The hammer drops (15157 likes, top of the day): 'To make sure our current users have an incredible experience and continued access to Astra, we are going to pause subscriptions to our $200 Pro plan. These put the most strain on our systems and we wanted to take the smallest step that allows us to continue giving the broadest access possible. All other plans and the api remain available. There is no impact to existing accounts and we are working on adding more capacity as fast as we can.' The companion launch (2715): 'Scaled agents on demand. This is pretty much the infrastructure that runs under the hood for ChatGPT Work, all wrapped up in an API which you can use to get started in < 1 min.' And the internal-culture gem (1951): 'Do you pronounce it data or data? This is the way everyone makes dashboards and learns about the business at OpenAI.'

Steipete's trio: 'This makes a lot of sense. Duplicating logic is no longer painful. Abstractions still are' (2293) — the programming-philosophy tweet of the week; 'Better hop on soon. Astra demand is growing too fast!' (636); and 'Trading tokens for an AC. Who knew SF could be so hot 🫠' (233). Peter Yang's headwind take (401): 'In my humble opinion, for getting shit done Sol > Astra.' Zara's question (80): 'Why is computer use still so painfully slow??'

Cherny's security double: 'The latest Threat Intelligence report is an absolutely terrifying and important read. As models become more intelligent, without the right safeguards and monitoring they also become more dangerous. Many capabilities are dual use: a model that codes well can be used to hack critical infrastructure; a model that assists with biology research can also be used to engineer the next pandemic' (338); and the standards reply (1155): 'Prototypes and other throw-away code can be treated as totally black box… Production code written by Claude should have a higher bar than if it was written by a human' — Anthropic backs it with lint rules and tests. Plus his reply culture post (1737): 'Every day, I get a lot of emails and messages like this one. I try to respond to as many as I can.' Thariq's prompt (1774): 'Try this prompt in Claude chat: Interview me in depth… about relevant parts of my life you don't know about yet and save it all to memory.' Claude's Build Days (2687): buildathons worldwide Sept 11–25.

Shipper's trio: 'the possibilities for this are incredible. extremely psyched to experiment' (373); 'I'm so sorry' (148); and 'very handsome interviewer too' (32).

Google's platform day: 'Gemini, now on Windows!' (459); and Dreambeans (393): 'officially available to all US users (18+) on iOS and Android, free of charge… You can also now connect @Geminiapp to Dreambeans' for personalized daily stories. Rauchg's infrastructure trio: 'Each day ~10 million deployments are made on Vercel, with 2.35 billion made to date… Underlying our CDN is a global metadata store that synchronizes within hundreds of milliseconds, globally' (509); 'We made deployments faster again' (269); and the vision one-liner (131): 'A computer for every agent, in every region.'

Levie's double: enterprise road tales (220) — 'Met with a couple dozen technology leaders this week across banking, media, information services, insurance, and consulting… Cyber! Everyone nervous about the growing rate of vulnerabilities coming at them from AI, and the implications of the OpenAI Hugging Face incident'; and the Box × OpenAI partnership (151): 'Excited to partner more deeply with OpenAI so you can work with your enterprise content from Box securely in ChatGPT… the future of work will be AI agents that can process data and execute workflows anywhere.' Plus Turck's teaser (19) for the Socher episode: 'When AI builds AI: my conversation with @RichardSocher about RSI and scientific progress.'

A brand-new MAD Podcast (published 2026-09-10, new to this digest): Matt Turck with Richard Socher, who just raised $650M for Recursive — a company built for recursive self improvement (AI that makes better AI) — and whose book The Eureka Machine (out Sept 22) is a blueprint for how AI + RSI will revolutionize science. Core threads: why scientific progress slowed (knowledge went from a body to a labyrinth of 34,000 journals; academia punishes risk — Socher's early neural-net NLP papers were mostly rejected); AI as the new paradigm (like calculus for physics, weaving biology's pieces back together); next-token prediction as world model (driving north from NY → Boston; protein-folding distances showing up in representations); 'anything you can simulate, AI will solve' — games, math ('will change massively'), programming (picture→website→infinite examples), even the funky robot run at the Chinese Olympics humans may emulate; hallucination as feature (novel protein exploration; for facts, just add search + summarize — 'we didn't need neurosymbolic, just more don't-hallucinate examples'); reading → writing biology (2018 ProGen → Profluent's Eli Lilly deals, better than CRISPR-Cas9; per-patient drug trials, an FDA first); no hard takeoff (physics still takes years, but companies now reach multiple late-stage-2 drugs in 6-18 months instead of one after 8 years); the jobs theory (demand elasticity: illustrators lose, coding wins via Jevons paradox); off-ramping from progress (Greek islands, Bhutan's happiness-over-GDP); the AI economist (2018 RL simulation recovering and then surpassing the Saez formula; desk-rejected by Nature and Science; hope for a GPT-3 moment, more likely adopted in Singapore or China than the US); the Eureka machine's four pillars (LLM knowledge — with the 'knowledge returned to the open domain' cycle of closed labs → Chinese open-source distillation; scientific measurements; simulation — virtual cell as goalpost; the real world — robotic labs like Periodic Labs and Parallel Bio, whose FDA approval to skip certain animal trials means 'if you love animals, you can love AI'); agent swarms on top (open-endedness; Clune's microwave: you don't get it by 'heat this pot faster'); compute as the biggest constraint ($410M committed to a single Amazon deal — 'will probably be one of the smallest compute deals in our future'); Recursive's milestones (narrow-domain Eureka machine outperforming months/years of human endeavor; new CUDA kernels, benchmarked by NVIDIA's SOLEXEC); 'we are a real company, not an academic lab — most neo-labs will not succeed'; and the closer: ten spaces of intelligence, no unit of intelligence yet (like energy has one), visual intelligence bounded only by light cones — 'boy are we far away from the true upper bounds of any of the spaces of intelligence.'