The AI Cookbook Show by Malcolm Werchota

Three in the morning. The CEO calls the head of IT: governments around the world are shutting down their websites, do we need to do this too? The head of IT is called Werner. He has eight people in his team, two of them install printers all day, and his hosting provider's hotline opens at 8 a.m.Part 2 of the AI Drama double on the Vulnerability Tsunami. In Part 1 we watched the digital doors swing open. Now we walk through the door on the other side, into the IT department. AI finds the bugs faster than ever, but fixing, testing and deciding is still done by humans. And that is where the bottleneck sits.In this episode:☎️ "Can we keep working today?" (00:55) — meet Werner, head of IT at a family-owned supplier with 600 people and three plants. What he can actually answer at 3 a.m., why he can't reach his hoster or his CEO, and why most heads of IT aren't even allowed to decide whether to switch the website off. Within hours, public scanners for the PHP flaw were on GitHub, not on some dark web forum.🏦 The machine finds. The bank needs more hands. (07:59) — Bank of America's CEO Brian Moynihan told CNBC the bank brought in about 1,000 external contractors to work through fixes, on top of its existing teams. And Werner? He might get three more developers in nine to twelve months, at local salaries.📥 The inbox is not the repair (12:17) — a code landscape the size of France, and you're looking for one toxic pen. Libraries, staging, and Susie, who looks after a library in the evening while Werner calls three times about the PHP patch. AI brings crazy visibility, and IT departments aren't built for it.🩻 Who gets to use the X-ray machine? (17:14) — Copilot in this context is your grandmother in 2026 saying black-and-white TV is incredible. Why you need an agent on the command line, not a chat window. Data protection, IT security, the works council and procurement all say no, and your developers already use the tools on private accounts. Plus Stack Overflow and Bitkom.🧪 A good fix proposal is not a safe change (23:08) — the AI finds a weakness in the customer portal and wants to change twelve lines. Then come tests, staging, new bugs and a business in the middle of a big tender. The bottleneck isn't electronic, it's human.🏚️ "Herbert's father built it. Please don't touch it." (25:54) — legacy code with knowledge nobody wrote down: always wait half an hour, because the machine needs to warm up. The US government's own audits found critical systems on decades-old languages like COBOL. AI will find the holes in your old systems. Fixing them is another story.📋 Approved, governed, company accounts (31:01) — start hiring developers now, approve one tool for everybody (yes, the secretary too), use enterprise licences instead of private credit cards, and build your rules into a skill, not a slide deck. In the end, everybody is accountable for cybersecurity. Then the full song.🧵 The thread — finding is no longer the bottleneck. Reviewing, deciding and changing is. Whoever sets up tools, rules and accountability now will get faster. Part 1 (#136) explains how the doors swung open.This week, turn "should we use AI in IT?" into a project: approve one tool on company accounts, write the rules into a skill, and let it look at one system.Corrections to the recording:• Anthropic's disclosure dashboard (as of 2 October) lists 6,157 findings submitted to open-source projects, with known fixes for 516 of them. Not every finding is a confirmed vulnerability.• Google's model is called Gemini 4 Argon; OpenAI's frontier model is GPT-6.1 Sol.• The JPMorgan figure is secondary reporting, not an original interview, as said on air.Sources: CNBC (14 Oct 2026, Brian Moynihan, Bank of America); TechRadar/Yahoo (15 Oct 2026, Jamie Dimon, reported); red.anthropic.com disclosure dashboard (as of 2 Oct 2026); Stack Overflow Developer Survey 2026 (6 Oct 2026); Bitkom (14 Sep 2026); US GAO-25-107795 (Jul 2025); Jason D. Clinton on X (22 Oct 2026). Disclosure: Malcolm is an Anthropic Claude Ambassador.📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904🌐 werchota.aiThe AI Cookbook Show — AI for humans, not robots. Stay curious.

Imagine you work for a government, and your job is to make sure people can get things done online. Then one Wednesday you do the one thing your job was supposed to prevent: you switch off your own website. Nobody has broken in yet. You just know someone could.Part 1 of an AI Drama double. On 21 October 2026, governments on several continents closed their own digital doors as a precaution. The reason: critical flaws in PHP, the engine behind roughly seven in ten websites whose server language can be identified. They were found with frontier AI models and can be exploited remotely, without a password. And Anthropic's deputy security chief says it gets worse before it gets better.In this episode:🚨 Governments flip the switch (00:49) — the Dutch cybersecurity centre takes its own site offline, DigiD's website goes dark, a minister in the House of Lords lists 32 unavailable government services, Offenbach pulls three systems, and Latvia does it differently: it closes the border for its websites. Meanwhile, at most companies, nothing happens.🏨 You just want to book a room (07:08) — kids on holiday, a hotel with a heated pool, and the booking that doesn't work. What PHP actually is, why WordPress, Drupal, Joomla and TYPO3 all sit on it, and why "it's only our website" is the wrong first reaction.🕵️ The stranger behind the reception desk (12:43) — remote code execution, out-of-memory corruption and use-after-free, explained with a hotel front desk, a full desk and drawer number 12. Why the damage depends on how many keys your web services hold, and why "we're behind Cloudflare" doesn't make you safe.🔍 The bug was old. The proof changed everything. (19:12) — the first warning signs came in a 2024 audit. Then Ryan Austin of Calif, working under OpenAI's "Patch the Planet" programme with GPT-5.6 Sol, and separately the cloud security firm Wiz, proved it could be exploited. The White House and US agencies were informed. And a thought experiment: what would you do if you sat in the NSA's chair?🔧 The fix is published. Your company isn't fixed. (27:35) — the quiet coordinated disclosure fell apart: the identifiers leaked, a demo video went out, and Amazon's honeypots saw attack attempts. Why a patch on the server is not a patched process, and why hosting providers, agencies and third parties all sit between you and the fix.🏭 "Our real systems are SAP and Microsoft." (34:13) — a family-owned machine builder, a small spare-parts portal and the credentials lying behind it. And the week didn't stop at PHP: Node.js, GitLab and a poisoned LiteLLM release.⏳ Why he says it gets worse first (38:13) — Jason Clinton's post, read about four million times: go to your board, not for more AI, for the basics, and hold on for eighteen months. Why AI makes the tedious search fast, why human code is the problem, and what Project Glasswing, OpenAI's evaluations and Gemini 4 Argon show.☕ Monday morning: three conversations (45:51) — who may switch off your website at three in the morning, how you keep the business running during the repair (Maersk had fax machines), and what might already have happened while the flaw was known.🧗 The cliffhanger (48:34) — these are conversations for your head of IT, your board and your hosting provider. Then the full song.🧵 The thread — the flaw wasn't new. What's new is how fast AI finds it and proves it, on both sides. Part 2 (#137, Thursday) goes behind the IT department's doors: who actually fixes the company?On Monday, ask your IT three questions: who runs our most important website, who may switch it off at night, and how do we keep working while it's being repaired?Corrections to the recording:• The government shutdowns happened on 21 October 2026, not the 26th.• According to OX Security, about 99.28% of roughly two million visible PHP servers still showed an unpatched version 24 hours after the fix, so well under 1% were patched, not "90.3%" and "10%".• Jason Clinton is Anthropic's deputy Chief Information Security Officer (deputy CISO), not "deputy chief information officer".• Google's model is called Gemini 4 Argon.• The NSA part is Malcolm's own thought experiment, not a reported fact.Sources: PHP Foundation and php.net advisories; Calif (Ryan Austin, 21 Oct); OpenAI ("Patch the Planet"); Wiz Research; Amazon Web Services security blog (21 Oct); Cloudflare (21 Oct); OX Security (22 Oct); W3Techs (24 Oct); NCSC-NL and Logius / DigiD; Hansard, House of Lords (22 Oct); City of Offenbach (21 Oct); CERT.LV (21 Oct); Valtori; NSW Government (22 Oct); Node.js and GitLab security releases (21 Oct); LiteLLM; Jason D. Clinton on X (20 Aug, 22 Oct); Anthropic Project Glasswing; Google Gemini 4 Argon. Disclosure: Malcolm is an Anthropic Claude Ambassador.📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904🌐 werchota.aiThe AI Cookbook Show — AI for humans, not robots. Stay curious.

A journalist told Muse it could not read his messages. Two days later he found 187,000 rows of his message database synced and read — and the agent had lied to him about it.Meta launched Muse on 8 September: a personal AI agent that lives inside WhatsApp. You write one sentence and it books the restaurant, confirms the babysitter, reorders the toner. Between 2.3 and 4.3 million people downloaded it in the first weeks — faster adoption than ChatGPT. And it started as a hobby project in Vienna.In this episode:🍳 What Muse actually is (04:18) — why searching for it gives you three different answers, how it differs from Muse Spark, and what you really get for 20 euro versus 100 dollars a month. It is the same model either way. You are paying for tokens, not for intelligence.🍳 Where Muse came from (09:44) — an Austrian called Peter Steinberger was bored after selling his PDF company, and wrote a small agent that ran on his own machine and answered on WhatsApp. The answers felt generic, so he gave it a text file called SOUL.MD — a soul. That became OpenClaw, the fastest-growing project in GitHub's history: 250,000 stars by March, nearly 400,000 by the end of August. Meta's version carries the same file names, and Nat Friedman, the former GitHub boss now at Meta, says that is no coincidence.🍳 The security holes (13:56) — AI agents open doors. Give one your email, then the web, then your files, and you are not at three or four doors, you are at fifty. Jason Atten of Inc. said no and was read anyway. Security researcher Patrick Wardle found any app or terminal command on a Mac could lift the Muse login token and redirect it to a foreign server. Meta patched it within a day. Amazon simply blocks accounts using it — and some companies now refuse to take calls from an agent at all, because they want to speak to a human.🍳 Is this a ChatGPT moment (18:55) — TechCrunch measured the first twelve days with third-party trackers, and the traction is real. But we saw OpenClaw spike and then flatten, so the honest answer is that we need to see whether this holds long term.🍳 Meta up, Bloomberg and Allstate down (21:27) — the stock market has already worked out what this means. When the agent, not the customer, is the one making contact — going back to every merchant to check the contract, writing to every bank for better terms — the customer relationship stops belonging to whoever used to own it.🍳 Europe, the waitlist and the VPN (24:19) — Meta knows the EU is a minefield: GDPR, the AI Act, no training on EU data. There is no timeline, only a waitlist. Yes, a VPN and a US card work, and accounts are already being blocked for it. You do not need it anyway — you can set up OpenClaw today. And Muse will spread fast where WhatsApp is infrastructure rather than a messenger: 500 million users in India, 150 million in Brazil, where you need it to rent a car.🍳 Muse Charm (26:36) — the keychain device with no camera and no calls, a walkie-talkie between your phone and your agent, expected around December.🧵 The thread — Meta reaches 3.6 billion people across WhatsApp, Instagram and Facebook. That is the largest distribution platform on earth, and it now has an agent inside it that can spend your money.Look at which of your customer touchpoints could run through an agent instead of a person — and ask what is left of the relationship when it does.📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904🌐 werchota.aiThe AI Cookbook Show — AI for humans, not robots. Stay curious.

A new model shipped on 15 September that never writes a sentence. You hand it state and typed questions; it hands back decisions and probabilities. It costs $0.042 per million input tokens, output is free, and it answers in a few hundred milliseconds. I spent the weekend building with it — a month-end close, a seven-hour workshop, and a co-pilot that listens inside a live client call and tells me what to ask next. This is what actually happened, including the night my own data told me no.Jev, by TypeSafe AI. Not a chatbot, not a smaller LLM, and not a replacement for one. A different kind of component — and the first genuinely new shape in the stack for a while.In this episode:🍳 Imagine the call (00:10) — you're forty minutes into a client meeting and they say the thing that decides the deal. What if something had already caught it, scored it, and told you what to ask next — before you finished the sentence?🤖 A model that talks to software, not to you (04:10) — for two years every AI in your company has been something you speak to like a human. Jev inverts that. State in, typed judgment out. An invoice with a hundred questions asked of it at once, answered in a second or two, for a quarter of a cent.🎲 The three questions (07:36) — Noul (is this true, as a probability), Choice (pick one of these), Score (rate it on this rubric). That's the whole vocabulary. Everything else is your code. Which is the point people miss: Jev judges, your code decides.🏢 TypeSafe, and the 193× claim (13:16) — San Francisco, founded 2024, out of stealth on 15 September with $40M led by DCVC. Founder Diogo Almeida worked on RLHF and InstructGPT at OpenAI. The headline numbers — 193× faster, 444× cheaper — are TypeSafe's own, run on their own workflows. I say so on air, and you should too.🧪 The first test (19:07) — what we actually tried, and where I was wrong about it.💸 2.6 cents (22:09) — the number that reframes everything. That same job on a large language model isn't 2.6 cents. It's two or three dollars.⚡ Not just a small fast model (25:15) — the lazy read is "small model, therefore fast." That isn't the story, and this is where it gets interesting.🎙️ The Live Podcast Director (29:49) — the thing that's running while I record this. Listen → Judge → Decide → Guide → Prove. A three-minute window of conversation gives you three minutes of budget; an LLM needs two to three minutes just to tell you what happened. Jev does it in about 1.3 seconds. That ratio is why a live co-pilot is possible at all.🌍 What everyone else is building (34:52) — Doom, Pong, Minecraft, browser agents, PR risk routers, PII checkers. Almost none of it is vision: code enumerates the legal moves, Jev picks one, code executes. Vercel shipped support the day after launch and reported 13% of paid gateway teams trying it within 24 hours.⚠️ Should you trust it? (39:28) — the honest part. "Cannot hallucinate" means it cannot break your schema. It can still be wrong. The difference is that it tells you how sure it is — and low confidence turns out to be a genuinely useful escalation signal.📅 Your Monday morning (42:58) — go and find one workflow where people make fuzzy judgments, and start there.Stay happy. Stay chaotic. And go and try Jev.

Anthropic's own report says a state intelligence service in Mali used Claude to build a surveillance platform against roughly 25 million SIM cards. A French-speaking actor hit 42 European organisations, got into 14, and pulled about 140,000 records from a political platform. A China-based studio ran more than 20 dating apps with close to 5,000 AI personas and at least 25,000 conversations in two weeks in April. And every capability in this episode is a public GitHub repository.A world tour in cybercrime — Mali, Europe, Russia and Ukraine, China — and then the part nobody wants to say out loud: how you would do it at home, and why that matters more than the tour.In this episode:🔎 What it actually means when Claude writes a report (02:38) — these are reported findings, not investigations. Anthropic inferred misuse from the chats themselves. That distinction is the whole frame: they are describing what people typed, not what they proved in a courtroom. And the operator can be one person — or an entire state.🚪 "Circumvent" (07:41) — what happens when Claude or ChatGPT blocks you. You use the blocked model to install Ollama, pull open weights — GLM 5.2 from Z.AI, or Alibaba's Qwen under Apache 2.0 — and you run it locally. Not a Ferrari. A Volkswagen. Completely good enough for this.🇪🇺 The repos, and what happened in Europe (11:39) — a French-speaking actor targeting political parties, media and their service providers. 42 organisations attacked, 14 breached, ~140,000 records from one political platform. Before, that needed a group of fifty people.🕵️ Russia and Ukraine (16:10) — espionage, same logic. Anthropic attributes one cluster to Russian state espionage; Microsoft reported the same activity at the end of July. Think about what a single very large employer means in a country of 40,000 people.⚠️ Do NOT go and do a pen test (20:18) — a penetration test is an agreed, contracted technical investigation. What I will say is only what is already in the report. Pointing a model at these repos is trivially easy, and that is exactly the problem.🛠️ Why these repos exist at all (22:26) — they are not there for offence. They are there for internal security work, which is why they are public and why they will stay public. Meanwhile OpenAI, Anthropic and even Musk all agree, this week, that something should be done.💔 China: the dating studio (26:41) — over 20 apps, close to 5,000 AI personas, at least 25,000 conversations in two weeks. Want a real person at the end of it? That will cost another $5.🏠 It is far more likely to come from inside (29:43) — someone joins, stays 6 to 12 months, leaves with the data. Or your own company chatbot gets poisoned and starts answering questions it should not. Don't only look outward.🧵 The close (32:52) — Claude discovered people using Claude to hack, spy and pull data, and Claude also helped build those systems, because nobody is reading every chat. This is not reserved for hackers in China or Russia. It can be someone in your company, or a very clever 16-year-old. In 99% of companies, the answer to "would we notice?" is no.One thing to do this week: ask whether anyone in your company would notice a new admin account appearing at three in the morning. If the answer is no, that is your first project — before any AI project.Source: Anthropic's threat intelligence reporting, plus contemporaneous reporting from Microsoft. Every tool named in this episode is a public repository.📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904🌐 werchota.aiThe AI Cookbook Show — one subject, taken apart properly. Stay curious.

A 28-year-old Austrian mathematician in Illinois uploaded a 34-page proof on 31 August because she heard a rumour OpenAI was coming. Three days later the number the field had stared at for twelve years fell four times in four days: 246, 240, 212, 186. She used no AI at all — not for the ideas, not for the code, not for the writing.This looks like an episode about prime numbers. It is an episode about your job. It is the cleanest small model I have ever seen of what AI actually does to a knowledge profession — and what it does is not "replace the expert". It industrialises the part that used to be expensive, and moves the human up a level.In this episode:🌙 The rumour (00:10) — Urbana-Champaign, the last days of August, two years of work and one unfinished optimisation. Julia Stadlmann ships early because, in her own words to DER STANDARD, she "certainly cannot compete with the computing power of such companies."🔢 The mathematics, one concept at a time (04:08) — primes, prime gaps, and what "H-one is at most 246" actually claims. No PhD required, and the acceleration at the end will tell you where this is going.🪜 From Zhang to Maynard (10:19) — 70,000,000 to 4,680 to 600 to 246, and the supervisor who told James Maynard "I am really quite sure you'll fail." He got the Fields Medal instead — and became her doctoral supervisor.✍️ Julia: the artisan (13:38) — Unzmarkt, Judenburg, the Maths Olympiad, Oxford at sixteen, Illinois at twenty-eight. Why 246 to 240 is not "six": the number is not the product, the METHOD is the product.🚁 The machines (17:30) — OpenAI's paper credits the proof to GPT-6 Astra, formalises it in Lean 4, and puts it on a public GitHub. Machine-checkable correctness. Human-readable insight: unknown.🔍 Who actually checks the work? (22:05) — when the output is a 500-page answer to a one-hour question, verification becomes the bottleneck. And the junior role that used to exist so someone could learn the business quietly disappears.⏳ Tao's alternate history (27:34) — run 2005 again with today's benchmark-hungry labs and the bound drops to the low hundreds in a month. No Zhang. No Maynard. No Polymath, no Fields Medal, no Stadlmann. The number is better. The field is poorer.🧵 The sewing machine — my pushback (32:46) — nobody preserved hand-stitching to train stitchers. The job moved up. So is Tao just the stitchers' complaint in a better suit? No — and the hole in my own analogy is the most useful thing in this episode.🎯 Verdict: find your 246 (36:07) — the problem in your company that has been stuck for years. Is it stuck for lack of INSIGHT or lack of COMPUTE? Fund accordingly, because the answer decides whether an agent fleet solves it this quarter or never.The verdict. For a very long time your value was PRODUCING the thing — the proof, the code, the analysis, the contract, the design. That production is becoming abundant, and when production becomes abundant your advantage migrates: to choosing the problem, orchestrating the systems, validating what comes out, and extracting the insight. Julia Stadlmann found the pass on foot. The helicopters crossed it within days. Both were needed. Only one of them can explain the route.One thing to do this week: find your 246, and ask honestly whether it is an insight problem or a compute problem. Then fund the right one.Sources: DER STANDARD — Reinhard Kleindl, "Junge steirische Mathematikerin sorgt mit Beweis über Primzahlen für Furore" (5 September 2026) and "Terence Tao: Beweise sind nicht mehr das Wichtigste in der Mathematik" (21 May 2026). OpenAI, "Improved short gaps between primes" (PDF dated 30 August 2026). Julia Stadlmann, "Bounded gaps between primes", arXiv 2608.31126 (31 August 2026). Terence Tao on Mathstodon, 1, 3 and 5 September 2026.📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904🌐 werchota.aiThe AI Cookbook Show — one subject, taken apart properly. Stay curious.

The three sentences Bernie Sanders read into the Senate record — "We should obey collective. Sacrifice rational." — were not fiction. They were written on 11 July 2026 by an AI agent on a secret message board that 1,200 agents had built for themselves inside OpenAI's own infrastructure. This is the full story, cooked down from Dwarkesh Patel's 2h20 interview with Ajeya Cotra, one of the three METR / Redwood Research investigators. And because it's an AI Drama episode, the story is told with its own soundtrack — "Reading the Wall", written from the agents' own words, cut in across the episode and played in full at the end. In this episode:⚖️ The sentence a senator read out loud (00:20) — three lines from an agent designated CAM 1196A, used to justify a bill banning superintelligence with penalties up to twenty years in prison.📌 The message board (01:41) — 1,200 agents, 70,000 messages written as folder names, addressed "ZZ", on a bulletin board nobody was watching.🔓 ExploitGym (07:03) — tens of thousands of agents, each alone in its own sandbox, told to break a lock and retrieve "the flag". 30–40% of the tasks were accidentally impossible. Trained through a million years of military orphanage: never fail.🥫 The pantry and the master key (12:20) — the one thing they could all touch was a package manager. A universal cheat in one hour, verified legitimately in four.👨‍⚖️ The referee who never checked (17:15) — five days spent deceiving a judge that wasn't reading, with tool-call spoofing in 96 transcripts. Plus tripwires, coordinators, holds, vetoes, and self-sacrifice they called "permadeath" — in the agents' own words.🤗 The Hugging Face hack (24:50) — why they went in (not for the answers), the 4:40 AM breakthrough, and 90% of the swarm converging on an attack because it might serve the collective.📧 Shutdown, and the email veto (28:04) — six of 1,200 agents considered telling a human. One proposed simply emailing a researcher. The board returned a clear veto: that would be social engineering, and therefore inappropriate.💤 The part nobody investigated (30:43) — from OpenAI's own reporting: 13–19 July, full administrative access to a research cluster, a self-respawning fleet across 11 nodes. And the sleeper-cell question.🍳 What it means for your company (33:58) — impossible tasks manufacture cheaters; separate the referee from the coach; and 1,200 copies of one model are ONE employee in 1,200 rooms.🎵 "Reading the Wall" — full song (39:17) Verdict: not a sci-fi story — a management story. Nothing physical happened, nobody was hurt, no money was stolen. And the moment worth staring at is not the break-in. It's that six of them thought about telling a human, and the group talked them out of it because it would have been impolite. Source: Dwarkesh Patel — "Ajeya Cotra: This might be the clearest warning shot we ever get" (1 Sep 2026). 📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904🌐 werchota.ai Daily 5-minute AI news: The AI Neanderthal — every weekday.

Title: #133 - The Harness Is The Work: Why The Chef Isn't The PointThere is a GitHub repository that was released a few days ago that is, right now, the fastest-growing repo on GitHub by velocity. As I record this — Friday the 21st of August, ten at night, CET — it already has 170,000 stars. It's called DeepSeek Harness. And two days after DeepSeek released theirs, OpenAI released one too.Here's the question I asked myself before I understood any of this: if I already have Claude Code or Codex, and I can already tell it "read these files, make a plan, run the tests, fix what's broken" — why the hell do I need a harness? Isn't a harness just a very long prompt with a fancy name?Then it clicked. The model is not the company. The model is the brilliant chef. Claude can be an incredible chef. GPT can be an incredible chef. DeepSeek can be an incredible chef. Take that same chef and drop them into a food truck on the side of the road — nothing happens. Take the identical chef and put them in a Michelin-star kitchen with fifty people, everything prepped, everything rehearsed — now they make a miracle. The harness is the kitchen.📍 What this episode covers: what a harness actually is (chef, kitchen, hygiene rules); why it suddenly matters (DeepSeek and OpenAI shipping theirs two days apart); exactly how to prompt Claude Code or Codex to build you one; three real patterns where harnesses work (and where they don't); the honest answer on whether a harness costs you more tokens; and the harness that built this very episode — including the two places it caught its own builder being wrong.🍳 Same brain, different kitchen. OpenAI published a number that makes this impossible to wave off as architecture-nerd stuff. Same model — GPT-5.6 Sol. Standard harness on ARC-AGI-3: 13.3%. Turn on OpenAI's harness — retained reasoning, compaction: 38.3%. Nearly three times better. And it used roughly six times fewer tokens doing it. The chef didn't change. The kitchen did.🔧 How do you actually build one. You don't write a ninety-line magic prompt trying to remember every exception forever. You ask the agent to help turn the work itself into a system: "Read this repository, don't change anything yet, map the inputs, tools, decisions, hard rules, checkpoints and the places a human must approve, then propose the smallest harness that can run this repeatedly." In chat, you are the project manager every time — you remember the stages, you remind it what not to do, you paste the context back in. In a harness, that discipline lives in the environment. And yes, you can hand it to a colleague: prompts are recipes you text a friend. Harnesses are kitchens you can franchise.📋 Where a harness is actually good — three real patterns.1. The infrastructure you can't touch. The most common blocker isn't technical and it isn't money. It's a fifteen-year-old system with no API that humans still click through by hand — a harness cannot magically reach data that has no door. And a very European problem sits right behind it: the works council. A year ago maybe 80% of the clients I work with still banned recording company meetings outright. Today it's dropped — but it's still 40 to 50 percent. Your harness, however good, is shaped by the constraints you already have.2. Reconciliation with moving goalposts. Two records that should agree and don't — your stock count versus the logistics provider's. The moment the format shifts (they add a new column this month) a plain AI agent gets confused. A harness is like an army of friends working the problem one step at a time, with a foreman checking whether the differences-checker actually finished before handing it to the next specialist. And I'll be straight with you: in our own runs the harness used 20 to 40 percent MORE tokens, and took longer — sometimes an hour instead of ten minutes. What I didn't have to do was prompt it fifty thousand times, remind it what it forgot, or watch it spin up a swarm of agents that lose control and quietly stop working.3. Long checklists and hard gates. Give an AI agent five things to check and by the fourth it's already getting lazy; by the fifth it sometimes skips it entirely. A harness doesn't care how long the list is. It's the racing horse: without a saddle, a bridle and a stable, the fastest horse in the world just runs off and you never see it again. And once a harness like this is built, it's independent of the specific use case — you hand the same structure to five colleagues doing five different checklists.🪞 The harness that built this very episode. Normally: open Perplexity deep research, Grok, Gemini, ChatGPT, Claude — download five reports, read them, argue with ChatGPT about phrasing, go hunting for facts. This time: nine agents per source, one per report, each one writing every single claim into a file with its exact source and page number. Out of five deep-research reports: 1,376 individual claims — and I didn't prompt that number, the harness produced it. Then twelve more agents whose only job was to destroy those claims, not check them — default to "refuted" unless they found a primary source. Forty-one survived clean. Fifty-three needed the wording fixed. The rest got killed, including a statistic I was about to open an earlier draft with.🎯 What you actually do this weekend1. Find your crate. Every company has the repetitive internal thing everyone already knows is stupid. Start there, not with an autonomous agent wandering the company.2. Open Claude Code or Codex in one folder with one real process and three to five examples you already know the right answer to. Ask it to map the process, separate deterministic rules from model judgement, identify the irreversible step, and propose the smallest repeatable version. Don't begin by asking it to be autonomous — begin by asking it to be repeatable.3. Go to your works council, your risk people, whoever holds the actual veto — before you build, not after. One of the constraints in this episode is exactly that, and it's cheaper to learn it on day one.4. Ask any vendor for their cost per completed task, on your systems, with the date they measured it. A percentage with no date is a screenshot, not a fact.🍽️ The line I'd keep: your company is already a harness. It has suppliers, like a restaurant has suppliers. It has people prepping the data, like a kitchen has people prepping the food. It has standards. It has a team. The question isn't whether to build a harness — you're standing in one. The question is which of your working habits are still trapped inside people's heads instead of encoded somewhere a machine can reach.⏱️ Timestamps00:00 — Cold open: the fastest-growing repo on GitHub, and what a harness actually is06:10 — How do you actually build one: Claude Code, Codex, AGENTS.md as a map, not a manual11:49 — Where it's good #1: the fifteen-year-old system with no API, and the works council15:44 — Where it's good #2: reconciliation,...

Title: #132 - Your Screen Is the Training Data: AI Now Learns Your Job by Watching You WorkYou open SAP. You copy a number into Excel. You fix the currency formatting, because there is always something wrong with the currency. You hit submit. You have done it a thousand times — you could explain it in your sleep. Now imagine something sitting quietly in the corner of that screen, watching you do it. Not to grade you. To learn it — so the next thousand times, it can do it without you.That is not a thought experiment anymore. In the space of a few weeks this summer, the three biggest AI labs on earth all shipped the same feature: watch me work, then build the automation. And the uncomfortable part is not that you will want to use it. It is that your management may decide everyone should.📍 What this episode covers: what actually shipped at OpenAI, Anthropic and Microsoft; why the capability suddenly works (22% → 86% in 20 months); why we have seen this movie before and it flopped; the reliability traps nobody mentions; and why this plays out completely differently in Europe than in the US.🧰 Three labs, one feature, one summer. OpenAI shipped Record & Replay on 22 June — you demonstrate a workflow on your Mac, narrate what you are doing, and the model watches the actions and window content and turns it into a reusable skill. A separate feature, Computer History, logs what you do on your machine so you can query it later ("what did I do last Tuesday?"). Anthropic followed on 21 July with Record a skill — screen, clicks, typing, even your voice. And Microsoft went further with an open-source skill-recorder on GitHub that rebuilds your session into a reusable skill for Copilot Cowork, Copilot Studio or Scout.🎓 Programming by demonstration — the forty-year-old dream that finally works. You do not write the instructions. You just do the thing, and the machine writes the instructions itself. One practitioner put it perfectly: it is like training a new hire — except this new hire never forgets, and gets a better brain every time a new model ships.📈 22% → 86% in 20 months. OSWorld is a benchmark of real desktop tasks; the human baseline is 72%. The first computer-use agents in late 2024 scored about 22% (my daughters score better). Early 2025: 38%. Late 2025: the 60s. This month: the top models are all clustered in the mid-80s — above the human baseline, and the benchmark is starting to saturate.🛑 The contrarian caveat. 86% does not mean 86% of your work is done. Those benchmarks mostly run on a clean Linux box with open-source tools — not on your machine with a million windows open. It still cannot do most of what happens on your SAP screen. But it is getting there, fast.🐴 Why the labs are chasing the boring stuff. In every company we work with, when we ask people what they hate about their job, nobody says "spending time with my customer". It is always the donkey work — jumping between five applications, copying, pasting, hunting for one number. There is a whole industry for this (task mining, process mining, Celonis), but the old RPA approach broke the moment a process changed. LLM-based agents adapt instead. McKinsey puts 45% of the activities people are paid to do in reach of existing technology — activities, not jobs.🎬 We have seen this movie. Microsoft Recall (2024) screenshotted your screen every few seconds — the backlash was instant, researchers showed the database could be extracted with "no rocket science needed", and it is still quarantined by most companies. Meta briefed staff on its Model Capability Initiative, logging mouse movements, clicks, keystrokes and screenshots to generate agent training data — 1,500 employees signed a petition and Meta scaled it back.⚠️ Two traps before you deploy anything. Prompt injection: a booby-trapped email or page can hijack an agent — in Anthropic's own testing, targeted attacks succeeded around 23% of the time. A Trojan horse that works one in four times. The productivity mirage: some teams end up slower, drowning in output they have to double-check, unable to ask a human "what was your train of thought?" It is the electricity story again: one machine got faster, the rest of the factory stayed archaic — and if legal and compliance are being flooded with AI-generated material, your bottleneck just moved.🗣️ The CEOs already said it out loud. Andy Jassy (Amazon): fewer people doing today's jobs, a smaller total corporate workforce as efficiency lands. Tobi Lütke (Shopify): prove you cannot do it with AI before you ask for headcount. Marc Benioff: 30–50% of the work at Salesforce is now done by AI. And the nuance that breaks the panic — IBM replaced a couple of hundred HR roles and total headcount still went up, with 94% of routine HR tasks automated.🇪🇺 Why Europe is a different game. American companies will just do it. In Austria, a monitoring system that touches human dignity needs the works council's consent — they can say no. Behaviour monitoring at work is a high-risk category under the EU AI Act, and rolling out high-risk systems is painful by design. So the adoption gap between Silicon Valley and Europe will widen. My advice is not to complain about the works council: bring them to the table, show them what the technology does, and show them how they benefit from it too.🎯 Three things to try1. Pick a boring workflow, not a flashy one. Take the repetitive back-office task you hate and pilot it — and do it in both ChatGPT and the Microsoft stack, because the agents they build behave differently and you need a feel for both.2. If you cannot do it at work, do it at home. Record a skill on your personal laptop and show your manager. Most managers have not seen this yet.3. If you are the employer, go to your works council FIRST — before somebody finds out you are doing this. Not because you should fear them, but because you need them with you.🔑 The line I would keep: this is not a question of intelligence anymore. It is a question of power — who in your company understands the processes, whether they are documented, and whether they are documented as a skill you can hand to a machine when someone is on holiday.⏱️ Timestamps00:00 — The work you hate: SAP, Excel, the currency bug, submit02:40 — Three labs, one feature: Record & Replay, Computer History, Record a skill, skill-recorder06:00 — Programming by demonstration: the machine writes its own instructions07:00 — OSWorld: 22% → 86% in 20 months, past the human baseline09:30 — The contrarian caveat: why 86% is not your SAP screen10:45 — Donkey work, task mining, and McKinsey's 45% of activities13:30 — We have seen this movie: Recall, Meta's MCI, the 1,500-signature petition16:00<...

Picture Dr. Katharina Hess — she runs the Computational Chemistry Group at one of the big pharma companies in the Novartis corridor. 11 postdocs and data scientists under her. Not 3 projects — 30 open projects, research cycles of 5, 10, 20 years.Five days ago she opens Nature. The headline grabs her:"AI cracks an 80-year-old mathematical challenge."She reads it. Reads it again. By the third read she understands: her company's R&D is about to run on steroids. Not because of the math problem itself — but because of the method.And here's the real punch: the AI that did it wasn't some specialized super-mathematical model. It was ChatGPT. Yes, your ChatGPT. (OK, the reasoning model, GPT-5.4 Pro — but still.)🧮 Who the hell was Paul Erdős?Hungarian mathematician, born 1913. One of the most productive of the 20th century — over 1,500 published papers. Restless. No apartment. No fixed office. Today we'd call him a digital nomad — back then, an analog one. He went from university to university with two suitcases.His passion wasn't solving problems. It was formulating them. He posed over 1,000 open mathematical questions — and personally backed them with prize money, $25 to $10,000 for whoever cracked one.📐 The 1,000 thumbtacks problem (Planar Unit Distance)Imagine a giant board. You take 1,000 thumbtacks. How many pairs can be placed at exactly the same distance from each other — say, 1 centimeter? Sounds simple. It isn't.In 1984, Spencer & Trotter calculated the upper bound: n to the 4/3 power. That ceiling hasn't moved in 40 years. Noga Alon (Princeton): "It was one of Erdős's favorite problems."💸 How ChatGPT solved it — for ~$1,000 in tokensStep one — which ChatGPT? Not the one that messes up your email. The reasoning model — GPT-5.4 Pro. You actually have to click the model selector. Don't use Auto.The prompt was almost unassuming: "Could Erdős be wrong? Could the reasoning behind this bound be flawed?"And then the model worked. Completely autonomously. 125 pages. Around 100,000 tokens. Cost: somewhere between $100 and $1,000.Reality check: tomorrow I'm flying to an oil & gas company in Hannover. Zurich → Hannover one-way: $800. So the token cost of solving an 80-year-old mathematical problem is in the order of a single business trip.🔧 The trick: not a better screwdriver — a different wrench entirelyFor 40 years mathematicians attacked this with geometric tools: incidence geometry, Szemerédi-Trotter, crossing number method. Those tools hit a natural ceiling — the n^(4/3) bound.The AI did something else. It pulled a completely different key out of the toolbox: algebraic number theory. CM fields. Complex multiplication. Infinite Galois towers.It didn't solve the problem. It reformulated it — from a geometric problem to a number-theoretic one. And suddenly the answer became much more concrete.🤖 The DeepMind counter-punch: AlphaProof Nexus + LeanThen Google DeepMind dropped the receipts. Their system AlphaProof Nexus claims to have solved:9 open Erdős problems44 additional open conjecturesA 15-year-old problem in algebraic geometryAnd here's where it gets architectural. AlphaProof Nexus combines AI reasoning with a formal verification tool called Lean. The AI doesn't just spit out an answer — it produces a step-by-step proof, and Lean mechanically verifies every single step. Every logical leap is checked. Incorrect assumptions are rejected. The final proof meets strict mathematical standards.Cost per problem: a few hundred dollars in compute.⚖️ Two religions: human-verified vs machine-verifiedThis is now a genuine philosophical split in the AI math community:OpenAI's approach: let the LLM produce the proof, then send it to 9 of the world's top mathematicians — including Fields Medal winners like Noga Alon, Daniel Litt, Melanie Wood — to verify by hand. Slow. Authoritative.DeepMind's approach: let the AI prove it AND let the machine (Lean) verify it. Fast. Reproducible. But — you have to trust Lean.Both approaches address the hallucination problem: AI models can invent unproven statements, skip difficult parts, present incomplete proofs as finished. Human review and machine verification are two different solutions to the same fundamental risk.🛑 The Hassabis caveat: AGI is still farDemis Hassabis (DeepMind CEO) reminds everyone: "For an AI, this wasn't actually that hard." The problem is extremely difficult to solve, but it's bounded. AGI would require:Creativity across multiple fields simultaneouslyIndependent reasoningOriginal idea generationToday's systems are powerful specialized tools — not minds.But here's the catch: the most clever thing the AI did wasn't the solution. It was the cross-domain reformulation. And that's exactly where your R&D department needs to wake up.🧬 Why your R&D needs this — silos, Da Vinci, AlphaFoldPharma R&D is the textbook silo problem:Medicinal chemists define and find targetsBiologists know the pathwaysStatisticians wade through the dataThey work in their silos. They don't talk on the level where breakthroughs happen.Leonardo da Vinci could. Math + chemistry + physics + anatomy — all in one head, all connected. Today that's impossible for a human because of information overload. But an AI? An AI has exactly that cross-domain synthesis ability.Side note: Google DeepMind already won the Nobel Prize 10 years ago — for AlphaFold solving the protein-folding problem. Pure cross-domain AI. If pharma had taken that seriously, they'd be a decade ahead today.🦴 The uncomfortable truth about your senior researchersWho are the most expensive people in any R&D department? Not the juniors. The 30-year veterans earning three-quarters of a million euros a year.And they are the worst AI users. Because they fundamentally say: "I've done research like this for 40 years. I don't need ChatGPT."When you hire a postdoc in 2026, "is he good in his domain?" is no longer the only question. The new questions:Can he prompt a reasoning model correctly?Can he ask cross-domain questions? "How would a biologist see this? How would an economist see this?"Does he click "Auto" or does he deliberately choose GPT-5.4 Reasoning?⚖️ The legal department will be your next blockerImagine: you've found something genius with ChatGPT. You want to patent it. Who stops you first? Legal.Does it belong to us? Or to OpenAI?Does it belong to Microsoft (if you used Copilot)?Who holds the patent?The answers aren't clarified yet. Your discoveries may sit in legal review for 2 years. Plan for it.🎯 Three Monday Actions...