AIron

One engineer had Claude Fable 5.2 run 1,000+ AI agents for 8 days. They built a prototype operating system that boots on a laptop. That’s one experiment from Anthropic’s launch post yesterday. A few more: • 1,000 Fable agents formalized Fermat’s Last Theorem in 30 hours. The first time took 11 days • It helped move a production service from Python to Rust. It now starts about 26x faster • It solved two open problems in theoretical physics, both about black holes • Over 90% of the code merged into Anthropic’s core codebase is now written by Claude, with human guidance Price: $8 / $40 per million tokens. Available now. Which of these would you actually trust it with?

What does a pangolin look like with 1,234 moving parts inside? This. White armor on the outside: 427 plates, each one hinged. Under it: gears, pistons, copper motors, cables and green LEDs. 10 seconds, one take: scales up, teardown, snap back, armored ball. The original series had a tarantula, a piranha and an octopus. I picked the one animal that was basically built for this.

Claude + Blender. No assets. No Tripo. No modeling by hand. One Python script built all of it: the cloth sim, the hat, the lantern, the rune, the fog, the camera moves. Even the sound. 12 seconds. One chat. What should he cast next?

Say your AI agent sends $2,400 to a new payee. Under Chase's new terms, you authorized it. Under Wells Fargo's, you stay responsible for everything it does until you tell the bank you revoked it. The Lockout, Part 2: the fine print. Would you sign that?

OpenAI rates its new model "Critical" for cyber risk. That's the top tier in its own safety framework. It shipped it anyway, to a short list of vetted security teams. GPT-6 Cyber went live yesterday through the API and Codex for approved "Daybreak Red" customers: red teams and pentesters. What OpenAI published alongside it: • one automated run produced 948 working Linux root exploits • 588 of the bugs behind them were previously unknown • the API bill for that run: $1,213,430, about $1,279 per exploit • 81 new high-severity Firefox flaws, now being fixed by Mozilla In OpenAI's scope test, the model stayed inside its authorized targets 99.93% of the time. OpenAI didn't say what happened in the other 0.07%. Same week, OpenAI raised its own bug bounty. A critical bug in ChatGPT or Codex now pays up to $30,000. Finding a root bug now costs about as much as a laptop. Defenders get that price first. The real question is how long they keep it. Who wins this race: security teams or attackers?

AI data centers were on the ballot in 41 US towns this week. They lost 40. Voters in 10 states backed bans, moratoriums or new limits almost everywhere, per Ballotpedia's count. Ohio alone had 21 measures. The one exception was Harwood, North Dakota. NBC News exit poll: about 76% of voters said they'd oppose a new data center in their own community. 56% called AI and data centers a crucial or major factor in their vote for Congress. Two days later, OpenAI published a "National Community Compact." The pledges: • cover the full cost of power and grid upgrades at its future campuses • publish expected water demand for local residents • stop asking mayors and city councils to sign NDAs before reviewing projects For two years the AI race was about chips and power plants. This week it reached the town hall. Would you vote yes on a data center in your town if it paid its own power bill?

Anthropic lost $80.7 billion in six months. The business behind Claude still brought in more cash than it spent. Both are true. Here's how. Anthropic filed to go public. The S-1 shows the first half of 2026: Revenue: $16.3B Operating loss: $693M Net loss: $80.7B Running the company came close to break-even. Then one line added $72.4B: the "change in fair value" of convertible notes. Convertible notes are IOUs that turn into shares later. As Anthropic got more valuable, those notes got more valuable too, and accounting records that rise as a loss. No cash left the company for it. Operating cash flow was about +$3.9B in the half, after burning $1.5B a year earlier. The quieter shift is compute. In 2025 it cost 160% of revenue. In H1 2026, 55%. Now the part most people will skip. The loss is mostly paper. The commitments are real. The amended S-1 lists about $895B in purchase and lease commitments, almost all compute and infrastructure, running as far as 2045. That's roughly 27x Anthropic's revenue for the first nine months of 2026. The income statement tells you how AI labs look today. The contracts tell you what they've promised to pay for tomorrow. Is $895B of compute a moat or a trap?

This AI kept thinking in circles. GPT-6 Astra changed 0.016% of its weights. Its math score went from 7.9% to 48.1%. The detail most people skip: 93% of its answers never reached the end. The edit mostly fixed that. Full breakdown

AI agents just got a kill switch. Is that enough? On Nov 3, CrowdStrike launched an "AI Kill Switch" in Falcon Guardian. What it does today: an authorized admin can block a misbehaving agent's access to the AI models that power its decisions, through the AI gateway. Every action is recorded in an audit trail. What's "coming soon": stopping agents directly on the device, plus cloud, identity, SaaS and browser controls. CEO George Kurtz: "When you deploy agents with a kill switch, you stay in control." My take: a stop button is the last line of defense. The first is giving agents only the access they actually need. If an agent can move money or touch customer data, is a human-pressed kill switch enough? Or do you want hard limits before it acts?

My friend swears he wrote this apology to his girlfriend himself. Read the first letters.