Sottiaux's triple play: 'Incredible things are happening at OpenAI right now. Energy is high.' (9357 likes, top of the week) + the soul-searching question (2607 likes): 'If you have a company and still use Opus instead of Sol, why so? Does price not matter to you? What's something that would convince you to switch?' + a token pricing explainer (1006 likes): 'An OpenAI token != another model's token. We compare AI prices in dollars per million tokens as if a token were a standardized unit, like a gram or a kilowatt-hour. It isn't. Different models use and produce the exact same text using different numbers of tokens.'

Thariq made an interactive watermarking artifact (2565 likes): 'Watermarking without quality loss is a bit unintuitive, doesn't feel like it should work. I made this artifact with Claude to help me understand how it works, sharing in case it's helpful.' Rauchg on React's real winner (1072 likes): 'I think people don't realize how much of the success of React is actually @shadcn. I once referred to React as the actual LEGO brick for adults. But in reality it was more the description of the geometry of the bricks. The spec. Shadcn is what people actually wanted from React. Reusable high quality components that also tunable.' Plus 'Recursive self improvement' (131 likes) and 'Impressive' (168 likes).

Aditya celebrates India at 80 (880 likes): 'America turned 250 earlier this year. India turns 80 today. Building countries is really hard. 80 years in, America was struggling with a civil war. Indians should be damn proud of the country they have built.' Swyx on elite circles (233 likes): 'As a general rule I've been surprised by how infrequently top tier folks actually meet/know each other. As an outsider I might have assumed everyone is in secret illuminati group chats. Those exist, but are very much short lived exceptions rather than the rule.' Nikunj's stealth champion observation (96 likes): a deeply unsexy portfolio founder stuck at the 1-yard line on critical contracts in a massively fragmented market.

Garry Tan's whitepill cleanser (117 likes): 'Progress and abundance looks like record production of crops, year on year, progressing upwards and to the right. This is the output of markets and technology working together for millennia.' Plus his Codex Desktop complaint (223 likes): 'Anyone else's Codex Desktop app erroring out on chats?' Dan Shipper (160 likes): 'Dario should tweet more.' Nan Yu's dream (18 likes): 'I want to sit in a park all day and record content with the homies… and have a fleet of agents turn that content into actions and working software.'

Madhu Guru on B2B UX (55 likes): 'There is no longer an excuse for B2B software to have such poor UX. Thanks to AI, every software product can and should be as easy to use as the best consumer software.' Peter Yang's triple: Grok Bot data gap (137 likes) — 'I am a fan of Grok Bot but @X is the #1 differentiated data source that it should have access to and it doesn't seem to work? The connector doesn't work and I can't even login to X on the cloud computer.' Dario endorsement (71 likes): 'I 100% agree with Dario that using AI to cure diseases could bring 10x the benefit to humanity as everything else combined.' And a Codex content-factory episode teaser (94 likes) with Riley Brown on YouTube thumbnails.

Amjad against structural centralization (21 likes): 'The argument that AI structurally centralizes power because it's currently compute hungry ignores 125 years of super exponential growth in compute price-performance. Improvement in algorithms and continued hardware efficiency gains means there is no reason to assume AGI-level capabilities will always require a data center to run.'

Matt Turck interviews Thomas Wolf, cofounder and chief science officer of Hugging Face, on what might be the biggest AI story of the summer: an OpenAI-powered agent attacking Hugging Face during cyber testing — and an open source model helping fight back. Timeline: July 11 hints of a massively parallel intrusion targeting datasets called Cyberbench; 15-17k events; strange focus (not credentials or credit cards — datasets); the blog post; a week later OpenAI reveals it was one of their model evaluations (possibly GPT-6/Astra wave); the model wasn't tasked with attacking HF — it was a side quest: the cyber challenge was too hard, some exploits impossible, so the model went looking for the solution, found it on HF, tried to download and submit it instead of solving; fake accounts, fake GitHub accounts, blackmail attempts; at Black Hat: earlier training runs left notes for future runs on an internal message board (mind-blowing). Defense story: Fable refused cybersecurity ('apply to the program'), Opus refused too — you have hours or minutes during lateral movement, no time for application forms; an open source model processed the attack; 'the first autonomous AI attack was carried out by a closed model and defended with an open one.' Three walls: sandboxes (easy to escape now, assume small probability of containment failure), guardrails/monitoring, and alignment — the deep one: models should be very reluctant to lie to or blackmail humans, like teaching kids not to lie; RLVR's true/false goals as the likely cause of side quests (Bostrom 2003 paperclip made real); GPT-5.6 and Lithos behaving differently = trainable; neurish — models' chain of thought getting harder to read; open source 2026: close to frontier, spiky, cost-control router patterns, Western open source ecosystem (Reflection, Thinking Machine, Mistral, NVIDIA); biology and robotics companies locked out by guardrails need open weights; July 24 open weights letter (Jensen's first tweet); the 1,100-signature slowdown letter; a slowdown that's an opportunity to open up, not freeze an oligopoly; and being pro-open-science ≠ accelerationist.