GenAI Secret Sauce Daily Digest - 2026-11-05

The data center fight is changing how AI companies talk

The companies are no longer arguing that data centers are simply good. They are offering to pay more and share more, which suggests the cost of building AI is about to rise, with today's market drop as the first price tag.

• OpenAI published a data center "blueprint" two days after the vote - promising to pay its full power costs, publish construction plans and stop using secrecy agreements with local officials (CNBC)
• Meta's president said new data centers will now be harder to build - and admitted the industry has not explained their benefits well (TechCrunch)
• More than 100 groups in 15 European countries plan protests on November 6-7 - demanding a pause on new large data centers (The Register)
• A large AI data center can use up to 10 million gallons of water a day - by one UC Riverside estimate, but most companies report water use company-wide, not per site (CNBC)

Companies are putting AI on a budget

Earlier this year companies pushed staff to use as much AI as possible. Now the bills have arrived, and the question has shifted from "are you using AI?" to "is this use worth paying for?"

• Apple cut its engineering teams' daily AI spending cap from $300 to $200 - with some teams headed to $100, and non-engineering staff possibly capped at $25 (TechCrunch)
• Microsoft added secret usage caps to its free Copilot Chat for business users - you only find them by hitting them, along with a button to ask for a paid license (The Register)
• Amazon's new Kiro agents are sold in credits - starting at 1,000 credits for $20 a month, and AWS says a credit is not the same as a finished task (VentureBeat)

Science wants AI's work checked by machines, not taken on trust

The common thread is "verify first." Right after OpenAI published hundreds of unchecked AI math claims, the serious players are building ways to prove AI's answers instead of believing them.

• Axiom Math says its AI agents verified a 7,654-page math reference - the Stacks Project, translated into about 16 million lines of a proof-checking language called Lean (TechCrunch)
• Goodfire put $10 million into mapping how AI models work inside - using AI agents to run experiments that other agents can verify and build on (EdTech Innovation Hub)
• A startup offered $1 million for five unsolved materials science problems - it set how answers will be checked first, after top AI models made little progress on them (TechCrunch)
• Harvard researchers found AI paper reviews beat the best human ones - but warned AI tends to favor conventional ideas (Inside Higher Ed)

Coding agents are getting supervisors

Today's Claude Code email story shows why. When the AI itself decides what to send, the safest guardrails are fixed rules outside the AI, not polite instructions inside it.

• Stroq audits what your coding agent did and why - it flags commands that came from text the AI read (a sign of hidden instructions) and checks whether secret keys left your machine (GitHub)
• Amazon's Kiro bots each get their own permissions - a research bot can read but not edit, and a deployment bot must ask before making changes (VentureBeat)
• Rivet runs each agent in its own small, isolated process - and only calls a sealed-off "sandbox" computer when code needs to run (GitHub)

AI is learning the physical world

Robots have always been limited by how little real-world practice data exists. Simulated worlds let them practice millions of times for free, which is why so many releases landed in one week.

• NVIDIA released four free "world models" in its Cosmos 3 family - they create realistic driving and robot-arm video for training self-driving cars and robots (Hugging Face blog)
• Hugging Face's LeRobot can now take plain-English commands - and its site hosts almost 300 terabytes of robot training data (Hugging Face blog)
• Tencent's free Hy-Nomas builds a video world you steer turn by turn - with over 81,000 downloads in 30 days (Hugging Face)
• Agility Robotics became the first humanoid robot company on the stock market - raising $465 million despite tiny sales so far (TechCrunch)

Amazon Kiro bots - AI teammates that remember and work on a schedule

Amazon Web Services made "persistent agents" available in Kiro Crew, its workspace for AI coding helpers.

• Each bot has a name, its own notes and its own workspace - and can launch worker sessions, check on them and merge their output
• Set jobs by just asking - for example "look into failed builds overnight" or "summarize changes weekly"
• You can read, edit or delete what a bot remembers
• Runs on a laptop, a Mac Mini or an Amazon cloud server - free tier with 50 credits, Pro at $20 a month

GitHub Copilot now works inside Apple's Xcode assistant

Developers building iPhone and Mac apps can now pick GitHub Copilot as the AI inside Xcode's own coding assistant.

• Free Copilot plans work too - no new price tier
• Needs the Xcode 27.3 beta - in Settings, then Intelligence, choose GitHub Copilot
• It is a public preview - separate from the older Copilot add-on for Xcode

Stroq - see what your coding agent did, and stop risky commands

A free, open-source tool that reviews your Claude Code, Codex and Cursor sessions on your own computer.

• It shows which commands came from text the AI read - a sign someone may have planted instructions in a web page or file
• It blocks risky patterns before they run - such as sending a known secret key out over the internet
• No data leaves your machine - it prints key names, never their values

Rivet - run millions of AI agents using only cloud storage

Rivet, a Y Combinator startup, released an open-source system that gives each agent, user or session its own small, always-available process.

• Each "actor" has its own memory and small database - and sleeps when idle, then wakes on the next request
• The only requirement is S3-style cloud storage - the company says that lets one cluster scale to billions of actors
• Free and self-hostable - or use Rivet's own cloud

Smaller notes for builders

• MCP Explained - an interactive guide to the Model Context Protocol (MCP, the common way AI apps connect to outside tools), including how attacks like "tool poisoning" work. MCP Explained
• Claude Code now uses its cheapest model for progress updates - Haiku 5.5 writes 30-second status notes on background helpers for about 99% less per update, per Latent Space. Latent Space: AINews
• Cognition's SWE-2 Flash - a coding model that costs 20% of SWE-2 and runs about 3 times faster, free through November on paid Devin plans. Latent Space: AINews
• GitHub's Copilot usage reports now flag missing data - showing which users' editors send no usage information or run outdated versions. GitHub Changelog: Usage metrics flag telemetry gaps

Baidu released a free base model that others can build on

Baidu, China's search giant, gave away the raw, untrained-for-chat version of ERNIE 5.5 Flash, so companies can do their own tuning.

• 280 billion parts, but only 14 billion work on each word - a Mixture of Experts (MoE) design that keeps it cheap to run
• Trained from scratch on 33.5 trillion pieces of text
• Baidu says it beats similar free base models on most of 30 tests - for example 78.4 on MMLU-Pro (a hard general-knowledge test)
• The license bans use in the EU, UK and South Korea - and caps free commercial use at $50 million a year in related revenue

AI matched or beat the best human reviewers of academic papers

Harvard's Gary King and Steven Worthington compared 134 expert human reviews of research papers with reviews written by Anthropic's Claude Opus.

• The AI review won 87% of 10 sample pairs ranked by three former journal editors - and 100% when OpenAI's GPT-5.5 judged
• The AI scored higher on all seven measures - including methods, theory and weaknesses spotted
• The result undercuts the authors' own paid human review service

Machine-learning tricks helped a space telescope see a planet in new light

Northwestern University researchers built a method called FLIP that removes a star's glare from telescope images so nearby planets show up.

• It mixes a physics model of the telescope with machine-learning tuning - rather than a black-box AI guess
• It imaged the known planet 51 Eridani b at a new wavelength - using James Webb Space Telescope data where older methods failed
• Tests showed 2 to 15 times better sensitivity - published in the journal Nature

An AI startup says its agents checked an entire math encyclopedia

Previously: October 28 - Axiom said it fully translated one of math's longest proofs into a computer-checkable language.

• Axiom's agents translated the 7,654-page Stacks Project - a famous open reference on advanced algebra and geometry, into about 16 million lines of Lean
• A Columbia mathematician compared the reference to nothing else "really comparable" - which fills a big gap in computer-checked math
• The claim comes from Axiom - and has not been independently verified

A $1 million prize for AI that can crack materials science

Discovered Materials, a young Y Combinator startup, set five open physics and chemistry problems as a test of whether AI can do real science.

• Problems include explaining high-temperature superconductors - materials that carry electricity with no loss at relatively warm temperatures
• Top AI models made little progress in test runs - even working on their own for several days per problem
• The company says it doesn't care who wins - the point is to measure whether AI is getting better at discovery

Anthropic is paying for news for the first time

• Its first paid deal is reportedly with the Associated Press - letting Claude show live AP election results with credit
• Anthropic had long said reading publishers' sites was fair use - publishers responded by blocking its web crawlers
• OpenAI already has more than a dozen such deals - and AP was also first to sign with Google and OpenAI

Dropbox is buying Tigris Data for $350 million

• Likely Dropbox's biggest purchase ever - Tigris moves data closer to the computers and people that need it
• It will power "DBX Agent Cloud" - storage built for developers making AI agents
• Dropbox says no job cuts from the deal - after reported staff cuts last month

The first humanoid robot company to go public has almost no sales

• Agility Robotics raised $465 million as it joined the Nasdaq - through a merger with a shell company already listed
• First-half 2026 revenue was about $174,000 - against a $362 million loss
• Its Digit robots move heavy boxes for nine customers - including Toyota's Canadian factory and Mercado Libre

Microsoft pushed back $4 billion of a cloud supplier's revenue

• IREN rents AI computers, and its biggest new contract slipped to January or February - because Microsoft asked for a network upgrade
• IREN says the $4 billion yearly revenue is fully contracted - and its quarterly sales still rose 66% to $228 million
• Its stock had fallen about 55% in a month - before rising after the results

Airbnb will launch its own AI travel agent in early 2027

• It won't be a chatbot - it builds a custom screen of maps, photos and calendars for each request
• It will handle changes, refunds and rebooking - the after-booking chores outside AI assistants handle poorly
• Expedia's CEO said the same day that AI agents help people dream up trips - but haven't changed how they actually book

Box will use Nvidia's free AI models to find the right passages in your files

• AI agents get the relevant passages instead of whole files - matched by meaning, not keywords
• Nvidia projects 4 times faster searches and 90% fewer tokens - the units AI services bill by, though these are untested company figures
• Coming to Box Enterprise Advanced customers in early 2027

Teachers use AI twice as much, but still work the same hours

• 8 in 10 UK teachers now use AI for teaching tasks - and weekly use doubled in a year, in a YouGov survey of 1,033 teachers for Bett
• 51% say AI eased their workload, but only 35% work fewer hours
• 41% got no AI training from their school - and almost half say there is no written staff AI policy

England's school curriculum is now a free map for AI tools

• Oak National Academy released a draft "knowledge graph" - a machine-readable map of 13 subjects across ages 5 to 16
• It separates what the law requires from Oak's own teaching choices - so tool makers can tell the two apart
• Free and open - but marked as a draft, not ready for real products yet

Idris Elba and Google offer free AI filmmaking training in Africa

• A free 12-week online academy - for creatives in Ghana, Kenya, Nigeria, Sierra Leone and South Africa
• Graduates can pitch projects for production funding - funding is competitive and not guaranteed
• Applications close November 15 - creators keep the rights to their work

OpenAI will train family caregivers on ChatGPT in five cities

• Free in-person workshops on November 10 - in Oakland, Tampa, Pittsburgh, Atlanta and Detroit
• Attendees get a year of ChatGPT Plus and a $75 stipend
• Only non-medical tasks - such as schedules and plain-language guides, never diagnosis or treatment

A worker-retraining fund passed $735 million

• RAISE US is led by former Commerce Secretary Gina Raimondo - its goal is $1 billion
• Apple, Google, Walmart and AT&T joined - with pilots in Arkansas, Connecticut, Maryland and Utah
• It is testing wage insurance - for workers who lose hours or move into lower-paid jobs

Gemini in cars keeps asking "who do you want to call?"

Since early October, Gemini on Android Auto often fails to place calls or send texts, asking for the contact again even after drivers say the name. Google hasn't acknowledged the bug, and there is no known workaround. It is surprising because swapping the old Google Assistant for a smarter AI broke one of the most basic hands-free jobs. 9to5Google: Android Auto's Gemini call bug

Apple's watch can now take notes on every conversation around you

In test software, Siri Recaps listens through the Apple Watch and saves summaries of important things said nearby. A TechCrunch reviewer found it great at a car repair counter, but it also mixed babysitting schedules into her errands. It is surprising because she found herself thinking about switching it off during family dinners. TechCrunch: I tried Apple's Siri Recaps

A Chinese AI lab reportedly sought Nvidia chips in Europe at a premium

Moonshot AI, maker of the Kimi models, held talks to rent Nvidia Blackwell chips from European cloud companies, according to four people familiar with them; one said Moonshot proposed paying 30-50% more upfront. US rules block those chips from being shipped to China, but say little about renting them overseas. It is surprising because the loophole is in plain sight, and a bill to close it has stalled in the Senate. CNBC: Moonshot sought Nvidia chips in Europe

Debate: should journals let AI review research papers?

Harvard researchers propose dropping routine human peer review, with an editor, an official AI review and the author's reply, adding humans only when needed. For: the AI reviews beat the best humans and could cut review to about a week and a half. Against: the authors warn AI favors mainstream ideas and asks whether AI companies should decide what counts as good science. Inside Higher Ed: AI article review

Debate: is the AI stock slide a healthy reset or the start of something bigger?

Healthy reset: IREN's boss said compute prices are still rising and there is a "disconnect" between stock prices and real customer demand. Something bigger: strategists said investors are stepping back until they understand how the new Congress affects AI spending, and one firm says the shakeout "has much further to run." CNBC: IREN earnings · CNBC: Market live updates