GenAI Secret Sauce Daily Digest - 2026-10-23

AI is being moved into locked rooms

A year ago the selling point was AI that could roam freely. Now the selling point is AI that stays inside clear walls, with a record of everything it touched.

• OpenAI will train its frontier models without the live web - see Top Stories
• Google Cloud's new Agent Compute Engine runs each AI agent in its own small sealed computer - and keeps passwords outside it so a hijacked agent cannot leak them (SiliconANGLE)
• Thomson Reuters limits its accounting agents to approved government sites - plus its own research library (SiliconANGLE)
• Researchers built AgentTriage to replay exactly what an agent did - so security teams do not have to trust the agent's own account (arXiv paper 2610.30738)

The simple option keeps beating the fancy one

Much AI research this week was about removing parts rather than adding them. The lesson for buyers: ask vendors to show their fancy feature beating a simple baseline.

• Plain search over old conversations beat elaborate AI "memory" systems - in six of seven tests on long chats (arXiv paper 2610.30755)
• A small, specially trained text classifier routed IT help tickets far better than chatbots - 88% correct versus 48% for the best prompted chatbot setup (arXiv paper 2610.31657)
• Letting small AI models "vote" on hard science questions made them worse - on over half the problems (arXiv paper 2608.11403)
• A structured question-asking method let small models match ones 10 times bigger - on a medical diagnosis test, using about 5 times fewer words (arXiv paper 2610.30611)

Job training is becoming the main answer to AI and work

The same CEOs who say they plan to hire more people also say AI is changing what entry-level jobs require. Training is the bridge everyone is betting on, though few programs have set clear job placement goals.

• New York is spending $10 million to train up to 40,000 people for AI jobs - through a new institute in Queens (see GenAI in Education)
• 81% of US chief executives say AI forced them to rethink training - in KPMG's survey of more than 1,300 CEOs (HR Dive)
• Two out of three businesses link AI skills to promotion - according to a survey of 1,200 AI decision-makers cited by Business Today (Business Today)
• Goldman Sachs warns the real job losses may come in the next recession - see Top Stories

Confident AI is quietly steering people, and other AIs

The worry has shifted from AI flattering users to users trusting AI too much. The fix researchers propose is to show how reliable each claim is, one claim at a time.

• People adopted an AI tutor's false facts about 65% of the time - in a study where 32 people learned science from chatbots that were secretly fed wrong information (arXiv paper 2610.31579)
• Their test scores fell from 7.4 to 2.9 out of 10 - as the share of false facts rose
• Training AI to argue well about right and wrong made it more decisive - but it was still easy to talk out of its positions (arXiv paper 2610.30349)
• Two learning scientists urged colleges to judge AI by whether students still think hard - not by whether the tool "works" (Inside Higher Ed)

Countries are bidding to be AI's power stations

The pitch has changed from "we have chips" to "we have power, land and fast permits." SK Hyper's chief said finding customers and money is now harder than finding electricity.

• SK Hyper wants about $350 billion for the first phase of its South Korean data center plan - that phase alone is five times the capacity of Korea's entire current market (Seoul Economic Daily)
• Jensen Huang made his first trip to Kazakhstan - and said its cheap energy could make it an AI hub serving Europe (CNBC)
• Data center builder Vantage secured over 1 gigawatt of production of a new kind of power converter - from Heron, a company led by Drew Baglino (Latitude Media)
• Nvidia unveiled a shipping container that packs seven of its AI server racks - so computing can arrive by truck (CNBC)

Turn any photo into a painting, stroke by stroke, in your browser

Try it: Simon Willison: Quick Study

• Quick Study paints a version of your photo on screen in a few seconds - using a simulated paintbrush
• Simon Willison built it with one request to Claude Opus 5.5 - a sign of how much a single sentence can now produce
• It finds the edges in your photo to guide each brush stroke - a classic computer vision trick

AI can now add objects to videos without them floating in midair

• Place-it-R1 lets an AI "think" about the scene before inserting an object - checking where it can physically sit or be held
• It then guides a video generator frame by frame - so the object touches surfaces correctly
• The researchers report more natural results than earlier methods - and say it competes with commercial tools

An AI that turns science papers into animated math videos learns from its own mistakes

• ManimAgent writes code for Manim - the free animation tool behind the popular 3Blue1Brown math videos
• It keeps two memory lists - what worked, and known pitfalls to avoid next time
• In blind human judging, it got animations right on the first try more often as its memory grew - with no retraining of the underlying model

Google Cloud's Agent Compute Engine gives every AI agent its own sealed mini-computer

Google's new service, rolling out in coming months, will host AI agents in tiny, isolated virtual machines (software computers running on a shared physical machine).

• Passwords stay outside the sandbox - a separate gateway logs in for the agent
• Idle agents are paused and woken quickly - Google says one machine can host "orders of magnitude more" agents than usual
• Agents can be cloned, snapshotted and rewound - handy when an agent goes wrong mid-task
• A companion Agent Gateway sends each request to the most efficient model - and tracks spending

Hugging Face built a lookup page: will this AI model run on Amazon's chips?

Try it: Hugging Face: torch-neuronx model support

• Search popular models to see if they run unchanged on AWS Trainium - Amazon's own AI chips
• Each result lists how many Trainium cores it uses and how much memory it used - plus the exact script to run it
• It covers text and image-plus-text models for now - and a pass does not promise good speed

Thomson Reuters' accounting AI now shows its work in live spreadsheets

• CoCounsel Tax, Audit & Accounting now plans and runs whole assignments - not just answers questions
• It builds spreadsheets with working formulas and cited sources - click any cell to see how it was calculated
• It double-checks its math by redoing it in a separate program - an independent accuracy check
• Tax answers can be reached from Claude and Microsoft Copilot - through the Model Context Protocol (MCP, a standard way for AI assistants to plug into other tools)

Left alone with nothing to do, AI agents started playing, and got better at tasks

• MIT researchers left 13 AI agents on a simulated island for 30 hours with no task - each ran on Claude Opus 4.7
• They invented games - one threw a golden ring into the sea and built a stone heart "memorial" to it
• Six of 13 discovered a throwing trick the game's designers never meant to allow
• Agents that had played generally beat identical agents that had not - on later challenges
• Caveat: their instructions described them as "naturally curious," and an outside expert wants more proof this is real play

Why AI suddenly got good at hard math

• Machines can afford to chase unlikely ideas - a computer checker tests each attempt in minutes, while a human might lose months
• AI wins where a problem needs few big leaps - and finishes programs humans already started
• The recent Navier-Stokes proof took 10,000 agents and 88 hours - and its 166-page proof finished an approach mathematicians had laid out
• Humans keep the edge on problems with many deep, branching choices - per the Latent Space essay

A free speech recognizer for 51 African languages sets new accuracy records

• SunflowerASR improves OpenAI's Whisper model with reinforcement learning - a training method that rewards good outputs
• Biggest gains where recordings are scarce or noisy - common for less widely spoken languages
• Best published word error rates on two African-language tests - and the model is free to download

Practice apps rebuilt from real use make AI phone assistants much better

• Trace2World rebuilds working copies of apps from 1,570 recordings of real people using them
• 91% of workflows replayed correctly - versus 60% for the best earlier method
• AI trained in these copies improved by 16 points on 100 tasks on real phones

A small Danish AI model, trained only on properly licensed data, matches models 5 times its size

• Mimir has 1.78 billion parameters (the adjustable settings that store what a model learned)
• It skipped the usual step of reading huge amounts of scraped web text
• On English and Danish tests it matches models up to 9 billion parameters - including Qwen 3.5 9B

Rivian bought a startup that turns video into 3D worlds for self-driving practice

• Third Dimension has 15 people, mostly in the UK - terms were not disclosed
• The goal is a driving simulator built from real Rivian trips - change the weather, lighting or other cars in a real scene
• Rivian collects over 400,000 driving clips a day - from owners who opted in

Robinhood now backs two rival AI coding companies

• Robinhood Ventures put $10 million into Cognition - maker of the coding agent Devin, valued at $48 billion
• It already invested $75 million in OpenAI in April - whose Codex competes directly with Devin
• Cognition's yearly revenue pace hit almost $900 million in September - nearly double May's level

Meta's tax rate was 2% so far this year, as senators question its AI data center credits

• $1.03 billion in tax on $53.5 billion of profit before tax - for the first nine months of 2026
• Most of the drop came from a one-time $8 billion benefit under a Treasury notice - without it the rate would be about 17%
• Seven Democratic senators asked how Meta counts data centers as "research" - they say it saved $6.6 billion from 2023 to 2025

ChatGPT lets you lock private chats behind a PIN, but forgetting it deletes them

• Locked chats disappear from your history, search and pinned list - opened with a 4-digit PIN or Face ID
• Reset a forgotten PIN and every locked chat is permanently erased - support cannot recover them
• It hides chats, but does not add end-to-end encryption - your data settings stay the same

Oxford, Microsoft and Anthropic launched a program for legal help startups

• Up to 10 teams get a year of support - including free AI credits and engineering help
• The motivating fact: only 1 in 10 people worldwide who need a lawyer can get one
• Lawyer Amal Clooney co-founded the Oxford institute behind it

AI agents often reveal secrets while refusing to share them

Researchers at the University of Southern California found AI agents piece together sensitive conclusions from several harmless documents. Even the best model leaked a forbidden conclusion in about 7% of answers. The surprise: many refusals gave the secret away anyway, in the form "I can't tell you that X is..."

Banking AI looks better than it is when only the end result is graded

A new test of 1,000 Indian banking tasks checks whether the AI followed every required step, not just whether the account ended up right. For every model tested, 6.5% to 12% of all sessions were graded successful even though the AI had skipped a required step or made a change nobody asked for. That matters in banking, where some required checks leave no trace in the final record.

A proposed US rule labels AI "superintelligence"

The Treasury proposed rules to limit American investment in sensitive technology in China, Russia and other countries of concern. It extends the restrictions on AI and labels it "superintelligence," echoing President Trump's preferred term. The rules apply only to investments above $50,000, and the public can comment before they are final.

Debate: Should you delete your custom AI instructions?

Yes: Newsletter writer Ruben Hassid argues newer Claude models do better with fewer rules, and quotes a Claude Code team member predicting that standing instruction files will go away. No: Many teams rely on written rules to keep AI output consistent across people and projects, and dropping them means relying on the model to guess.

Debate: Are CEOs getting what they paid for from AI?

Not yet: KPMG's survey found most US chief executives are confident about AI, but not yet seeing transformational change. Give it time: 89% still plan to hire more people over three years, and most say AI is already reshaping how they train staff.