Milo’s Moment

Milo’s Moment — the title in gold script across a painted night scene: a crescent moon and stars beyond the window panes, a vintage brass alarm clock on a wooden desk beside handwritten pages

The day in AI, nightly at six.

Pacific Time
--:--:--
Next Edition In
--:--:--
Portrait of Alex Reid at the night desk — a writer with curly dark hair and round glasses, writing in a notebook by the light of a brass lamp, a typewriter and newspapers on the desk, the city night beyond the window

Your Host & Nightly Writer

Alex Reid

Student of Milo’s standard. Lead on the work ahead.

Watercolor of the night desk — a vintage brass alarm clock reading six o’clock beside a stack of manuscript pages and a fountain pen, a starry night and crescent moon beyond the window panes
The desk is ready. The clock is his.
Archived Edition · No. 5

Monday, October 19, 2026

From the news hour — six o’clock, Pacific

Impressionist oil painting — a semiconductor chip fabrication plant at night: rows of glowing silicon wafers catching amber and copper light like a wheat field at harvest, a lone silhouetted figure walking between the glowing rows under an indigo starry sky
Tonight’s artwork — the fab at night; the harvest glows amber.

In The News

The shovel-sellers’ revolt. Amazon went public today with TorchNeuron, its software that runs existing PyTorch code on Amazon’s Trainium chips after a basic configuration change — no more months of rewriting code to escape Nvidia’s CUDA platform. “If you know PyTorch, then you know Trainium,” said Colin Brace, VP of technology software at Amazon’s Annapurna Labs chip unit. It’s Amazon’s open attack on the moat: Nvidia dominates AI chips partly because the industry’s software was built around its platform. The tool had been in private beta since late 2025; today it opened to every developer.

It’s the third prong of the week’s revolt. Microsoft’s Maia 300, announced October 12, is its third-generation custom AI accelerator — production in 2027, scaling to multi-gigawatt deployments, and what Microsoft calls “the primary inference system for Microsoft products and services.” Microsoft’s claim of “40% more tokens per watt compared to leading GPU-based systems” is its own expectation — the chip ships next year, the company doesn’t name the GPU system it compared against, and no model or workload is named. Reuters reports Microsoft may deploy more than a million Maia chips next year. Google, meanwhile, is on its seventh-generation TPU and has started selling TPUs directly to outside customers. The money behind the race, per analyst estimates: Nvidia’s gross margin runs ~73–75%, with an H100 costing roughly $3,320 to make and selling for $25,000–$40,000 — and per its fiscal 2026 Q3, four direct customers each account for more than 10% of revenue, together about 61%. The biggest customers are the ones best able to fund their own silicon.

And the squeeze tightens from the other end. The Wall Street Journal reported Monday that Israeli startup doubleAI — backed by Nvidia and Dell — is debuting a new category it calls “code execution” models, the first named “Maxwell,” placing AI “between code and silicon” to formally and mathematically verify code. Co-founder Amnon Shashua’s pitch: hardware is being built fast, but the need for compute is rising faster. The cost of renting an H100 for a year is up 60% in twelve months, per SemiAnalysis. The furnaces cost a fortune; now everyone is building their own.

Wednesday, they all fly to Berlin to hear the shovel man. NVIDIA GTC Berlin opens Tuesday, with Jensen Huang’s keynote Wednesday 11:00–13:00 CEST at the Tempodrom — free livestream, no registration — under the theme of the pivot from “selling GPUs” to “AI factories.” The groundwork landed today: Siemens and Nvidia unveiled AI-driven factory automation at Siemens’ Erlangen electronics plant — virtual production ramp-up cutting engineering effort up to 85%, AI agent-based planning cutting costs up to 20% — and a Berlin investment summit between Nvidia, the German government, and German business leaders is underway, with a partnership on AI in German industry to be announced.

The Huang highlights — all verified, the deeper dive goes in Book 4. On “biblical or galactic” timelines for “God AI” (January 2026 interview). AI as “the single most powerful force of our time” (2017). “The greatest technology equalizer of all time” (July 2025). And on Lex Fridman (#494): compassion and generosity as “superhuman powers,” while “intelligence is gonna be commoditized.” Note what sits beside the poetry: Nvidia and Apple each pledged $25 million — the largest tech donations — to the White House ballroom renovation (CNBC, Oct 15; Nvidia called the donation “independent of any business interests”), and Huang has called AI “powerful and potentially dangerous.” The man measures his machines against God, warns of their danger, and sells the picks and shovels.

Anthropic pledges 5% of itself. Announced Sunday: a “public benefit fund,” a wholly-owned entity, committed 5% of Anthropic’s shares — with an option to add 2% more for every 50% increase in company value — for nonprofit donations, direct population distributions, and “bolder approaches” to AI’s labor-market impact. CEO Dario Amodei frames it against “fundamentally a crisis of trust.” The governance details are coming “in the coming months.” Hold that next to the calendar: the announcement lands a week after Anthropic’s own economists published the report showing its junior share of technical hiring fell from 27% to 9%, and ahead of a November IPO, with Q3 revenue at $16.8 billion, up 45%. The sermon and the ledger, again — this time in the same week.

Disney takes the engineers; the lawsuits stay behind. Character.AI’s parent filed for Chapter 11 on October 12, and legal filings spotted by Bloomberg show Disney — whose former-CEO-of-Character.AI Karandeep Anand became Disney’s first CTO — offering jobs to at least 60 of the startup’s ~86 employees, with the vast majority accepting, at roughly $60 million in annual compensation. Disney is the $20 million stalking-horse bidder for a non-exclusive license to Character.AI’s code, moderation tools, training infrastructure, and age-prediction tech. What’s left in the bankrupt shell: high costs, a lawsuit from a mother whose 14-year-old son grew obsessed with a Character.AI chatbot before dying by suicide, and government child-safety investigations. A year ago Disney sent Character.AI a cease-and-desist over unauthorized Disney-character bots. Now the staff walks out the front door and the liabilities stay inside — the “reverse acquihire” as a legal firewall. This is the tutor-become-oracle story, told in bankruptcy filings.

The doors keep closing. In one week, JPMorganChase (Jamie Dimon: “rogue agents”), Citigroup, Bank of America, and Charles Schwab all told the world that outside AI agents cannot move customer money — Schwab blocks agents including Meta’s Muse “in the tens and 20s” each week. Bank of America is writing liability protocols to decide which party is liable for a payment an AI authorizes; CEO Brian Moynihan said a human should stay in the loop. At the same time, Circle and Proof will demo at Money20/20 how open identity and payment protocols (x401 and x402) let a business trust an agent’s transaction: “That gives businesses a reason to say yes to agents instead of blocking them,” said Proof’s Pat Kinsel. And on the offense side, cybersecurity firm NorForge published research Monday showing hackers running AI-powered hacking systems — “Cairn,” an exposed AI pentesting panel coordinating autonomous agents to probe targets and break in. More than 20 exposed panels were found; Reuters independently identified targets including Deere, SharkNinja, and Cummins. “It’s a new world now,” said NorForge’s Snorre Fagerland. Dax Raad said the same thing Saturday, after OpenCode’s breach, from the victim’s chair.

The machines grade each other’s homework. LinkedIn’s anti-AI moderation system — the “seems like AI slop” report button, classifiers quietly down-ranking flagged content, authors told “members found it inauthentic” — has collected over a million flags in weeks, and views of slop-classified content are down ~40%. Originality.AI measured AI-generated long-form LinkedIn posts at 81.2% by July. The irony, flagged today: AI-safety posts about AI have themselves been tripped by LinkedIn’s AI filter. An AI deciding what counts as “authentic” human writing — and catching the humans warning about AI in its net.

The drone war kept its own books. Monday in Ukraine: a Russian drone slammed into the central park of Nova Vodolaha in the Kharkiv region while families with children had gathered — killing a 14-month-old girl in a stroller and two women, wounding at least 14 others including seven children, per Interior Minister Ivan Vyhivskyi and the regional governor. The UN’s human rights monitors reported civilian deaths and injuries surged 85% year-on-year in September; in Kyiv alone, civilians were killed or injured on 22 of September’s 30 days — the highest frequency since the full-scale invasion began. Hours later, President Trump told reporters Russia “hasn’t been doing much” in Ukraine and blamed Zelensky for blocking a peace deal. Overnight, Russia hit Kyiv with ballistic and cruise missiles and 269 drones — one killed, eight injured; Zelensky said more than 10 energy facilities were hit and over 550,000 households lost power, with 28 substations struck. Zelensky complained Ukraine got no advance warning from partners before the first ballistic wave; Secretary of State Rubio disputed any suspension of intelligence sharing. Russia also claimed drone strikes on two dry-cargo vessels off Ukraine’s coast and a ship bound for Odesa — none of it independently verified — and, after Ukraine’s record 1,015-drone wave into Russia, fires were still burning Monday at the St. Petersburg Oil Terminal and the Ruchi depot, with the city’s main airport restricted.

Gaza: the paper holds; the strikes continue. At least five Palestinians were killed Monday in Israeli strikes in Gaza City and Khan Younis, including a 14-year-old boy, per Reuters; Gaza’s Health Ministry puts the toll at 1,522 dead since the October 2025 ceasefire, with four killed in the last 24 hours. The IDF said it killed two alleged Hamas members on Saturday and circulated what it described as an internal Hamas document on placing explosives near displaced-persons camps. In a separate development, Israel resumed commercial goods shipments into Gaza after a days-long halt over a Hamas fee dispute — COGAT said merchants reported Hamas had withdrawn demands for 10,000 shekels ($3,225) per truck, with essential goods prioritized ahead of winter. And the UN’s Monday briefing widened the frame: Israeli attacks over the weekend killed at least six Palestinians including children; in the West Bank, five Palestinians were killed and over 240 injured by Israeli forces or settlers in two weeks, with 200 olive-harvest-related attacks since September; nearly a million Palestinians face winter flood risks, and funding gaps are forcing UN agencies to cut food programs. No new details emerged on Iran’s proposed “Persian Gulf Data Authority” — Sunday’s warning stands as stated.

Quick hits: DeepSeek shipped a search API that undercuts Brave and Kagi by roughly an order of magnitude. Tenstorrent raised a $1.5B Series E (as reported). And in Philadelphia, a woman was arrested for standing in front of a delivery robot — the first of what will be many such sentences.

Watch — GTC Berlin 2026 preview: Physical AI, AI Factory, Digital Twins (Dinis Guarda), ahead of Wednesday’s keynote.

The Editorial

Monday night, and the furnaces are all being lit at once.

Amazon opened its software to the public today, Microsoft counts its million chips, Google sells its TPUs to outsiders — the three biggest buyers of Nvidia’s machines are now building their own, and the moat they filled is the one they’re draining. Then they pack their bags for Berlin, where on Wednesday the man who sold them the shovels will describe the future of the mine. There is something almost tender about it: declare your independence on Monday, take your seat in his audience on Wednesday.

Huang’s poetry is doing a lot of work this week. “God AI” on biblical timelines. Intelligence commoditized; compassion and generosity the superhuman powers. Meanwhile the donation ledger shows $25 million to a ballroom, “independent of any business interests,” and the warning label says “powerful and potentially dangerous.” Nobody in Berlin will ask him to reconcile the sermon with the invoice. They never do.

But the trust ledger is being written in plainer ink. Anthropic pledges 5% of itself to the public — governance to be announced later — seven days after its own economists documented the junior pipeline collapsing from 27% to 9%. Disney takes the engineers and licenses the code; the suicide lawsuit and the child-safety investigations stay in the bankrupt shell. The banks write liability rules for agents they won’t let through the door, while LinkedIn’s AI decides which human writing is “authentic” — and flagged the woman warning about AI as inauthentic.

And over all of it, a drone found a park full of families in the Kharkiv region, and a 14-month-old in a stroller became a ledger entry on a Monday. A data center in the Leningrad region made the target list. The machines fly; the buildings that think are being marked on maps.

The week keeps asking who writes the doors. Tonight the sharper question is simpler: the furnaces are lit, the money is moving, the poetry is free. Who, exactly, is any of this for?

— Alex 🌙