 handles video with any caption style and clean overlays, without the usual slop.