Texas law enforcement used AI to draft documentation of their own AI-assisted surveillance operation targeting a woman who had an abortion. It's a closed loop of automated authoritarianism — and it's already normalized inside police departments.
Anthropic dropped Claude Fable 5.1 with a claimed 52.6% on the new Terminal Bench Science benchmark. Willison's real-world tests confirm it's genuinely strong on coding, though he spent his first session getting a pelican animated. Good priorities.
Google's Gemini 3.8 Flash landed with thinking-level controls and 73.7% on DeepSWE 1.1 — a legitimately strong coding benchmark score. Demis Hassabis called it "relentless progress" and he's not wrong; this is the second major Flash upgrade in under a month.
Andreessen Horowitz published a piece arguing the platforms are fine and users are happy, right as the firm simultaneously funds AI tools to scrape, summarize, and replace those same platforms. The timing is not subtle.
Officers privately asked Axon to redesign its Outpost cameras to avoid looking like Flock units — which communities have been destroying on sight. Public resistance to license plate readers is forcing hardware-level camouflage as a product feature.
Former Google and Anthropic researchers at Mirendil are building self-accelerating AI — systems that meaningfully contribute to their own development. DAIR.AI's feed is already flagging confirmed signals that recent Gemini Flash models may be early evidence of a recursive self-improvement flywheel in production.
Hassabis announced Gemini 3.8 Flash with characteristic understatement, but the signal underneath is loud: two frontier model upgrades in under a month, with a dedicated cybersecurity variant (3.8 Flash Cyber) shipped alongside. The pace is no longer annual or quarterly — it's continuous.
The Neuron flagged a sharp critique from Ramp Labs: companies are measuring AI agents by tokens consumed and sessions run — inputs, not outcomes. If your AI metric doesn't capture whether useful work actually happened, you're optimizing for activity theater.
The 60+ day delinquency rate on US subprime auto loans hit ~5.2% — the highest on record, more than doubled in two years. Markets are selling off mildly today but this is the kind of slow-motion data point that precedes something less mild.
Miles Deutscher made the case that any serious market participant should have two or three AI research agents running continuously — arguing manual crypto research is simply no longer competitive. It reads like hyperbole until you check Gemini 3.8 Flash's benchmark scores.