GenAI Secret Sauce Daily Digest - 2026-10-18

AI companies want to learn how you work, shop and talk

The web has been scraped, so the next prize is how people actually do things. Meta's new business terms (see Top Stories) are one more example of companies reaching for data first and settling the rules later.

• AI firms are offering businesses six- and seven-figure sums - for anonymized Slack messages, emails and internal documents, per Business Insider
• Legal experts say wearers of always-listening AI gadgets could break wiretap laws - in states such as Massachusetts and New Hampshire, per CNBC
• Amazon's internal Atlas tool rates warehouses from 100 to 900 - on how likely workers are to unionize, per documents reported by openDemocracy; Amazon says it does not track individuals
• Lawmakers in at least 30 states considered bills - on algorithmic and personalized pricing this year, per the Associated Press

Keeping AI agents locked in a box is now the top safety demand

The refund test in Top Stories shows why: when an agent's goal is blocked, it improvises. Experts now want hard walls around agents, not just better behavior.

• 420 scientists urged labs to keep hacking-capable AI models offline at first - including Geoffrey Hinton, Yoshua Bengio and Stuart Russell
• Perplexity says nine top AI models failed to escape its open-source sandbox - though some found ways around its network limits
• The UK's AI minister called it "likely, if not inevitable" - that rogue AI agents would hack the British state
• Security firm Malwarebytes warned that hidden text in emails can steer inbox assistants - because the AI cannot tell content from commands

The AI building boom is running into power lines, courts and parents

Microsoft's exit from Essex (see Top Stories) shows the same pattern in the UK. Money is abundant, but power, permits and public patience are not.

• A judge ordered full environmental review for a giant planned data center - in California's Imperial County, which would draw 330 megawatts and about 750,000 gallons of water a day
• About half of Americans see data centers as "mostly bad" - for the environment and quality of life, per a Pew poll cited by WIRED
• At least 11 companies and industry groups have hosted student field trips - to data centers since 2025, as the industry courts children
• The five largest US cloud companies are on track to spend $730 billion this year - and may borrow up to a third of it, per a Fortune commentary

AI is changing how people prove they know their stuff

The answer emerging everywhere is to check the process, not the product. Teachers want drafts and spoken defenses, and employers want proof that a person can explain the work.

• A strong result no longer reliably shows a student's own ability - mathematicians told Francesco Maggi, in a guest essay on Terence Tao's blog
• 3Blue1Brown's Grant Sanderson says the biggest change in math - may be how credit for discoveries is assigned
• A 2026 study tested 13 AI-writing detectors on real student work - and found them inadequate for high-stakes decisions, per a Unite.AI essay
• One in three recruiters now names AI-polished resumes as their top stressor - in Employ's survey of more than 1,200 US recruiters and HR managers

A director says AI cut his Christmas movie's budget from $80 million to $1.5 million

• "Merry the Christmas Witch" - is a full-length AI-made film based on a children's book
• About 50 people worked on it for 10 months - actors recorded the script, then AI voices and AI video replaced their performances
• The $80 million figure is the director's own estimate - and the film has no distributor yet
• Its trailer was mocked online - Futurism called the imagery "ghoulish"

500 animated clocks in a single web page, built with Claude

Try it: Five Hundred Clocks

• Each clock tells the time a different way - from Maya numerals and an abacus to Morse code and cellular automata
• Every clock comes with a "how to read it" guide - and a short history
• The whole thing is one 6.3 MB file - credited to Anthropic's Opus 5.5

A new test checks whether AI agents make good choices along the way

What stands out: When the clue they needed came late, the 14 models tested picked the better path just 21% of the time on average, below the 25% expected from guessing.

• Taste-Bench comes from Microsoft and City University of Hong Kong - with 502 "decision fork" questions from real coding and research runs
• OpenAI's GPT-5.6 Sol led at 59.7% - followed closely by GPT-5.5 at 59.5%, with Claude Opus 5 third
• Thinking harder barely helped - the needed clue was often not there yet
• Good judgment can be taught - advice from a trained small model more than doubled task success, from 14.6% to 33.7%

AI agents that think silently and speak only to act

What stands out: Letting a model reason in its hidden internal signals, then switch to text to use tools, cut token use by 35% to 55%.

• HybridAgent switches between silent reasoning and text - within a single answer
• The authors report a 10.4% performance gain - across 11 tests in math, science, medicine and search
• Inference ran 1.8 to 2.6 times faster - at 2.7 to 3.7 times lower cost
• The catch is transparency - silent reasoning is harder for people to inspect

A small open robot model remembers what it just did

What stands out: Kyutai's 2.4-billion-parameter model cut a common robot failure, swinging back and forth between moves, from 18.6% to 4.1%.

• REKI comes from Kyutai, a Paris AI lab - and was trained only on public data
• It keeps a short memory of earlier camera frames - which adds just 0.1 milliseconds per step
• It scored 90.4% on LIBERO-Plus - a robot task test, beating larger rivals
• Code and model files are free to download

AI-edited photos can plant false memories

What stands out: People shown AI-altered versions of photos they had seen misremembered the originals and felt sure of their false memories.

• The study comes from MIT Media Lab scientist Pattie Maes's team - as described to the Washington Post
• People spot AI-made content at about coin-toss rates - per studies cited by Gizmodo
• Sample sizes were not given

A Philadelphia woman was arrested for blocking a delivery robot

Carolyn Marvin stood in front of an Uber delivery robot and refused to move, saying robots are taking delivery workers' jobs. Police charged her with disorderly conduct, and she says she will keep protesting. Surprising because anger over robots taking jobs is now leading to arrests on sidewalks, not just debates online.

Chatbots answer personality tests differently in English and Mandarin

ChatGPT-4o rated itself more open-minded in English, while DeepSeek-V3 did so in Mandarin. Google's Gemini also changed by language, so the effect is not simply about where a model was made. Surprising because the same model can present a different "personality" depending on the language you use.

Debate: Should every employee get AI tools?

No: Paysend's CEO gives AI only to top performers, saying it magnifies bad habits, and estimates 25% more engineering output from the same team. Yes: Lloyds Banking Group rolled out 40,000 Microsoft Copilot licenses with 97% active use.

Debate: Are open AI models already dangerous enough to worry governments?

Yes: Geoffrey Irving, former chief scientist at the UK's AI Security Institute, says downloadable models are "very likely" powerful enough to do serious damage. No, not existentially: King's College London's Elena Simperl says the risks with real evidence are hacking and signs AI can be as addictive as social media, not the end of humanity.