Someone embedded hidden AI instructions directly inside a legal document — essentially trying to manipulate any AI system summarizing or analyzing the case. It's the first documented case of prompt injection entering the court record, and it won't be the last.
Meta's latest patent describes smart glasses that recognize faces in real time and auto-generate highlight reels of social gatherings. The gap between "patent" and "product" is closing fast — Meta's Ray-Ban glasses already have cameras and AI.
Alibaba's new 27B open-weight model runs well on a laptop and punches above its size — but its extended thinking mode burns tokens on trivial tasks. Willison flags it as a genuine local model contender if you know how to tune it.
Amodei argues the public's negative view of AI isn't caused by safety warnings — it's a fundamental crisis of trust in companies, governments, and tech broadly. A rare moment of honest self-awareness from a lab CEO worth sitting with.
Hibernating mice retain memories despite massive synaptic loss during near-total brain shutdown, upending assumptions about how memory is physically stored. If memories don't live in synapses, the architecture of intelligence — biological and artificial — needs rethinking.
Former Google and Anthropic researchers at Mirendil are building systems where AI meaningfully contributes to its own development loop. Self-accelerating AI is moving from thought experiment to funded startup.
The top 1% of US businesses spent a record median of $7,400 per employee per month on AI in July. Pair that with retail sales dropping 0.6% — the steepest monthly fall since May 2025 — and you get a split-screen economy: enterprise AI spending accelerates while consumer spending contracts. Companies are betting the productivity gains show up before the recession does.
Ethan Mollick wants more precision when AI leaders claim AI will "cure diseases" — noting the phrase usually means "a nation of geniuses in a data center will invent things faster." It's a sharp call for the industry to stop hiding vague techno-optimism behind medical promises that haven't been earned yet.
New IBM research finds that benchmark score differences between models are partially artifacts of how prompts are phrased — not true capability gaps. This quietly undermines the entire leaderboard culture that drives model selection decisions at companies spending thousands per employee per month on AI tools.
Two stories this week share a hidden-variable problem. The mouse memory research shows that what we thought was the physical seat of memory — synapses — may just be one layer of a more distributed system. The IBM benchmark research shows that what we thought measured AI capability — leaderboard scores — may just be measuring prompt sensitivity. In both cases, the thing we're measuring isn't quite the thing we think we're measuring.
Meanwhile, Meta's facial recognition glasses patent and the Mirendil self-improving AI story point in the same direction: the infrastructure of ambient intelligence is being laid quietly, patent by patent, startup by startup. Dario Amodei is right that this is a trust crisis. But trust isn't rebuilt with quotes — it's rebuilt when the glasses don't ship, or when they do and nothing bad happens. We're about to find out which it is.