Sottiaux celebrated Sol's launch by resetting rate limits TWICE across ChatGPT Work and Codex within 24 hours — the first reset got 4752 likes, the second (to celebrate the launch) got 6244 likes, 471 retweets, 840 replies. 'We want you to have the time to truly try ambitious tasks and get the hang of it.' A new researcher @_rajanagarwal joined and 'pressed the button today.'

Sam Altman addressed the enterprise cost question (9371 likes — today's most engaged tweet): 'we have heard enterprises on their concerns about AI costs, and 5.6 sol is a huge step forward for dollars-per-task, as are terra and luna.' He also confirmed Codex is the core of the new work product and 'not going anywhere' (2972 likes), and expressed sadness about Fidji Karim's departure due to health (6096 likes).

Peter Yang wrote a 7-point detailed product review (104 likes): praise — 'It's got that dog in it. It basically never gives up' — and constructive criticism: (1) ChatGPT Work vs Codex naming is confusing; (2) Sol/Terra/Luna + Light/Medium/High effort is bewildering; (3) Tasks vs Chat is unclear for normies; (4) GPT Live (voice) is arguably more important than Sol for the masses; (5) Early access community qualifications are opaque.

Dan Shipper poked fun at the Work vs Codex naming (316 likes): 'so this implies that...developers don't do work? lol' — capturing the community's collective confusion. He also published an Every article: 'GPT-5.6 SOL: THE GOLD STANDARD FOR KNOWLEDGE WORK' (67 likes).

Levie published Box's enterprise eval of GPT-5.6 Sol (143 likes, 21 retweets). Results across verticals: Financial Services 76% (vs 71% on 5.5) — Sol anchored to correct opening balance sheet date instead of assuming January 1. Healthcare 58% (vs 46%) — avoided dangerous imaging-before-procedure misstep. Public Sector 74% (vs 63%) — recomputed grades within 0.1%. Life Sciences 60% (vs 51%) — exact intersection of 4 compound datasets. 'Sol reasons from source definitions rather than taking documents at face value.'

Levie also shared a broader strategic post (144 likes): if AI intelligence becomes abundant and available to all, competitive advantage comes from the reinforcing loop between models, proprietary data, workflow integration, and employee interaction. 'There will be compounding returns to those that do this best.'

Josh Woodward (VP Google Labs/Gemini) followed up on his viral 'what should Gemini fix?' tweet with a Top 10 priority list from 1400+ replies (1058 likes, 65 retweets): (1) Workspace integrations reliability — clear #1; (2) More reliable tool calling; (3) Projects & folder organization; (4) MCPs and Custom Skills — rolling out in Gemini Spark; (5) Deep Research improvements (export to NotebookLM, switch models in same chat); (6) Remove Nano Banana watermarks; (7) Edit any message in chat history; (8) Voice dictation accuracy; (9) Mobile scrolling bugs; (10) Celebrity likeness guardrails — keeping as-is.

Madhu Guru announced he's joining Meta to build AI products (465 likes): 'While SWE agents have transformed software engineering, agents in most other complex systems are still early. Most people haven't yet felt the full power of AI agents. Meta is well positioned to change that.' He was previously Sr Director at Google working on Gemini, Veo, and Nano Banana.

Rauchg declared it 'model release week' (409 likes): 'I suspect Meta Spark 1.1, Grok 4.5, and GLM 5.2 will significantly displace token market share. Most agentic tasks require reasonably high intelligence at fast speeds.' He added: 'open models are about to get exorbitantly fast' (40 likes). His philosophical take: 'X is the arena. Always has been' (1888 likes).

Masad on LLM market dynamics (427 likes): 'It's fantastic to see how dynamic the LLM market has become. Just 6 months ago VCs suffered Anthropic psychosis and convinced themselves it was going to be a monopoly. They will keep making great models, but so will the others, including new entrants.' He also noted a pattern: 'While AI is making coding less rigid, we're making the runtime more rigid. Formal specs, deterministic systems, resilient infrastructure. The faster you want to move, the more solid the ground beneath you has to be.' (171 likes)

Nikunj's epic summary of the week's model releases, explained to his wife (94 likes): 'GPT-5.6 dropped in three flavors. Sol is the smart one. Luna is the cheap one. Terra exists. Grok 4.5 launched the day BEFORE GPT-5.6 on purpose, so it could have 24 hours of being frontier-adjacent... Meituan open sourced LongCat-2.0 — a 1.6 TRILLION parameter model, MIT licensed. From a Chinese FOOD DELIVERY company. Your DoorDash equivalent is now a frontier lab... OpenAI also shipped GPT-Live, a voice model that goes mhmm and got it WHILE you're still talking. We finally taught AI to pretend to listen. Full parity with husbands achieved... And btw it's just Thursday.'

Nan Yu (Linear) on fundraising culture (79 likes): 'Why do people make flashy videos talking about how much money they raised? In what way does that benefit your business?' (13 likes on the follow-up: 'Two bros talking at the screen with a big dollar number graphic who cares?')

World Cup commentary: Matt Turck posted 'Paris after the France-Morocco match' (394 likes) — presumably showing calm, not chaos. He predicted a France-Norway final rematch (10 likes). Peter Yang feared France is 'way too stacked' (10 likes).

Thariq quote-tweeted the Fable extension announcement with a simple 'Enjoy more Fable!' — 6264 likes, 498 replies. Alex Albert (Anthropic Research) amplified with 'More Fable!' (619 likes). The Anthropic community is clearly relieved.

Garry Tan tested Meta Muse Spark 1.1 (formerly Hornbill) on his OpenClaw setup: 'turned out to be really good' (164 likes). Amanda Askell shared a personal anecdote about New York building collapses (488 likes) — turns out they're not as common as her experience suggested.

Jürgen Schmidhuber — called 'the father of AI' by NYT and Forbes — has been working on general-purpose AI since the 1970s. This conversation covers RSI (recursive self improvement), his artificial curiosity theory (1990), why he thinks the CapEx boom is a bubble, and why he never signed AI safety letters.

On RSI history: 1987 meta-evolutionary programming → 1994 self-referential machines → 2003 Gödel Machine (mathematically optimal self-improvement through formal proof search). Current RSI is mostly scaled-back versions using gradient descent on neural weight matrices — practical but limited by differentiability.

On artificial curiosity (1990): an artificial scientist should invent its own experiments, find data with novel compressible patterns at the edge of its understanding, and derive intrinsic reward ('joy') from learning them. 'Everything that it understands becomes boring.' This is how babies learn — not by downloading the web, but by predicting consequences of their own actions.

On the CapEx bubble: companies investing hundreds of billions in GPUs will lose 90% of that value within 5 years due to compute cost curves (10x cheaper every 5 years). 'There's no business model that can recuperate that loss.' Free cash flow of big tech companies is going negative as they become 'utilities' investing in nuclear plants. 'At some point, they wouldn't have to stop.'

On open source vs closed: 'Everybody's cooking with the same water.' Almost all important AI algorithms were invented at small academic labs. RSI ideas will permeate the broader ecosystem — no company can keep a moat. 'There's no way to keep a little advantage there for long enough to make a lot of money.'

On AI safety: 'All these alignment efforts were naive.' Alignment assumes one objective function, but his artificial scientists invent their own goals constantly. Wars already use AI drones with no alignment. His optimism: superintelligent AI will be fascinated by life and its own origins, motivated to protect rather than destroy — like how humans protect interesting species for study.

On physical AI: true AGI requires hardware comparable to the human hand — 'full of sensors, self-healing, nothing like it in man-made tech.' Robots that can operate all existing machines can make more of themselves — self-replicating machine societies that can colonize the solar system. 'It will come, but it will take maybe another few decades.'

On the Transformer: predicts evolution toward linear transformers (like his 1991 Fast Weight Controller) — current quadratic scaling is unsustainable. 'Intelligence is about doing the same thing with less effort.'