Four X signals: Europe has no frontier lab, GPT Astra drives a humanoid through an unseen kitchen, and Claude draws a pixel-art party

In the 24 hours ending September 27, 2026, four items from two of the fourteen AI and tech accounts this channel tracks were worth carrying, and every one of them starts from a post. Ethan Mollick counted Europe's missing frontier lab and Mistral's retreat from building one; a Stanford team put GPT Astra in charge of a humanoid in a kitchen it had never entered; Mollick returned with how Claude's personality changed between Opus releases; and Simon Willison had one model draw a pixel-art kākāpō party that another model filmed. Retweets without comment, replies, jokes, and posts about something other than AI or technology are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

Europe has no frontier AI lab, and Mistral is selling a Chinese model X post · 2,507 likes, 116,149 views at capture · September 26

Ethan Mollick set the arguments about causes aside and named the outcome, which he called shocking: Europe "does not have a single frontier AI lab, nor even a near-frontier lab nor even an effort that could likely lead to building up a frontier lab in the future" 1.

A follow-up post the same afternoon took up Mistral, the French company the reporting he cites calls Europe's great hope in the AI race. Mollick wrote that Mistral has largely pivoted from creating frontier models, and that its current models sit far from the frontier even among open-weight releases 2. Mollick attached a South China Morning Post article on Mistral's announcement of August 11, when the company said it would start offering other companies' models inside infrastructure kept under European control, and named GLM-5.2 as its model of choice 3.

GLM-5.2 is an open-weight model from the Chinese company Zhipu, also known as Z.ai, which the article counts among China's six "AI tiger" companies and which has sat on the United States entity list since 2025 over concerns about military modernisation [cite:3]. The article's own framing is that European political and commercial incentives pull in different directions here, and that the push for technology sovereignty ends up running through a Chinese model [cite:3]. Both of Mollick's posts are his reading of that reporting, which was published on August 21, so what they describe is where Europe stood in late September rather than anything that changed this week.

A frontier model now drives a humanoid around a kitchen it has never seen Quote post · 779 likes, 72,562 views at capture · September 27

HomeBody, a September 2026 project from Stanford's Movement Lab with Caltech, puts GPT Astra in charge of a Unitree G1 humanoid 4. The robot first explores the kitchen on its own, collecting phone video, camera frames, LiDAR scans and joint poses, and the same model turns that record into a digital twin of the room. At task time Astra picks from a short library of skills — navigate, pick, place, open a drawer, reach into a drawer — and reads the result of each one before choosing the next, all of it without environment-specific training data or extra policy learning 5.

The question the authors set out to answer is whether a frontier vision-language model still needs a learned action model underneath it. HomeBody drops that layer and lets Astra call the skills directly through structured tool calls, which makes the model replaceable and the skill library extensible [cite:4]. The demonstrations are a clean-up that spans the room and a request to fetch medicine from a drawer the robot had only seen earlier; the local stack runs on one laptop with an RTX 4090 GPU while Astra runs remotely [cite:4]. The project page also lists what the setup costs: building the digital twin takes time and API calls, the finger servos overheat during long runs, and Astra's reasoning latency adds pauses between skills [cite:4].

Ethan Mollick quoted the announcement with his own reading of it. "It is strange how much LLMs turned out to be the key to such a wide range of problems that would not, initially, seem to be problems that a model of human language would be able to solve," he wrote 6. He then invited readers to argue in the replies about "pure LLMs" against multimodal models with tool use, saying he takes the caveats 7, and closed the thought a few hours later: "Anyhow, it is almost like some sort of general intelligence, but made with artificial means. Very hard to know what we might call that" 8.

Model personality matters in a way benchmarks can't capture X post · 333 likes, 18,109 views at capture · September 27

Ethan Mollick described a change in the Claude Opus models that he noticed across releases. The versions between 4.7 and 5 "just didn't feel 'Claude-y' anymore, more like an watered-down Fable," he wrote, and Opus 5.5 "feels like working with ol' Claude again" 9.

Mollick offers a guess at the cause in passing, wondering in parentheses whether the intermediate models were trained on another model's output, "teacher models?" [cite:9]. He reports no evaluation behind the observation, which rests on how the models feel to work with from one month to the next.

Nothing in the post ties that shift to a measurement, and Mollick supplies none. What he is describing is the judgment a subscriber makes when they decide which model to keep paying for — how it drafts a reply, how often it needs correcting, whether the exchange is pleasant. The post drew 48 replies [cite:9].

A model drew a pixel-art kākāpō party and a coding agent filmed it for a keynote X post · 75 likes, 12,892 views at capture · September 27

Simon Willison closed the WeAreDevelopers World Congress North America with a talk that referenced this year's record-breaking kākāpō breeding season, and built the animation for his final slide by dropping three kākāpō photographs into Claude with a prompt asking for an HTML5 canvas animation of at least twenty pixel-art kākāpō jumping with confetti 10. Willison published the conversation transcript and the finished page, which plays the animation and throws more confetti each time a viewer clicks it 11.

Getting the animation into Keynote took a second tool. Willison downloaded the HTML file and asked a local Claude Code session for a fifteen-second video of it, clicking a few times to trigger the confetti and spreading those clicks around the page 12. Claude Code drove the page with Playwright and recorded the result; Willison published the script, which clicks ten times between the third and thirteenth seconds of the recording [cite:10].

The write-up carries the prompt, a share link to the conversation, the interactive page and the Playwright script, which is everything a reader would need to run the same two-step pipeline on a subject of their own [cite:10] 13. Willison's own account of why it worked is that Claude Opus 5.5 "is surprisingly capable at pixel art animation" [cite:12].