Sam Altman drops four words (32803 likes, top of the week): 'we made a chip and it is fast.' Claude unifies memory across chat and Cowork (6846 likes): 'Hand Cowork a task and it starts from what Claude already knows from your chats: the project you talked through, your manager's preferences, or the client from last quarter.' Everything remembered is a list of topics in Settings — read, edit, or delete each one (549 likes); sensitive topics like health or religious beliefs stay out unless enabled (394 likes). Cat Wu confirms (588 likes): 'Thanks to your feedback, we've unified memory across Chat and Cowork.' Boris Cherny ships (1429 likes): 'A small improvement: memory is now simpler and more powerful.'

Sottiaux's Team plan (2926 likes): like the Pro $100 plan but for teams and small companies — all ChatGPT/Work/Codex features, Google Workspace/Slack/GitHub/M365 connectors, SAML/SSO/MFA secure workspace, centralized billing, usage analytics and spend controls, and NO 5h limits. Plus 'Good products take time. At least 34 days.' (2034 likes) and 'I want to see the polymarket' (367 likes). Swyx's PSA (5 likes): don't use Codex 'locked use' right now — it relies on unstable macOS features and locked him out of his Keychain twice this week.

Thariq teases (1118 likes): 'excited to share more on how we're making Claude Code more hackable soon.' Google Labs ships Putty (1234 likes): 'a collaborative vibe coding tool that lets you build tools and websites together in real time. Ready for your imagination to go multiplayer? Join the waitlist.' Rauchg's double: Run SDK (419 likes) — 'secure eval for dynamic Code Mode execution. When agents write code, you don't always need a full sandbox. You can run their code in a lightweight QuickJS secure context. Faster and more cost-efficient. npm i run'; and Vercel Connect GA (214 likes) — 'The hardest problem in building agents is secure connectivity to services and data. Run vercel connect create notion, then get an MCP client you can query on behalf of the authenticated user.'

Levie on the applied AI gap (122 likes): 'There's a wide gap between the AI models and the underlying workflows of an enterprise, which leaves a ton of opportunity for applied AI companies. The world doesn't just want raw models and agents; it wants problems resolved and outcomes achieved. The premium will sit with the companies that can diffuse this intelligently.'

Peter Yang open-sources /fuck-cancer (1389 likes): 'an AI skill that helps patients and caregivers navigate cancer diagnosis and treatment and advocate for themselves and their loved ones' — built from what patients told him after sharing his mom's story: 'You have to be a huge patient advocate. Ask questions and push for answers, biopsies, and proper testing.' The skill creates and updates a single-source-of-truth brief: patient info, next steps, what we know, medical term definitions, and an update log (101 likes). Plus his DevDay excitement (117 likes): 'ChatGPT / Codex is basically the OS I use for everything I do on my computer.'

Garry Tan's trio: the enigmatic 'lmao' (4686 likes); 'You need to be cross-harness memory maxxing' (510 likes); and the clout-chaser clapback (60 likes). Madhu Guru's evals Part 9 (83 likes): the Eval Roadmap Problem — evals fail because teams treat them as static artifacts while user expectations evolve; early user wants a 5-page earnings summary, later users want competitive positioning across 50 filings. Nikunj's El Niño monitor (8/3 likes): real-time updates from government sources, impact per region, historical records, glossary — built with Codex and Railway. Plus the VC dinner joke (14 likes). Steipete's week in review (280 likes). Dan Shipper's trio (37/51/71): GET DOT, a device to test, and an upcoming Every vibe check. Aditya on datacenter resentment (18 likes): 'AI today primarily helps knowledge workers and the highest paid segments. I suspect the big change will happen when AI is able to find cures for diseases that ail everyone.'

Training Data interviews Parag Agrawal — Twitter's former CEO, now founder of Parallel Web Systems — on scaling agentic search for the agentic web. Core thesis: 'human click data is a bug; agents doing work with search should rely on agent feedback, not human feedback.' Parallel started with the bet that agents would search 1000x more than humans ever have, requiring reinvented technology AND business models. Key threads: unlearning post-PMF Twitter lessons for a pre-PMF company building for a customer that doesn't exist yet; search as a billion-to-billion matching problem (hundreds of billions of pages × queries → the right thousand tokens); launching a search AGENT first (not a search engine) — trading the index for inference-time compute, competing with outsourced human curation; early customers in insurance underwriting, sales enrichment, finance data pipelines; quality-cost-latency with latency ignored for two years, then TurboNow — 'the fastest, highest quality agentic web search on the market by a lot'; agents as a different customer class — fewer typos, better-specified, longer queries; the pre-AI human slop (SEO affiliate pages) as catering to lazy humans vs. agents getting authoritative excerpts directly; the announced Google Cloud partnership (search and grounding provider for enterprise agent APIs alongside Google Search); why model companies' pretraining crawls don't build fresh indexes (they're not completionists); search multipliers: a chat prompt = 5-10 searches, background portfolio agents = millions; the broken economics of the ad-funded web when humans don't show up — differential pricing must be rebuilt for agents; and the Shapley value framework for publisher compensation — remove a source, measure quality drop, price it against the compute needed to claw it back; the macro math (2-10% of LLM inference spend allocated to web data beats all web data business models outside walled gardens); 12-24 months to meaningful dollars for a wide range of content owners; and the company's original name: Shapley Inc. (shapley.ai was available; shapley.com is a parked porn domain).