Sunday, May 24
AI Stumbles, Surveillance Creeps
The gap between AI ambition and real-world execution is widening in public view. Starbucks has scrapped its AI inventory tool (opens in a new tab) across all of North America, a quiet admission that the technology wasn't delivering operational value at scale. Ubisoft is faring no better creatively — an insider reports that generative AI integrated into Far Cry 7 (opens in a new tab)'s development pipeline looks, in their words, like garbage. Meanwhile, Google's new Gemini Omni model, billed as an anything-to-anything AI capable of generating video and audio together, is simultaneously impressive and unsettling — a reminder that the frontier keeps moving even as enterprise deployments stumble.
Passive surveillance is getting harder to opt out of. Researchers have demonstrated that ordinary WiFi signals can now identify individuals (opens in a new tab) with near-perfect accuracy using gait and body shape analysis — no camera required, no consent requested. Separately, the FBI is pushing for near real-time access (opens in a new tab) to the United States' sprawling network of license plate readers, a request that would give federal agents continuous visibility into the movement of millions of Americans. Together, these developments mark a practical shift: the infrastructure for pervasive, ambient identification already exists, and the institutions that want to use it are moving faster than the legal frameworks designed to constrain them.
Hobbyist hardware experimentation is quietly pushing the boundary of what local AI inference looks like. A builder has managed to run a 1-trillion-parameter version of Kimi K2.5 on a single-GPU system (opens in a new tab) by stacking 768GB of cheap Intel Optane DIMM memory, achieving roughly 4 tokens per second — slow, but functional, and achieved without a data center. On a different front, Apple has open-sourced LiTo (opens in a new tab), a 3D AI generation model, signaling that the company is willing to contribute generative tools to the research commons even as it keeps its core products closed. These moves collectively push inference and generation capabilities toward the edge, where cost and accessibility matter far more than benchmark scores.