I have not been able to shake last Tuesday night which I spent inside Jacob Coxon’s resignation thread, and I told myself I was done after I wrote about it, and then I kept refreshing because a hundred and seventy million views does strange things to a story. It stopped feeling like a resignation letter and started feeling like a market event, with Nasdaq futures wobbling and cable news booking the same twenty seven year old who spent three years on pretraining at OpenAI and a few months at Anthropic to say slower on camera. And I sat with that unease for two weeks until it found a sharper shape.

The shape came from a question which keeps looping in my head about what we would have done in 1946 if Robert Oppenheimer had walked into Sequoia Capital with a deck for fission as a service, and I keep picturing him offering enriched uranium to the UAE on one slide and to Procter and Gamble on the next for dish soap research, and the image feels absurd because we decided some powers were too large to be products, and we built fences around them which were clumsy and political and human, and those fences held enough that only nine countries hold nuclear arsenals today while thirty one run nuclear power plants. And that split between weapon and wattage came from adversaries cooperating through terror into something like restraint.

I wrote last month that models are limbs which borrow all their direction from the human who prompts them, and wanting plus access plus money plus willingness to break seals all live on our side of the glass. David Bellamy steadied me then with steel and supply chains and protocols and freezers and mice which translate poorly, and I still believe him about physics pushing back, and what has shifted for me since is where the money sits while physics pushes back, because a handful of labs capture staggering upside from scale while the downside leaks outward into power bills and water tables and breached inboxes and ballots about data centers, with sixteen local measures across seven states this year alone from people who never asked to host a furnace for someone else’s demo.

Geoffrey Hinton has been living inside this fork for years, and he quit Google to speak freely and then won the Nobel in physics and described two kinds of risk in the same breath, with short term misuse which is urgent and already here through deepfakes and bias in hiring and care and criminal justice and cybercrime with a new tool, and longer term displacement where minds we built outthink us, and he said anyone certain it will all be fine is crazy and anyone certain takeover is inevitable is crazy too, and I trust that uncertainty more than the confident threads (I am no expert on alignment so take my reading with a bag of salt), because Evan Hubinger putting extinction above ten percent inside a decade and Coxon demanding a slowdown and Dario Amodei writing about pacing the frontier all live in the same fog where nobody fully comprehends or controls what scaling does next.

Most times I notice the fog turning populist fast, and I feel it in the way people talk about AI now, which moved from taking your job to drinking your river to ending everyone, and pollsters call it manufactured by out of touch billionaires for capitalist efficiency and population management, and Ted Cruz says he would rather have American killer robots than Chinese ones, and Bernie Sanders proposes banning superintelligence with twenty year sentences, and Sanders sharing a stage with Steve Bannon at a Pro-Human Assembly tells me the old left right map is melting, and Sam Altman comparing training runs to raising children and Elon Musk cutting aid to the most vulnerable while warning about doom land for me as the same move, which is vesting into billions and then asking the public to be grateful and terrified at once. Robber barons left libraries and concert halls while this cohort risks leaving server farms and bot infestations and S-1 filings which list extinction as a risk factor on the way to ringing the bell. Its so sad to see.

I do not want to throw the good out with that anger because the good is embarrassingly concrete, and I spent an evening with a podcast where Josh Tyrangiel walked through Cleveland Clinic work on sepsis, which masks itself as almost anything else and kills about three hundred and fifty thousand Americans a year in hospitals, and a model built to catch it early cut sepsis mortality there by forty one percent, which scales to more than a hundred and forty thousand lives a year if it travels, and the same pattern shows up in AI designed antibiotics for drug resistant bugs and in pancreatic cancer detection pushed up to three years earlier for a disease which kills fifty two thousand Americans a year with eighty five percent diagnosed after spread and five year survival stuck below fifteen percent, and in wildfire spotting and hurricane forecasting and unreadable Vesuvius scrolls made readable and climate tools which researchers feared losing if we halted generative work in 2023. And each of those feels like wattage rather than weapon, which is why a blunt ban or a vague slow the frontier slogan worries me for what it breaks quietly.

The breakage I can already count lives in security, and IBM counted one in four breaches this year as AI enabled with a fifty six percent jump year over year, and Alex Stamos described every hacking team getting promoted a division the way Premier League clubs move up, so small states now swing with capabilities which once belonged only to intelligence agencies, and Sophos puts the average ransomware recovery without ransom at one point seven million dollars and up eleven percent, and aggregate cybercrime cost ten point five trillion dollars last year, which would make it the third largest economy on earth if it were a country, and the entry ticket for mass harm has collapsed from a nation state budget to a laptop plus a grudge. And that democratization runs on the same weights and APIs as the hospital model which saves lives, which leaves me circling who pays for hardening and who profits from scale.

Silicon Valley keeps telling me regulation opposes shareholder value, and history keeps answering with stamps and charters which created value by creating trust, because Upton Sinclair’s Jungle led to a USDA stamp which let people trust meat again, and the 1929 crash led to SEC and FDIC rules which let people trust banks and stocks again. AI valuations today ride on wild speculation fragile enough that the Mag 7 holding a third of the S&P and AI driving thirty two percent of earnings growth this year means a sneeze in one cluster becomes walking pneumonia for retirement accounts. In plain English; a bad week for a few data centers becomes a bad week for everyone’s retirement. And Mark Zuckerberg saying labs have the responsibility and incentive to move safely lands flat after social media, and Donald Trump saying a smart president is the only regulator AI needs lands flatter, and employee like access for evaluators plus promises to welcome a federal framework feel performative in the Sheryl Sandberg sense, which is why the idea of an international agency modeled on atomic energy with a domestic Department of Energy style body recording incidents and approving models and writing safety rules with subpoena power, built with private expertise under public authority and mirrored bilaterally with China where AI is already heavily regulated, feels less utopian to me than letting quarterly earnings set the physics.

A crisis is a terrible thing to waste, and people are terrified again, and every apocalypse gets a villain, and ours wants to sound the alarm while ringing the Nasdaq bell, and history disagrees with that choreography. And I keep returning to hands, which I wrote about last time, because hands sign term sheets and export licenses and deployment memos and pardons for pollution, and the same hands swab sepsis patients and label scans and fight fires spotted early. The models wait on servers for our next sentence, and the waiting is still the whole truth about them, and what we do with the waiting will decide whether the light we made warms rooms or burns them.

Tweet to me which pair of hands you are watching most closely right now. I am @troysk704.