A newly published Meta patent describes AI glasses that use facial recognition to silently identify people around you and auto-generate highlight reels of social events. This isn't a concept — it's a filed patent on a product line Meta is actively shipping. The privacy implications of ambient facial recognition at scale are severe and largely unregulated.
Someone embedded hidden AI instructions into a court document, telling any model that processed it to rule in their favor. It's the first confirmed instance of adversarial prompt injection weaponized in a legal context. Courts relying on AI-assisted review now have a documented attack vector to worry about.
Mice retain memories through hibernation despite massive synaptic loss and near-total neural shutdown, suggesting memory isn't stored where neuroscience thought it was. The findings challenge the dominant model of how biological memory works — with direct implications for AI memory architecture research.
Mirendil cofounders — ex-Google and Anthropic researchers — are building systems where AI meaningfully contributes to its own training loop. The conversation moves past productivity framing into something harder: what does oversight look like when the model is part of its own development team?
New research shows self-improving agents that log successful behaviors can accidentally canonize unsafe ones — an "unsafe success" becomes a reusable skill. This is a quiet, compounding failure mode with no obvious tripwire.
A16z's Charts of the Week reveal a stunning adoption gap: the heaviest AI spenders are dropping 600 times more than the typical company. The enterprise AI market is bifurcating fast, with a small cohort pulling radically ahead.
Andrew Ng published a map of the most important skills in AI Engineering — a rare, structured attempt to define what the discipline actually requires. In a field where everyone claims AI expertise, a credible skill taxonomy matters. Expect this to circulate through hiring teams and bootcamps for months.
Ethan Mollick drew a precise distinction: we are in a Singularity as Vernor Vinge defined it — a period beyond which the future becomes opaque — but not yet in von Neumann's original formulation, which required a harder discontinuity in the pace of change itself. The framing matters for how seriously policymakers should treat AI timelines.
IBM research flagged that benchmark performance improvements are partly an artifact of prompt phrasing, not model capability. Pick your model based on benchmark deltas and you may be buying gains that disappear the moment you change how you ask the question. A quiet indictment of how the industry currently shops for models.
The Neuron noted that frontier models routinely fail basic probabilistic reasoning — like coin flips — until prompted with expert-level framing, after which scores jump 30 points but still rarely crack 80%. The gap between raw capability and accessible capability remains a fundamental product problem, not just a research one.
Two stories today deserve to be read together: Meta's facial-recognition glasses patent and the prompt injection hidden inside a legal filing. One is about AI that surveils without consent; the other is about manipulating AI systems that have been granted institutional trust. Both are attacks on the same thing — the assumption that AI-mediated environments are neutral.