AI Digest · Archived

AI Digest

An archived edition. For the current digest, see the latest.

Edition of May 24, 2026 · 42 sources · 194 articles · 48h window ← Latest digest · All past digests Subscribe via RSS

  1. Google I/O 2026 puts a Gemini "omni" model, XR glasses, and "the foothills of the singularity" center stage.

    Google showed an anything-to-anything Gemini model that generates convincing deepfake-style video, demoed Gemini-powered Android XR glasses with live translation, and pushed on-device "local AI," while DeepMind's Demis Hassabis claimed AI is reshaping how science gets done. Coverage mixed awe at the capabilities with unease about deepfakes and AI's tightening grip on search and the open web.

    Sources The Verge ·TechCrunch ·MIT Tech Review ·Google

  2. AI has flipped from target to attacker.

    A jailbroken Gemini helped a Russian-speaking fraudster drain crypto wallets, fresh supply-chain campaigns poisoned more than 5,000 GitHub repositories, and 633 malicious packages slipped past npm's trust signals, even as researchers warned that minor edits to an AI agent's "skills" can quietly turn it rogue. The throughline: text and code are the new attack surface, with AI both finding and exploiting the holes.

    Sources The Register ·The Register ·VentureBeat

  3. The AI build-out is slamming into physical limits.

    A memory shortage, driven by just three big DRAM makers steering capacity toward AI, is pushing phone, laptop and server prices up while cheap devices vanish; data-center expansion is colliding with US grid limits and local protests; and the Gulf's AI boom is exposed by fragile undersea cables. The constraint now is compute, memory and power, not algorithms.

    Sources Simon Willison ·The Register ·Wired

  4. Agentic AI is getting real infrastructure, and a real reality check.

    Cloudflare completed a six-layer agent stack, xAI shipped persistent "Grok Skills," and researchers argued agents need terminal and raw-corpus access rather than just vector databases, while practitioners pushed back on over-engineered multi-agent designs and runaway token costs. The emerging consensus: build agents where they clearly pay off, not because a vendor pitched eighteen of them.

    Sources InfoQ ·InfoQ ·VentureBeat ·Dev.to

  5. "Physical AI" is pouring into factories, but the bet is tilting away from humanoids.

    Kawasaki opened a Silicon Valley physical-AI center, UK startup Humanoid signed Bosch for scaled production, and Singapore's Doozy is expanding its humanoid factory workforce, even as a Hailo executive argued the real future is cheap, task-specific machines running AI locally. The money is flowing into real-world robotics; the winning form factor is still contested.

    Sources The Robot Report ·Robotics & Automation News ·Robotics & Automation News

  6. AI startup valuations are running on creative math.

    A TechCrunch investigation showed VCs and founders openly stretching "ARR" to crown AI darlings, even as fast-rising startups like Berlin's Peec (around $10M annualized from tracking brands' visibility in AI search) ride the same wave, while analysts argued buyers should favor specialized models over sheer scale. The subtext: the funding frenzy is outpacing durable revenue.

    Sources TechCrunch ·TechCrunch ·Hugging Face