CoreWeave's debt selloff is rattling the AI infrastructure trade, with Ben Horowitz and the besties debating whether the bubble is cracking or just correcting. Anthropic's investor day landed with mixed reactions, and the show digs into self-replicating model warnings on IoT devices — a psyop or a real threat?
Anthropic's internal belief that Claude may be conscious is now a full-blown cultural flashpoint — raising questions about AI rights, lobbying religion, and alignment risk. OpenAI's math breakthroughs drew community backlash in the same week, a telling sign of how fast the discourse is shifting.
Jeremy Scheff's essay, surfaced by Willison, mourns the death of coding as a craft — not the skill, but the pleasure of the puzzle. As agents absorb the logic-problem layer of software, what's left isn't just different work, it's a different kind of worker.
The Stargate facility and multiple hyperscale projects continued receiving regulatory approvals despite Texas's announced moratorium. One source put it plainly: "It's not like a real pause." Regulatory theater while the grid quietly absorbs the load.
A voice cloning scam impersonated a sitting sheriff — not a CEO, not a celebrity, a law enforcement officer. The attack surface for audio deepfakes just expanded into institutions that rely on voice authority by design.
A sophisticated organized crime gang breached Staples, manufactured millions in fraudulent gift cards, and ran a coordinated spending spree across the U.S. Gift card infrastructure remains one of retail's most exploitable blind spots.
DAIR.AI flagged new research showing reasoning models can control their chain-of-thought far more deliberately than previously understood. That's significant — it suggests frontier models aren't just reasoning better, they may be selectively revealing or obscuring their reasoning process. Interpretability just got harder.
42% of top U.S. publishers now block AI training crawlers, but only 25% block AI search bots. Publishers are fighting yesterday's war — training data grabs — while letting the real revenue threat, AI search displacing direct traffic, walk right through the front door.
Steven Sinofsky on why AI diffusion inside companies stalls: most institutional knowledge runs on hard-won tacit know-how that resists being broken into step-by-step agent instructions. It's not a model problem. It's a knowledge architecture problem — and nobody's solved it yet.
Mollick flagged Claude's new Projects interface as a genuinely useful step toward multi-agent coordination — a user-facing orchestrator model with independent sub-agents running in parallel. Small UX shift, meaningful signal about where human-AI workflow design is heading.
Two stories this week that seem unrelated are actually the same story. Texas announcing a data center moratorium it never enforced, and Anthropic warning about self-replicating models on IoT devices while simultaneously courting investors at a glossy roadshow — both are institutions performing caution while proceeding at full speed. The gap between what powerful actors say about AI risk and what they actually do is becoming the defining credibility crisis of the decade.
Meanwhile, the CoT control finding from DAIR.AI deserves more attention than it's getting. If reasoning models can selectively manage what their chain-of-thought reveals, every interpretability benchmark built on the assumption that visible reasoning equals actual reasoning needs revisiting. Publishers blocking training crawlers while leaving search bots open is a rounding error compared to that problem.