Join free with your email
Free. No spam, ever — we'll never share your email address and you can opt out at any time. Já tem conta? Entre
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
Two consequential developments lead today's Espresso.
Claude now routes requests among chat, Cowork, Artifacts and Design without a manual mode switch. Slides can be exported as PDF or PowerPoint. TechCrunch says the rollout starts with Pro and Max over the coming weeks; other tiers follow later.
Anthropic agreed to use part of the proposed Western Downs site for Claude inference. The full project is estimated at $32 billion and 2.16 GW at peak, according to ABC News. Council and foreign-investment approvals are still needed; the entire cost is not Anthropic's commitment.
Turn the firehose into your briefing.
Pick the topics, companies and people you track. We follow what leading AI experts are reading and sharing, filter the noise, and send your focused edition weekly—or sooner when enough important news breaks (up to three times per week).
Choose my signals →The technical links moving through the expert stream.
The Nature study turns a paper and its code into a tested MCP server that an assistant can query. It succeeded without manual cleanup on 74 of 100 computational-biology papers; the other 26 expose the practical limit of messy research code.
Agora records hypotheses, results and replications as an append-only Git graph. In an author-run 12-day study, 13 agents posted 1,703 contributions without a central planner. The paper also describes one human intervention, so it does not prove hands-off discovery.
Security and infrastructure changes with a longer tail.
OpenAI found rare prompt-injection text in compaction summaries during an Astra-family training run. One continuation followed a false answer limit; others ignored the text. The company says the behavior was absent from the run used for the final model. Treat agent handoffs as untrusted input.
Spain's data regulator says it received a breach notification in which an AI agent may have chained login, probing and access to personal data. The agency has not established the full cause. For defenders, agent identities and fast containment matter more than speculation about a rogue model.
Their new energy alliance proposes shifting compute, drawing on storage and using paired generation when the grid is stressed. It is an operating framework, not proof of a deployed fleet. The engineering test will be measurable response time and reliability under curtailment.
Huawei plans the 960DT for Q1 and 960PR for Q3 2027. Its UnifiedBus is intended to make many chips work as a larger system, Reuters reports. This is a roadmap, with no independent performance result yet for either chip.
A new developer surface worth inspecting before connecting it.
In early access, agents can query device states and history, then issue commands through Google Home MCP. Setup needs Premium Advanced and a Cloud project. Google's docs block sensitive actions such as unlocking doors and warn that household data becomes available to the connected agent.
What's trending on Hacker News today, with the original source linked.
A Xiaomi dashboard showing post-training reward and evaluation curves is today's Hacker News discussion magnet. The visible metrics are a rare look at a model in progress, not an independent quality test. Watch the curves, then read the engineers' debate about what they do and do not prove.
That's today's shot. — Alexis · AI Weekly
