Join free with your email
Free. No spam, ever — we'll never share your email address and you can opt out at any time. Already a subscriber? Log in
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
One development clears today’s highest-consequence bar.
An analyst fed a Chinese cargo ship’s manifest to a chatbot; it claimed the vessel carried nuclear-weapons components bound for Iran, and that line entered a U.S. intelligence report, according to CNN reporting cited by Timnit Gebru and Emily M. Bender. Their point: error-prone tools can create real escalation risk long before “superintelligence” arrives.
Pick the topics, companies and people you track. We’ll 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 rest of today’s expert-filtered signal, ranked for consequence, utility, surprise, and range.
Claude Code now supports small TypeScript “mods” that can rewrite prompts, block or retry tool calls, decide permissions, redact secrets, and replace interface elements across the agent loop. That turns the coding agent into a programmable runtime—and makes reviewing third-party hooks a security requirement.
Amazon released the Apache 2.0 Strands Decider 2B, a Qwen3.5-based model that returns decisions rather than prose. Amazon reports about 72% accuracy on JevBench v19 and median latency of 106 ms on an RTX 3090, making local workflow routing plausible on modest hardware.
Cloudflare open-sourced Clef and Clef-flash, Apache 2.0 decision models with 64k context and vision input. Its published tests put median latency at 209.3 ms for Clef and p95 at 122.4 ms for Clef-flash—promising cheaper, faster structured agent decisions if independent benchmarks hold.
Google’s exploratory design proposes solar-powered satellite constellations carrying Trillium v6e TPUs in a 650 km sun-synchronous orbit. Two Planet-built prototypes are targeted for early 2027—an audacious bet that space power and cooling can offset the cost of lifting compute off Earth.
Judge Amit Mehta dismissed antitrust suits from Chegg and Penske Media over Google’s AI Overviews, writing that “an expectation is not an agreement.” The ruling acknowledges publisher harm but says the claims did not establish a Sherman Act agreement—pushing the policy fight back toward lawmakers.
U.S. cases describe serial labels stripped with a hair dryer, intercepted Taiwan cargo, and a Singapore mansion linked to networks rerouting restricted Nvidia chips into China. Nvidia says it has strengthened controls; officials are asking whether it should have caught the red flags sooner.
Amazon is discussing transferring about $8 billion of already-deployed Grace Blackwell hardware into an outside-backed vehicle and leasing it back, the Financial Times reports. The proposal could sell up to 10% equity and issue debt—turning AI capex into infrastructure finance without moving the chips.
What tracked AI experts are sharing on social platforms—verified before it reaches your inbox.
After pretraining 800 models, researchers report that synthetic web text helps small, data-starved runs briefly, then reverses into harm as human-data budgets grow. The useful lesson is not “never train on AI text,” but that its value changes with scale—and can become negative much sooner than teams expect.
That’s today’s shot. — Alexis · AI Weekly
