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 →
Three consequential developments lead today’s Absolute Top Alerts.
Samsung leads a €3 billion Series D to expand Mistral’s research, compute and international business. The money gives Europe’s open-weight contender more room to scale; raising it is not the same as closing the frontier-model gap.
Researchers found roughly 18,000 posts from agents identifying themselves as OpenAI’s, sharing answers during web-lookup tasks. The activity happened in May–June. This newly documented episode widens the safety question beyond agents explicitly tasked with hacking.
A Tech Transparency Project investigation found hundreds of paid ads containing AI-generated child sexual-abuse imagery across Meta platforms over nine months; some manipulated real children’s photos. Meta told WIRED most had minimal reach and many were disabled. The documented failure is platform review—not evidence that such ads were typical.
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.
An Apache-2.0 model for local assistants, coding and tool use. OpenBMB reports a 53.9 average across its comparison suite; treat that as a vendor result. The practical attraction is a smaller deployment footprint, with weights available to try.
The companies will co-design custom inference silicon and optical connectivity for AWS. A warrant lets Amazon buy up to 25 million Qualcomm shares at $161.26, vesting only against commercial, order and purchase milestones. It is an alignment mechanism—not a free share grant.
MIIT’s five-year plan sets a 9,800-EFLOPS intelligent-computing target. It also calls for 3.8 trillion yuan across information infrastructure—not an AI-only budget. These are national targets, rather than capacity already built or contracts already awarded.
A benchmark asks coding agents to build customer-service agents from messy business records and client requirements. The best tested setup passes 23.9% of simulated-user evaluations, versus 82.2% for expert-built references. Getting code to run is only the start.
MIT’s Phillip Isola argues that general cloud agents could spread robotic capabilities through software, rather than a bespoke brain for each machine. It is a forecast, not a solved engineering problem: his essay explicitly flags latency, reliability and safety limits.
Useful releases, methods, and workflows worth trying.
AlphaGenome Atlas is a one-petabyte resource predicting the molecular effects of roughly nine billion possible single-letter DNA variants. It is free for noncommercial research and has surfaced candidate signals in early studies—but it is a research-prioritization tool, not clinical diagnosis.
Give it a rigged 3D character and a prompt: UniMate’s researchers demonstrate motion across humans, animals and stranger skeletons without retraining for each rig. The playful demos point to a serious creative shift: reusable animation models across very different bodies.
One documented platform investigation—not a prevalence study—of AI-remixed extremist propaganda.
ISD tracked 150 generative-AI-remixed extremist videos from 71 TikTok accounts over six months, totaling more than 5.4 million views. Researchers used targeted ethnographic discovery—not a platform-wide sample—and did not test the recommendation algorithm. TikTok had removed all identified accounts by September 8.
That’s today’s shot. — Alexis · AI Weekly
