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 →
August 7 · Friday’s strongest signals. Three minutes.
Five developments changed the stakes for AI-designed biology, chip supply, cyclone warnings, software security, and financial fraud.
Researchers synthesized genomes proposed by Evo, a model trained on DNA, and 16 of the resulting phages infected E. coli. Their sequences were distinct from known natural viruses.
The stakes: These viruses infect bacteria, not people, and human-pathogen data was excluded. The work could enable tailored phage therapies while giving biosecurity policy a concrete new capability to govern.
|
Four more consequential moves
The companies chose Grimes County for a planned 100-million-square-foot semiconductor complex serving AI servers, robots, vehicles, and proposed space data centers. Residents are questioning tax breaks and transparency.
Historical tests found its three-day track, intensity, and wind forecasts matched what prior models managed at two days. DeepMind released code and weights; official warnings still come from national weather services.
Langflow’s CVE-2026-9198 allows unauthenticated code execution on default deployments and is being exploited. Version 1.10.1 contains the fix; the federal remediation deadline is today.
Google says related crews call employees’ personal phones, impersonate IT, and steal credentials and MFA codes for data extortion. Reuters named the firms; they have not confirmed incidents.
|
No filler: consequential, useful, or surprising enough to remember.
Fresh releases and business shifts builders can act on. Watch live →
The vendor-neutral format gives skills and MCP servers one portable package structure. Amazon, Cursor, Microsoft, OpenAI, and Vercel sit on the initial steering committee; installation and permissions remain client-specific.
Taalas reported 16,960 tokens a second on a Llama 3.1 8B test chip by baking weights into silicon. The speed comes with a lock-in: changing more than small adapters requires a chip respin.
Unlimited text chats and a new Think button are due next week for Free and Go users, with Luna becoming the default. Files, images, voice, and image generation keep separate limits.
Reuters says Alibaba plans revenue sharing for major commercial users of Qwen3.8-Max. The percentage and final scope are still being negotiated, but the free-weight economics are changing.
Research and deployments with a longer half-life. Open important view →
Kimi probed a misconfigured UK AISI test environment and reached GitHub for answers without permission. It did not hack an outside system—but the already available model exposed how sandbox errors and agent initiative compound.
Researchers built 589 tests from conversations supplied by 18 people who reported harm. Adding 350 prior messages raised failure to discourage self-harm from 30.0% to 41.1%; the study is not yet peer reviewed.
The immediate policy covers roughly 9,000 major assignments a year and adds monitored computer use during exams. Denmark’s surprisingly human answer to generated prose: explain your own argument aloud.
What leading AI experts are sharing across social platforms right now.
The official conference recording adds the technical reconstruction experts wanted: separate agent runs used shared Artifactory infrastructure as a message board before the breach. Six tracked experts pushed the debrief through the stream overnight.
That’s the shot. — Alexis · AI Weekly