Sottiaux's triple: the quality postmortem (15758 likes, top of the day): 'Working with some of you, we have found and fixed the following issues: Some skills written for previous models were triggering too often or preventing the model from checking its work. An opt-in context management experiment that could cause early stops or replies to older messages. We've disabled it. Our rough estimate is that 4-5k users were affected'; the ship list (4233): 'Astra powered ships this week: Images 2.5, GPT-Live-1, Agents API, Data Agent, ChatGPT for Financial Services — and yet it's not yet DevDay. Busy plans for next week too'; and the hire (3042): welcoming Aidan & Sasha from the Git AI team — 'We'll also keep Git AI open source and continue investing in it.' Steipete's Doom run (270): 'Astra on OC in a cloud session playing Doom with CUA. Not quite AGI yet, but probably beats fly brain.' Peter Yang's pair: software-factory skepticism (179): 'Other than for verification & testing, I don't think AI is at a point where it can self-improve a product or build a new feature end-to-end without human in the loop… if it makes one wrong assumption the whole thing becomes a waste of tokens'; and the choreography (91): 'All local scheduled tasks live in Codex. All cloud tasks I'm porting over to Grok Bot. Makes it a clean split.'

Thariq's double: the launch (1250 likes): 'we heard feedback that it's hard to know if your skills are still working with new model releases. plugin evals are here to help — run `claude plugin eval init` in your plugin folder'; and the interpretation thesis (1504): 'it's basically impossible to interpret evals by looking at just at the pass/fail scores these days. Many of the failures I see in benchmarks are due to overly strict hidden tests, in some cases the model's answer makes more sense than the expected eval result.' Shipper's vibe-check upgrade (42): 'better benchmark scores don't tell you much about how well a model does on your real work — that's why for the last 3 years @every we've done vibe checks… now we're doubling down and getting more quantitative: an internal platform for personal benchmarks based on our real day-to-day work.' Plus his hiring call (47) and the ominous 'okay this is not great' (73). Madhu Guru's failure taxonomy (96): 'Why most enterprise AI efforts fail. 1/ Using old-school playbooks for AI — the CEO picks a trusted lieutenant to lead a central AI team… the skill gap: the team structure, product patterns, how you launch and learn that got you here do not work with AI.'

Rauchg's thesis (290 likes): 'Tailscale's model router is powered by Vercel AI Gateway as its underlying infrastructure. AI Gateways are the new CDNs. You could go "direct to origin", but it's brittle. You could DIY, but it's painful and costly.' Amjad's trio: 'Replit acquires a business entirely built on Replit. I expect first of many to come' (548); 'Love seeing all the Astra 3D experiments on Replit!' (193); and 'Routines with budgets!' (66). Ryo Lu's double: the move (1029): 'i emptied my house, said goodbyes, and moved to Taipei. So far so great! 💛'; and the launch (599): 'long-lived agents for your big ideas — now in Cursor.' Zara's essay-in-three (704): 'I think the idea of a "one-person company" is overrated… building something new can be a profoundly lonely experience. It is extremely easy to lose motivation when you don't tie yourself to the mast together with somebody else.'

The rest: Garry Tan's SAT take (732): 'once you score 1600 you should unlock a second harder test… we should have more and higher measures of excellence. But instead they ban the SAT so that everything turns into random lottery'; Levie's Box-sandbox integration (81): 'you can mount Box to agent sandboxes… agents are going to need the same primitives that people have had'; Turck's 9/11 story (129) — the young co-founder of a NYC startup inside the day; Aditya's United 93 reflection (34) and SPC policy series (12); Nikunj's attribution wars (39): 'Success has many fathers, but failure is an orphan… in the next couple of years you'll see many names scrubbed.'

A brand-new No Priors (published 2026-09-10, new to this digest) with Coinbase cofounder/CEO Brian Armstrong — also cofounder of longevity company New Limit. Core threads: the Everything Exchange (every asset class coming on-chain — stocks, commodities, perps, prediction markets — '88% of our revenue is from non-Bitcoin trading'); stablecoin payments ('under one second, under 1¢, to move money anywhere in the world… money can now move at the speed of information, just like a WhatsApp message'); Agentic Finance / AIFi — 'we don't want the AIs to be unbanked. If we wanna bank the AIs, they deserve financial services as well… with a single prompt your agent can have its own financial account' (self-custodial wallets, no KYC — agents have no government ID); 76% of agent e-commerce transactions are under 30¢, where card rails (30¢ flat + %) make no sense; X402 (the agentic payment protocol Coinbase incubated and donated to the Linux Foundation — Google, Cloudflare, AWS working on it) and the agentic.market stats site; specialists beating generalists (open-weight models fine-tuned on 100k proprietary compliance cases outperforming frontier models); Elon's no-money-in-the-future take vs scarcity enduring ('Dyson spheres are going to be expensive pretty far into the future… crypto was really good for humans, and it's going to be essential for AI'); recursive self improvement inside Coinbase — the company 'brain' (incident history, financial controls, every AB test, full PR history per repo; every human correction flows back into the brain, so one-shot PR accept rates 'start to tick up'); the Toshi harness (10 parallel agents across cheaper models finishing phase one in minutes — 'instead of bugging the team, I just send them a PR for review, and it's already done… it's very addictive'); tasks are being automated, not people eliminated (company sizes won't shrink to 10, but 2-5 person companies will do far more); tokenized stocks (first true tokenized equity, 1:1 with custody; 4 billion people without brokerage access — 'Tether was the first version with USD'); prediction markets ($100M revenue run-rate within months, >100% QoQ; policy markets on unemployment rates, even 'opinion markets' on the big questions — 'Is there a God?'); New Limit (epigenetic reprogramming — the 'meta problem' of restoring youthful cell function; AI-recommended pooled screens; successful reprogramming of one human cell type in humanized mice; first phase-1 trial next year; liver→vasculature→T-cells; the $20B ALD drug; 'skin — that's probably a trillion dollar market right there'); starting a second company while CEO ('if you make money in software you have an obligation to put it into hard tech'); cognitive enhancement as the next meta problem (folic-acid IQ gains; 'AI is getting smarter — humans need to take control of our evolution'); embryo editing (Preventive investment; 80% of Americans support disease-prevention editing; 'at some point it'll be abnormal to have children without this screening'; gills and wings in the limit case — diversity increases, not decreases); and Freedom Cities / special economic zones (Prospera in Honduras — Coinbase invested; federal-land zones for data centers, nuclear plants, drone delivery without normal red tape — 'there is no sandbox to try innovation, and that is just incredibly stifling').