Levie celebrates the Cursor outcome (165 likes): 'Amazing outcome. Cursor executed the applied AI strategy flawlessly. Most people completely underestimated the market size in AI coding. The biggest developer tool exits in history were on the order of low billions. And it was already assumed to be a heavily competitive space when they really got going.' Madhu Guru on Cursor's underrated cultural impact (19 likes): 'AI products were stuck in the chatbot phase searching for a product meta. Then cursor for x came along and inspired a whole new product pattern.' Plus his differentiators framework (84 likes): 'When everyone can build with AI, your differentiators are product sense, domain knowledge, distribution, execution.' And the Jevons analogy (39 likes): cheaper tasks → more economically viable use cases.

Swyx on Databricks' Series M (91 likes): 'The M in their $188B series M stands for we are going to kill so many meetings.' Dan Shipper pushes back on Turck's rocketship-or-death framing (55 likes): 'You can be an AI-native rocketship without being in a permanent fundraising situation and sacrificing gross margin — but the rules for building a company like that are very different.' Nan Yu (59 likes): 'There's no force in tech as destructive as the PM promo packet.'

Josh Woodward ships 3.7 Flash into the Gemini App (735 likes, top of the day) and upgrades Pomelli (182 likes): 'turn those amazing photoshoots into short videos or gifs with that same ease' — popular with small businesses. Google Labs rounds up its experiment family (110 likes): Pomelli for marketing campaigns, Flow for visual storytelling. Sottiaux ships restaurant reservations in ChatGPT (583 likes): 'Looking for a quick restaurant reservation is now super easy.' And asks the community (666 likes): 'What's a hard problem codex solved for you this week?' Rauchg claims (305 likes): 'Vercel is the fastest [AI Gateway] infrastructure in the world.'

Garry Tan on GStack (103 likes): 'The most surprising thing about using GStack pre-Fable 5 and after is now for many one-way-door questions you might get back in Claude Code, you can actually just say Take all recommendations and be happy.' His Steve Jobs story (138 likes): font city names weren't bad, but 'at least call them world-class cities' — that's how Paoli and Wynwood became New York and Geneva. Nikunj's /goal one-shot (11 likes): 'may not be the MOST token efficient thing, but it's a thing of beauty to see it one-shot an extremely detailed spec (with generous CLI tools) in 14h.'

Steipete's two workflow upgrades: (111 likes) 'Added a short instruction to our shared AGENTS MD file to upload videos to each PR that changes UI state.' And (160 likes) 'We moved the team over to build openclaw with openclaw. Being able to share agent sessions as URLs is a superpower.' Matt Turck's workday evolution (166 likes): 'Before AI: [decision][process][process]... 10pm: still going. With AI: [decision][decision][decision][decision] 3pm: [brain empty][need coffee][staring at wall].' Peter Yang digs into X's open-source anti-slop algorithm (35 likes): a behavioral model called TweetSpamBot analyzing up to 512 recent account actions — posting bursts, quote-post behavior, timing, dwell patterns.

Amjad on personal apps (82 likes): 'Even if you don't plan on publishing to the App Store, building personal apps via TestFlight is really great.' Aditya celebrates India (22 likes) and praises his SPC partner (78 likes): 'as investors we must deeply embody the traits we look for in founders.' Nan Yu's AI-shape take (5 likes): 'AI is not jagged, it is just AI-shaped. That's like saying dogs are jagged because they are similar, superior, or inferior compared to humans on specific tasks.'

Matt Turck interviews Mitch Troyanovsky, cofounder of Basis — the unicorn whose agents run autonomously for hours or days and complete entire tax returns end to end. A reference episode on long-horizon agent building: whispering context into microphones (speech beats writing for context transfer); accounting as 'intelligence over the economy' — a compression of real-world events into structured decisions; agents defined as a spectrum of agency, with long-horizon meaning coherence past the LLM's working memory; autonomy meaning 'I'm done' as a first-pass deliverable with decisions and assumptions for review; verification borrowed from how human firms already work (trial balances summing to zero, Excel without errors, independent review, judge-based rubrics); data scarcity and long feedback loops (a 1065 takes a human 20+ hours, thousands of inference steps, 500-1000 documents); process over outcomes — 100 passing evals don't generalize, and citing primary sources matters even when the answer is right ('if a person gets it right by going to Wikipedia, the accounting firm wouldn't hire them'); behavior specs — markdown files that are both spec and rubric, written by accountants + ML researchers, deliberately NOT shown to the agent; the magic-box mental model; LLM intuition built by automating your own work ('nothing paradigm-shifting has changed since o3'); an open-source specs standard with Braintrust; ontologies as the agent's world (coding agents don't own their runtime training data — non-coding agents do); canonical docs vs Gong history; hiring language architects and agent managers (systems thinking, law as practice); the DI (deployed intelligence) team; closing the self-improvement loop by year-end; English being more precious than code ('context affects performance, code does not'); assuming the harness gets swallowed by models in under five years; and why technical moats aren't real moats — business moats are.