Join free with your email
Free. No spam, ever — we'll never share your email address and you can opt out at any time. Déjà abonné ? Connectez-vous
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
Build a personalised Espresso around the topics, companies, and people you actually care about. We’ll keep the rest out of your cup—and send it weekly or whenever enough important news brews.
Build my personalised Espresso →1. The moat is moving below the model. Custom silicon and local execution let companies optimize cost, latency, and privacy as one system; access to a strong model alone is becoming less defensible.
2. The next productivity gain is less setup, not better chat. Persistent memory and one-video robot learning both attack the recurring cost of teaching an AI what you mean.
3. AI’s adoption ceiling is becoming social and physical, not technical. Cheap intelligence still needs power, public trust, and a credible answer for the human work it displaces.
Three consequential moves made Absolute Top Alerts: OpenAI’s custom inference chip challenges Nvidia, Skild teaches robots from one video, and Apple ships its first 2nm M6.
OpenAI reports 1.5–1.9× more work per watt across three public models, and SemiAnalysis says selected lab runs beat July Rubin figures. It is real silicon—but still engineering samples on easier workloads, with no harder AgentX result yet.
Skild showed S1 one video, then had it pot a plant, cook a pancake, brew coffee, and assemble a kit—tasks it says were unseen in training. The company reports a 7× unseen-task gain; independent replication is still missing.
M6 is Apple’s first 2nm chip; M5 Ultra is its first quad-die M-series design, with up to 512GB unified memory. Apple claims 4.5× M3 Ultra’s peak GPU AI compute. Both ship in new Macs rather than sitting on a roadmap.
AI news filtered through what leading AI experts are reading and sharing—we follow them so you do not have to.
Chat and Cowork now write to the same memory, and it is on by default for Free, Pro, and Max. You can edit or delete topic files; sensitive topics stay off unless enabled. A chat detail can now shape later cloud-executed work.
Portable Computer runs Perplexity’s agent locally on DGX Spark with a 27B Qwen or PPLX model, keeping files and routine work on-device without credits. It can escalate only when authorized; RTX Linux PC support is coming, not yet shipped.
Z.ai told Bloomberg that the anonymous Ox Alpha model is a new GLM iteration and said it would release the weights later August 26. Its free OpenRouter preview topped usage charts, but the checkpoint was still a promise when this issue was reviewed.
Global Energy Monitor says US gas projects under construction jumped 76% in six months and now double China’s capacity. Roughly half the wider US pipeline is tied to data centers—locking AI’s growth into emissions, fuel prices, and local power fights.
OpenAI says a Russian-origin cluster promoted a fake Israeli “expert community” across major platforms. In its sample, 34 of 36 institute articles were copied elsewhere. The reach was small; the elaborate mix of plagiarism, false experts, and AI promotion is the warning.
In Pew’s 3,488-adult survey, 34% had used chatbots for health; a quarter used them to interpret symptoms. Among those users, 47% called the information very helpful. That measures adoption and perception—not whether the medical answers were correct.
Amazon told workers and requesters the 20-year-old microtask platform will close September 30. “Artificial artificial intelligence” supplied labels, surveys, and evaluations across machine learning; its end displaces workers while closing a foundational chapter of AI’s hidden human labor.
What is surging across social platforms today—verified against the post before the conversation outruns the facts.
The prompt extracts a photo’s texture, light, and rhythm, then reorganizes them into a minimal natural silhouette. It is less “copy this style” than a shareable design procedure—and the post includes the full prompt, examples, critiques, and improvements.
That’s today’s shot. — Alexis · AI Weekly
