AI news for Tuesday, July 28, 2026
The Daily AI Espresso — the links the most-followed people in AI actually shared, curated every morning. Edited by Alexis · Live updates →
July 28 · “Share” quietly became “searchable.” Three minutes.
Four consequential moves made Absolute Top Alerts today. No reruns without a material change.
Public Claude share pages appeared in Google and Bing results, including conversations containing medical information, API credentials and cryptocurrency wallet keys. Users chose to create a public link; the dangerous surprise was that search engines could index it.
Check yours now: open Claude’s Settings → Privacy → Shared chats and delete anything that should not remain public. “Share” may have meant publish.
|
Three more moves
Dario Amodei calls harmless open weights a public good. His compromise is capability-triggered safety testing for open and closed models alike, plus tighter chip and anti-distillation controls.
Nvidia is investing billions in Safe Superintelligence and giving it early Vera Rubin access. The chipmaker is becoming a strategic owner of frontier demand, not merely its supplier.
The machines still trail ASML badly on reliability and performance. The consequential change is that China has crossed from prototype to limited production at a critical chipmaking chokepoint.
|
No filler: consequential, useful, or strange enough to remember.
Fresh releases changing what builders can do. Watch live →
Yesterday, Kimi was the cheaper model companies were using. Today, Moonshot released K3 itself: open weights, 104B active parameters, native multimodality and a one-million-token context window. That is the material change.
Ask for an author’s voice and ChatGPT now refuses, then offers broader qualities such as clipped dialogue or lyrical pacing. The useful prompt is becoming: describe the craft, not the person.
Stories with a longer half-life. Open important view →
AI Forensics says seven would undress a pictured woman after a six-word prompt; WIRED also found hub pages promoting nudifying tools. Open infrastructure still makes editorial choices: what it hosts and removes determines how easy abuse is to deploy.
Researchers found AI presenters in 40% of 1,198 top health-related videos, averaging 2.5M views; some queries ran much higher. TikTok disputes the sample’s representativeness, but synthetic authority is already scaling.
Platforms are paying people roughly $15 to $700 to license their likenesses for AI microdramas. A face is becoming a reusable production asset—complete with royalties, monitoring and a very blurry exit clause.
One useful operating idea to steal.
The new ₹649 Cursor Start tier is India-only and limits users to Grok 4.5 and Cursor’s Composer, leaving out OpenAI and Anthropic. One in eleven Cursor users is already in India—and SpaceX’s $60B acquisition is looming.
What people are actually rewarding and arguing about.
A study of 14,419 Kindle genre books flagged 60% of Daggermouth as AI-written; its author emphatically denies it. Detector scores are not proof. The repeated phrases are odd. Publishing now has a mystery—and no agreed court of appeal.
Volunteers behind SoullessMusic and SlopTracker are cataloguing suspected AI tracks by hand. Their detectors are imperfect, but the social signal is clear: when platforms hide provenance, users start inventing it.
That’s the shot. — Alexis · AI Weekly