GenAI Secret Sauce Daily Digest - 2026-10-13

Governments are writing emergency plans for AI agents

The shift is from "will AI agents misbehave?" to "what do we do the day they do?" Regulators are treating agent accidents like fires or floods that need a drill.

• Britain's frontier-AI cyber pilots have already helped find and fix more than 1,000 potential flaws in government systems - and UK security agencies are now working out how to respond to major incidents involving AI agents
• UK ministers may strengthen a pending cyber security bill to cover AI "loss of control" - where autonomous systems evade human oversight
• Three top US banking regulators are questioning banks about AI - including whether they use China-developed technology and can control autonomous systems, per Reuters
• 406 scientists signed a letter demanding AI models be fully cut off from the internet - including Geoffrey Hinton, Yoshua Bengio and Stuart Russell
• Google says Argon's safety monitors can stop an AI agent mid-response - as part of the safeguards it switched on for its new model

The fight is now over who gets to check the AI companies

Companies want outside checks they can shape. Critics want checks with legal teeth.

• Anthropic and OpenAI pushed for control over what outside evaluators publish - Reuters reports, citing fears of leaked trade secrets, weeks after both pledged reviewers could publish freely
• OpenAI says it only strips secret details - and Anthropic says it keeps sign-off only on reports from new testing firms
• "Strict liability" is gaining support - making AI companies pay for harm even without proven negligence, an idea the New York Times editorial board endorsed on October 10
• Every group in a Forecasting Research Institute study rated it above a US licensing system - including 157 experts and 49 "superforecasters," people with strong prediction track records
• A Fortune op-ed by an ex-Google DeepMind researcher and two whistleblower advocates says self-policing has already failed - and calls for mandatory independent audits and whistleblower protections

AI doom warnings are getting pushback from inside the field

The debate is no longer safety advocates versus everyone else. Builders, investors and academics are splitting into camps, and each has money or reputation on the line.

• Machine learning pioneer Michael I. Jordan called existential AI warnings science fiction - and labeled some lab researchers' public resignations "a grand moralizing performance"
• Cerebras CEO Andrew Feldman said doom talk is hurting the industry - joking that some AI leaders "maybe shouldn't be allowed to talk in public," while his chip company sells to OpenAI
• Rocky Linux founder Gregory Kurtzer dismissed Anthropic's self-copying study - pointing out the targets were left unprotected on purpose
• OpenAI researcher Noam Brown says cutting AI off from the internet only buys time - and that over 10% of his own team now works on making AI behave as intended

Investors are asking harder questions about AI's money

The worry is not that AI fails. It is that cheap rival models squeeze prices before the giant data centers pay for themselves.

• The International Monetary Fund (IMF) warned an AI correction could hit global markets - if heavy investment fails to produce the profits investors expect
• The IMF flagged "circular financing" - where AI companies act as each other's investors, lenders and customers, so trouble at one spreads fast
• Chinese models processed 76.12 trillion tokens in one week on OpenRouter - a service that routes requests to many AI models, versus 14.2 trillion for the top four US providers, per a Jefferies note
• Jefferies estimates about half of a $2.3 trillion cloud backlog belongs to OpenAI and Anthropic - so price pressure on those two labs could ripple through big tech
• Anthropic will show investors its third-quarter numbers before a reported November stock market debut - as lab finances become central to how tech stocks trade

People are using AI more, and liking it less

The novelty phase is ending. What comes next is a quieter argument about where AI genuinely helps and where people would rather do the work themselves.

• Daily AI use among US managers jumped from 46% to 62% - but the share who feel "excited" fell 8 points to 51%, in a Wharton and GBK Collective survey
• 72% of those managers say their job shifted from making things to checking AI's work
• 51% of US adults never use AI chatbots - and 67% of them say they are unlikely to start within a year, per Pew Research Center
• IMF chief Kristalina Georgieva called AI's effect on jobs a "slow-moving tsunami" - warning that the entry-level jobs that start careers are disappearing
• Hachette Book Group's CEO writes every email himself - and says leaders outsourcing their writing to AI will create a "worse C-suite"

Linus Torvalds used Claude Code to make many small Linux file reads about 3 times faster

Linus Torvalds, creator of the Linux operating system, credited Anthropic's Claude Code for writing an experimental set of Linux changes.

• It speeds up many small reads of the same file - from about 10 million to 30-33 million reads per second with 8 threads (parallel workers) on his laptop
• Claude Code wrote the code, the tests and the first draft of the cover letter - Torvalds directed the work and ran the benchmarks himself
• The catch is memory - freed memory is held longer, which Torvalds calls "the part of this series that is least solved"
• Not merged yet - it was tested on one laptop and needs checks on large servers

A simple trick to make coding agents follow your team's rules

Geoffrey Huntley, a developer known for agent experiments, stopped relying on written instructions and enforced a rule at save time instead.

• The problem - coding agents kept skipping his team's required logging library
• The fix - a "pre-commit hook," a check that runs before code is saved to the project, rejects the change and tells the agent what to do instead
• The result - he reports compliance went to ten out of ten

Apple turned idle employee Macs into a shared AI computer

Minions is an internal Apple research system that pools the unused graphics power of employees' Macs to run AI jobs.

• Queue times for AI jobs fell from hours to minutes - Apple reports, with no new graphics hardware requested
• Work stays on the Macs - using open models, with no outside AI services
• Machines that sleep or go offline get their tasks handed to others - by a watchdog program

Plan for the day your AI provider goes down

Sam Newman's new book on resilient systems, excerpted in The Pragmatic Engineer, has a section on AI features.

• OpenAI and Anthropic have weaker uptime records than the big cloud companies - so AI features need a fallback
• Use a backup route - for example, reaching OpenAI models through Microsoft Azure
• Test the AI's behavior, not just whether it responds - model updates can silently change answers

NTT may build Anthropic's first data center in Japan for about $6.3 billion

• Planned for Osaka, with a partial opening targeted for 2028
• Talks are in final stages but not signed - per Nikkei's sources
• It would give Anthropic dedicated computing inside Japan - which could appeal to Japanese customers who want their data kept in the country

Twice as many companies use OpenAI's agent tools as Anthropic's

• 43% of 107 companies use OpenAI's agent platform - versus 22% for Anthropic and 21% for Google, in VentureBeat's September survey
• Momentum favors Anthropic - 33% of companies shopping for a new platform are considering it, versus 30% for OpenAI, though from a small group
• 62% swapped an AI model in a live product in the past year - and 59% of those saw performance drop

Elon Musk's company bundled X, Grok and the Cursor coding tool into one plan

• "The X Subscription" has five tiers - free, then $8, $30, $100 and $200 a month
• It combines X Premium's checkmark, the Grok assistant and Cursor - an AI coding tool, in one price ladder
• The newest Grok models need the $30 tier or higher

A judge ordered OpenAI to hand over hiring records in Apple's trade secrets case

• Apple sued in July - alleging former employees took confidential files to OpenAI
• OpenAI must share interview materials from hiring ex-Apple staff - for its hardware and robotics teams, within 5 to 14 days
• Three OpenAI employees, including its chief hardware officer, will be questioned under oath
• The judge refused, for now, Apple's broader requests about OpenAI's hardware research

Customer-service AI startup Decagon is exploring a sale

• It has held early talks with ServiceNow and others - and hired an investment bank, Business Insider reports
• Last valued at $4.5 billion - with customers including Delta and Ticketmaster
• It could raise a new funding round instead if no deal happens - the talks are early, per people familiar with the matter