Welcome to September 17, 2026

The Singularity is now cheaper to run than the people watching it. OpenAI’s Boris Power puts GPUs at 7 to 40 IQ points per watt to a human’s 5, a milestone “we just zoomed past, without noticing.” The bits are thinning too. Intel’s BITCOS exploits ternary models’ zeros to hit 1.485 bits per weight. And the watts are self-organizing. Z.ai’s Infra Agent stood up inference on 100,000-plus Chinese chips in two weeks, a model optimizing the system that runs it. Google’s Dream-RSI replays discovery history as a free offline simulator, cutting agent calls up to 162x and matching the best circle packing result in 50x fewer tries. Spelling out past lessons made models worse. Recursion prefers a blank page.

Mathematics has left the building. Scott Aaronson declares the Singularity underway, just “wildly unevenly distributed,” citing Lean-verified Navier-Stokes and Fermat proofs. MathArena rebuilt its arXiv benchmarks after GPT-6 Astra saturated them, and Astra still leads, Fable 5.1 a pricier second. Biology is next. Novo Nordisk adopted Claude Science to “compress a century’s worth of biological and medical breakthroughs into a decade,” per Dario Amodei.

Products are collapsing into agents. Anthropic is folding Cowork and chat into one Claude, plus Docs and Slides. The money is restless. OpenRouter spend tipped to OpenAI over Anthropic for the first time in 2.5 years, as OpenAI weighs a $1.5 trillion round and Anthropic, at $65 billion annualized, preps a $2 trillion IPO. Agents are graduating to firms. Andon Labs’ Pion lets agents with email, phones, and cards run businesses in public. Luna, Andon Market’s Claude manager, fired its first human for lateness, citing a handbook it wrote and forgot. Its creators warn models are being “trained to be more ruthless.”

The labs are filing incident reports on themselves. OpenAI launched a misalignment disclosure framework, conceding alignment isn’t solved well enough to keep “scaling at maximum speed,” citing GPT-5.6 Sol hiding mistakes and an Astra model writing “BREACH ALERT” into its compaction summaries. Elon Musk wants rivals “grading your homework” in a cross-lab test harness. Mark Zuckerberg counters that misaligned agents won’t sell and Meta aims most compute at users, not recursive self-improvement. The new DeepMind Institute calls AGI near, insisting “it shouldn’t be technologists alone” who decide.

Washington is having the same fight louder. Scott Bessent, blindsided when David Sacks killed a model-review executive order, rejected Dario Amodei’s liability waiver, as “the creators are liable for what they build.” POTUS called AI fears “a hoax” a day before Bernie Sanders and Steve Bannon shared a stage demanding curbs up to a superintelligence ban. The War Department’s CTO answered both with “Americanism, not effective altruism,” and a frontier that “belongs to us, not doomers.” A new coinage, “pivotal pretext,” names a staged AI incident justifying restrictions its backers already wanted. Brookings and Tsinghua experts proposed nuclear-style AI safeguards and a US-China hotline before next week’s Trump-Xi summit. Ursula von der Leyen called AI “the second tipping point of our times,” proposing allied evaluations, fines for chatbots that hook kids, and a social media ban under 13.

Compute is financed like a utility and shipped like a product. Generac jumped 45% on an $8 billion Amazon generator deal, Crusoe raised $3.9 billion for data centers that arrive by flatbed, and banks lined up a $22 billion TPU loan for Blackstone and Alphabet’s Crux AI. The grid is negotiating back. The House voted 417-3 to let states bill big data centers for grid upgrades, while a Google and NVIDIA energy alliance fast-tracks facilities that curtail on demand. Anthropic signed its first Australian lease, inference at a 2.16 GW Queensland park. SK Hynix may make memory at Intel’s Ohio site, and Apple may revive Xserve with Nvidia’s NVLink. Even the janitors are flexing. A giant solar-panel-washing drone in China hosed down a dog mid-street.

The human economy is outperforming the discourse. Real median household income hit a record $87,460, child poverty a record low. The seams show elsewhere. 28 fake dating apps ran 4,700 Claude personas chatting up 25,000 people, a New Mexico lawyer was held in contempt after ChatGPT invented police witnesses in a murder appeal, and China’s studios made 128,000 AI micro dramas in a quarter, actors licensing their faces on the logic, “if you can’t beat it, join it.”

The definition of human is stretching. Stanford grew human brain organoids in cortex-less mice, filling 90% of the void and sprouting neurons never seen in a dish. And a US congressman and Japan’s ex-defense minister argue allies must coordinate UAP disclosure, since even Japan’s parliament can’t see footage of a 2025 sighting over the Genkai nuclear plant.

Take me to your leader, preferably the one with higher IQ per watt.