Today’s narrative
Today's most consequential document is a pair of papers from the UK AI Security Institute and the US biosecurity nonprofit SecureBio, testing the two safeguards governments and labs lean on most to keep AI from helping someone build a bioweapon: withholding code, and filtering dangerous sequences out of training data. Both held up in a narrow technical sense and both turned out to be cheap to defeat. A frontier coding agent rebuilt an open protein-design tool from its published paper alone, at 96.5% fidelity, for under $20 in compute. Fine-tuning a filtered protein-language model on the exact viral data that had been withheld restored full capability for about $80. Neither paper claims the sky is falling — the hardest tool tested held up, and both teams flagged real limits in their own methods — but the clear verdict across both is that these safeguards degrade a capability rather than remove it, and today's agents can already claw back a meaningful share of it for a few hundred dollars or less.
A second throughline is about who the market is rewarding for AI capital spending, and who it's punishing. Amazon's October 29 earnings showed capex more than doubling to $75.9 billion net in a single quarter, flipping trailing free cash flow to a $42.4 billion deficit — even as AWS grew 39% and CEO Andy Jassy said capacity remains largely sold out through 2027. Microsoft, reporting the same week, posted free cash flow of $18.63 billion, comfortably positive, and watched its stock rise rather than fall, in part because roughly three-quarters of its capex now goes to short-lived chips rather than long-duration buildings. Alphabet and Meta, by contrast, both saw their stocks and cash positions take a hit from the same AI buildout this earnings season. The same week, Amazon's own engineering organization was dealing with a smaller-scale version of the measurement problem: a VP memo killed an internal leaderboard that had rewarded raw AI token usage, arriving within 48 hours of Anthropic shipping new tools to show enterprise customers what their AI coding spend is actually buying, and of analytics firm Jellyfish reporting that AI coding spend per engineer has roughly tripled in six months with little change in what engineers produce.
A third pairing is about labs making claims about themselves that are hard for anyone else to check. Anthropic published guidance telling companies to fold multi-agent architectures back into a single agent loop, backed by a Claude Code system-prompt reduction and one unnamed customer's results — but the only real-model testing behind its own reference implementation, as published, was two runs on 12 tasks by one person. Separately, a report from The Information says Moonshot and DeepSeek trained their newest models largely on output distilled from OpenAI's GPT-6 Astra, extending a pattern of similar accusations OpenAI and Anthropic have each leveled at Chinese labs since February — unverified by any outside party, arriving the same week Moonshot's Kimi K3.1 Pro slid on a newly hardened benchmark it still leads among open-weight models.
Elsewhere: the Dutch cabinet told parliament it plans to cut 25,000 central-government jobs by 2035 with AI as a stated tool, but its actual AI deployment plan isn't due until the first quarter of 2027 — meaning the cuts start well before the government has committed to saying what the AI is for. And the US Treasury proposed a rule that would scrap compute-based thresholds for screening outbound AI investment into China in favor of a blanket prohibition, a change lawyers say could sweep in ordinary software companies far removed from frontier labs, with comments due December 7.
Top stories
Signal movers
You’re receiving this because you subscribed at machine.net. Unsubscribe · Manage preferences