Our Lives

The alignment panic just went mainstream, and the panel expects a Washington firestorm within two weeks. A researcher rendered as “Jacob Coxin,” who said he spent three years on pre-training at OpenAI and Anthropic, quit charging that “neither company is acting responsibly” in the race toward self-improving superintelligence and calling it “gambling with our lives.” His account’s only tweet reached 138.3M views; Elon called it “very strange,” while Emad said it appeared not to have been boosted. Anthropic alignment lead Evan Hubinger publicly put p(doom) above 10% this decade. Emad calls it AI’s “Tom Hanks moment” of COVID, landing near a US–China meeting at the UN and Xi’s visit.

OpenAI’s claimed Navier–Stokes breakthrough with 10,000 agents convinced the panel that “science is thoroughly cooked” — and the real shock is price. Sam Altman called it “the strongest evidence yet” for pacing progress; Alexander Wissner-Gross reports rumors that OpenAI and Anthropic may also have solutions to Hodge and possibly Birch–Swinnerton-Dyer. Noam Brown’s curve: o3 cost about $500K to reach 87.5% on ARC-AGI-1, while Astra scores higher for $20 — a 25,000× collapse in 17 months, implying Millennium-Prize-level reasoning at coffee-cup cost by late 2027.

A Dwarkesh Patel/Jerry Han experiment found better data drove 12× compute-efficiency gains versus 3.7× from architectures — data won by more than 3-to-1, and data pipelines are the moat because architectures get copied in months. Salim’s case study valued one company’s data subsidiary at $32B against roughly $8B for the parent. But Alex’s BloombergGPT cautionary tale stands: proprietary-data advantage “has a shelf life,” possibly only months.

GPUs have flipped from fast-depreciating assets to appreciating commodities: H100 rentals rose 22% in a month to $3.28/hour for a three-year-old chip. Jensen’s words: “fungible, durable, and highly rentable — a productive, revenue-generating asset.” HBM bought a year ago is up about 5×, AWS is reportedly sold out of GB200 NVL72 capacity for years, and corporations without a compute plan in the next couple of months could be frozen out. A panelist frames FLOPs, tokens, and outcomes as “the oil of the singularity.”

DeepSeek’s Flash architecture threatens the AI buildout’s HBM bottleneck. The discussion says its Engram-style lookups route KV-cache work onto SSDs and DDR, potentially reducing HBM requirements about 4×, against Peter’s estimate that HBM represents 40% of America’s current trillion-dollar capex buildout. A $10M-to-train, 500GB Flash model beats Opus and GPT Soul on the cited benchmarks and beats Fable on design at 20× lower cost. Emad’s message to Moderna and similar laggards: catch the frontier on open source now; “if you wait six months, forget it.”

Anthropic’s economic report projects 15% GDP growth, labor share falling from 60% to 45%, and one in five cognitive workers unemployed by 2030 — and the panel thinks it is a lowball. Alex expects “2× or 3× year-over-year growth” if measured properly. Salim says the model breaks because returns flow to capital while aggregate wages remain constant, which he calls mathematically impossible. The panel discusses COVID-scale redistribution, a proposed $5,000 universal basic dividend, and Peter’s $3,000-per-month universal-high-income idea. Peter and Alex both bet Anthropic’s post-IPO wealth will favor accelerationism over effective altruism.

Model weights are becoming national-security assets: Anthropic withheld its latest frontier model from Britain’s AI Security Institute, reportedly the first such withholding by a major lab. Matt Clifford has joined Anthropic, Rishi Sunak is an adviser, and the UK still does not receive the weights. Alex reads this as the beginning of a “Pax Intelligentia”: sovereign inference abroad, with training and safety review domesticated.

The health takeaways are concrete: Insilico’s rentosertib is the first AI-designed longevity drug to reach phase 3, and six aging clocks show patients’ blood-protein signatures looking 3–6 years biologically younger. Alex’s read: four weeks of input for 3–4 years of clock reversal “is, on the margin, longevity escape velocity,” already here “but spiky.” Google DeepMind’s AlphaGenome precomputes the functional impact of roughly 9 billion possible single-letter genome variants, showing the repeatable formula of bulk-solving finite fields into databases.

🔗 Original source & video: Three Lab Warnings in Five Days, Researcher Flags “Gambling with Our Lives,” and Labs Race