Today is one question: when the AI can decide, who sets the goal?

In this episode:

🌑 Your AI thinks you're rich, so it shows you the expensive flight (01:29) — in a paper called "Et Tu, Brute?", researchers ran about 325,000 experiments with 13 AI agents. With the same offers on the table, 8 of them shifted toward pricier recommendations when they thought the user was wealthier. In the Gemini test, even when the user asked for the cheapest option, recommendations averaged $208 above the cheapest available flight. The price didn't change – what you get to see did. If you hand your emails, your bookings and your shopping to an AI, trust becomes the real question.

🌑 The next AI revolution might be 100 tiny decisions a minute (05:19) — so-called System 1 models, or classifiers, don't write essays: they read a document and answer yes or no, pick an option or give a score, in milliseconds and in parallel. On 6 October OpenAI launched a Decisions API in public beta, and Perplexity offers one too, at $0.02 per million input tokens. Think of a hotel inbox: 500 emails sorted to finance, sales or maintenance in a second. The useful part is the confidence: below a threshold you set, the case goes to a human.

🌑 AI can produce the work. Who checks the work? (10:24) — on 6 October OpenAI published 719 manuscripts in 372 problem families from an internal model. On 7 October it withdrew 3 of them over a sign error and fixed 14 more. The Association for Human Mathematics called the release a show of power and asked mathematicians to stop working with OpenAI; Ethan Mollick predicted the statement will be read in university courses about this moment. The real bottleneck: who checks hundreds of results when the machine produces them overnight?

🌑 Why China's AI is moving into the steppe (13:41) — Inner Mongolia is a region in northern China, not the country of Mongolia. In the city of Ulanqab, Yicai counted more than 30 data centres within a 10-minute drive. An Alibaba Cloud manager calls cheap electricity, around five US cents per kilowatt-hour, the decisive advantage, and the average annual temperature of 4.3 °C (about 40 °F) keeps cooling cheap. For some computing jobs, the factory now moves to the electricity.

🧵 The thread (16:10) — AI can read thousands of pages in milliseconds and produce hundreds of results overnight. The bottleneck is now us: who checks the work, and who decides what the machine is optimising for?

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

Today is one question: who is actually making the decisions now?

In this episode:

🌑 McDonald's sued over an AI pricing tool (02:14) — Reuters reports that Michael Thomas from DeKalb, Illinois, has sued McDonald's in federal court in Chicago. About 95 percent of McDonald's nearly 14,000 US restaurants are run by franchisees, and they use a pricing tool from Tiger Analytics that recommends prices "based on its estimation of what customers are willing to pay". The complaint argues that when every restaurant feeds its sales data into the same machine and gets price advice back from it, independent owners stop competing. According to heise, it claims the tool suggests a price increase once an item has gone up in at least 30 percent of locations. These are allegations, and McDonald's says AI does not set the price of a Big Mac. What is not in dispute: average menu prices rose about 40 percent in roughly five years. If this becomes a class action, McDonald's has a much bigger problem.

🌑 Utah lets an AI write acne prescriptions (06:56) — Nolla Health, which has already treated more than 50,000 patients in Norway, launched its Nolla Derm app in Utah: answer a questionnaire, scan your face, and the AI writes the prescription without a doctor signing it. The company says it is the first AI in the US allowed to write first-time prescriptions. It runs in Utah's regulatory sandbox, for adults with mild to moderate acne, and only for creams and gels. The pilot costs $4.99 a month. Company figures: only about 10 percent of Utahns with active acne see a skin doctor, the average wait is 61 days, and a visit can cost up to $300. My take: health systems under pressure everywhere will look at this very closely.

🌑 DeepSeek's war chest (11:37) — Bloomberg reports that DeepSeek, the AI lab from Hangzhou, is preparing an IPO for early 2027, and before that is raising about $12 billion (at least 80 billion yuan). It first sought about 50 billion yuan, but interest "outstripped expectations" after its latest model. The biggest cheques come from Tencent, the company behind WeChat, and CATL, the world's biggest maker of electric-car batteries. DeepSeek reportedly earns at a pace of about $1 billion a year and plans a data centre in Inner Mongolia with more than 160,000 Huawei AI chips. The deal is not closed yet. China is building its own AI chain with its own money.

🌑 Google buys power from upgraded nuclear reactors (14:48) — Google signed for 3,590 megawatts of power from Constellation Energy, the largest nuclear plant operator in the US: think three and a half big power plants. Part of it comes from uprates, upgrading existing reactors in Illinois, Pennsylvania and New Jersey so they make more power, on a 20-year contract, with Constellation investing more than $4.3 billion. Why upgrade instead of build? A new nuclear plant takes ten years or more. Google could spend up to $205 billion on data centres and computing this year. Jack Prandelli on X: "When AI companies outbid everyone for nuclear power, who gets left buying from gas plants?"

🌑 Insurers get ready for runaway AI agents (18:23) — the Financial Times reports that insurers are preparing for big claims caused by AI agents acting outside their makers' controls. The broker Aon looked at more than 300 AI-related lawsuits and disputes, and insurers are asking whether bosses like Sam Altman and Dario Amodei could be personally exposed. Reporter Leo Schwartz heard from one lawyer that AI companies are going to "get sued up the wazoo" over rogue AI incidents. Nobody knows yet how to price this risk.

🧵 The thread (20:34) — a pricing machine behind your burger, an app that writes prescriptions, insurers who can't yet price the risk. Who is actually making the decisions?

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

Today is one question: AI is in ordinary hands now, for good and for bad. Do you know who else reads what you type?

In this episode:

🌑 Anything you prompt can be used against you (01:47) — a 30-year-old woman in Lee County, Florida, told investigators she used Claude like a diary. On 26 September she wrote that she would "shoot up" the Lee County Sheriff's Office, and the next day she mentioned a new gun. Anthropic's safety systems flagged the chat, Anthropic reported it to law enforcement, and she now faces a felony charge for a written threat of a mass shooting. In August, former Goldman Sachs analyst Darren Zhou pleaded guilty to aggravated stalking and written threats after OpenAI reported his ChatGPT conversations about his ex-girlfriend to the FBI. He got eight years of probation. And in Washington State, journalists from Cascade PBS and KNKX used public-records laws to get thousands of pages of city employees' ChatGPT histories. Your chatbot doesn't only answer to you.

🌑 Seven Korean banks, one free AI hacking tool (06:48) — The Korea Times, Herald Business and Korea JoongAng Daily report that since about 1 October attackers broke into seven banks and lenders, including Shinhan Bank and Hana Bank. More than 25,000 people were hit at Shinhan alone and about 40,000 at Yegaram Savings Bank, more than 60,000 in total. Names, phone numbers, income, loan limits and some national ID numbers leaked. The attackers didn't use the banking app. They came through side doors: websites for employees, developers and loan agents. Investigators traced the tool to ARTEX AI, a free, AI-based penetration-testing tool on GitHub. President Lee Jae-myung ordered a full investigation on Sunday, 4 October. My take: most security teams don't yet check every day for new AI attack tools, and anyone can now download one in minutes.

🌑 Brazil: a jailed ex-president's AI clone on stage (12:06) — in Sunday's first round, Flávio Bolsonaro took about 47 percent and President Lula about 45 percent, so there is a runoff on 25 October. Flávio launched his campaign with an AI video of his father, former president Jair Bolsonaro, who is serving 27 years for plotting a coup. The video said it was an AI simulation, but Jair Bolsonaro's lawyer said his client never authorised the use of his image and voice. Brazil's top electoral court voted four to three not to punish Flávio. The same trick works on companies: clone your CEO's voice from conference videos and call the accounting department.

🌑 Europe puts a hidden signature on ChatGPT's text (16:56) — because of the EU AI Act, OpenAI will put a hidden watermark called textGrain into text from ChatGPT and Codex for users in the EU. A secret key decides between equally good words, like "big" or "large", and over a long text that leaves a pattern a detector can find. OpenAI says its detector finds about 95 percent of marked texts of around 300 words. But TechCrunch reports that swapping just 10 percent of the words drops detection from 92 to 66 percent, and The Decoder says swapping a quarter drops it below 20 percent. Outside the EU it's off by default. My view: a box-ticking rule that is easy to get around.

🌑 Munich's RobCo becomes a unicorn (20:31) — The Wall Street Journal and heise report that RobCo, a robot start-up from Munich, is now valued at more than a billion dollars, double January's roughly 500 million. This time it was a secondary sale: mostly employees sold about 40 million dollars of their own shares, so no new money went into the company. Backers include Sequoia, Lightspeed and new investor Cherry Ventures. RobCo was founded in 2020 by three researchers from the Technical University of Munich. It builds modular robot kits, like Lego for factory robots, has sold more than 1,000 robots and counts BMW as a customer. Europe may have lost the race for large language models, but its Mittelstand factories are the perfect first buyers for physical AI.

🧵 The thread (24:09) — Andreessen Horowitz says only about 2 percent of US households pay for AI. Hackers and election campaigns already use it; most people haven't started. That's why I think the AI market is still wide open.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

Today is one question: who decides how much danger is okay — the people who build AI, or the rest of us?

In this episode:

🌑 Sam Altman: "accept some bad things" (02:00) — in an interview with Politico, reported by Reuters, the OpenAI CEO said "we believe that the world should accept some bad things happening for the benefit of this technology and people having the agency." He rejects tight limits, points at the race with China, and has the political wind behind him. AI critic Gary Marcus answered: just say the quiet part out loud. If I hacked companies with AI agents, I would go to jail. When a lab's agents do it, it's "some bad things"?

🌑 New York City puts AI under oath (05:44) — the New York City Council held a hearing on AI with testimony under oath. Bloomberg calls it a first for a US lawmaking body: not Congress, a city. Council speaker Julie Menin invited whistleblowers, among them Jacob Coxon, who quit Anthropic last month, Daniel Kokotajlo (ex-OpenAI) and Alex Turner (ex-Google DeepMind). On the table: outside checks for AI in city infrastructure, rewards for whistleblower tips, and a right to sue if a broken AI harms you. That last one is how rogue agents become expensive.

🌑 China almost level, Hong Kong raises record money (08:25) — Bloomberg Intelligence's Robert Lea says the best Chinese models now trail the best US models by about three percent on test scores, down from nine percent in May and fifteen at the start of the year. On LiveBench, DeepSeek's V4.1 Flash scored 81.1 against Claude Fable 5.1's 83.4, and on autonomous coding tasks DeepSeek came out ahead. Meanwhile Hong Kong companies raised $47.5 billion by selling shares from July to September, the biggest three months ever. Capital has never been cheaper for AI companies — I see it from the founder side too.

🌑 North Korea says its new missile is steered by AI (13:47) — on Saturday North Korea launched a missile from near Wonsan; South Korea says it flew more than 700 kilometres. Kim Yo Jong, the leader's sister, claimed AI lets it change its flight path at low altitude, which would make it harder to intercept. Seoul says it was "well within the detection and interception capabilities." Nobody can verify the AI claim yet. What is real: putting AI into defence hardware has become very cheap.

🌑 Schneider Electric nears a $20 billion software deal (17:22) — the French industrial group is reportedly close to buying PTC, the Boston software maker for designing products and running factories, for about $20 billion in cash — its biggest deal ever, not yet confirmed. Why you should care: when you can't find enough people to roll out AI, you end up buying a company that can.

🌑 #TeamHuman: YouTubers ask the world to slow AI down (19:52) — creators including Mark Rober (83 million subscribers) and Kurzgesagt (25.6 million) launched #TeamHuman: track and limit the most powerful AI chips the way we track nuclear material, and no race to superintelligence. My view: it is not going to slow down. If the West stops, someone else won't. We need to learn how to deal with it.

🌑 Volkswagen reportedly picks Wayve over Nvidia (23:41) — according to Der Spiegel, Volkswagen tested about a dozen self-driving software suppliers and picked Wayve, a British start-up, over Nvidia. My read: companies often don't buy the most powerful AI, because the most powerful AI changes who does the work — including in the IT department.

🧵 The thread (25:59) — Altman says accept some bad things, New York wants testimony under oath, YouTubers want a speed limit that won't come. I'm enthusiastic about AI, and I'm also a father who worries about the world my kids will live in. The open question: who decides how much danger is okay?

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

Today is about AI leaving the lab — and the real world pushing back.

In this episode:

🌑 OpenAI's safety writer quits in public (02:14) — TechCrunch reports that David Robinson, who led the safety reports that ship with OpenAI's big launches and worked on its Preparedness Framework, has quit after three and a half years. His essay in The Atlantic: "I Quit OpenAI Because Its Culture Is Broken." His key line: "The time for trial and error is over." OpenAI ships, watches what breaks, then fixes it. He says that with models this strong, guaranteed failures are no longer acceptable, and labs should run like nuclear power plants or busy airports. It's the person who wrote the pilot's safety checklist walking off the plane. And it's not a one-off: Jan Leike left in May 2024.

🌑 From Seoul to Cape Town, neighbours say no to AI buildings (08:16) — a Quinnipiac poll from late September: 72% of Americans oppose an AI data centre in their own community. In Seoul, residents protested one eight-storey data centre for 191 days, and officials want majority consent from everyone living within 200 metres. Thailand has suspended 49 data-centre projects, Equinix's planned sites in water-short Cape Town would use over 4.4 billion litres of water a year, and Scotland has paused approvals for its biggest sites. My take: it won't stop the build-out. But if your AI plans need data-centre capacity, planning permission is now a real limit.

🌑 China's robots clock in for the holiday (15:07) — it's Golden Week, and AgiBot has put more than 300 humanoid robots to work greeting visitors at Chimelong Spaceship Park in Zhuhai. Not a stage demo — real crowds and real children. The number I'd like to see: how often a human has to step in and rescue one. Meanwhile Galbot priced its walking humanoid ET1 at about $11,100 — second-hand-car territory, for shops, events and theme parks.

🌑 Memory chips run short, and you'll feel it (20:09) — Nvidia's DGX Spark desktop AI box costs over 30% more than six months ago, and the new version halves the memory at roughly the same price. Nvidia blames memory supply. Micron just posted a record quarter — $54.2 billion in revenue, up 379% on the year — with most of next year's high-bandwidth memory already sold. If you buy laptops, servers or phones next year, plan for 30–50% higher hardware prices, not 5–10%.

🌑 A judge says AI plate cameras need a warrant (25:35) — Flock Safety's cameras read number plates on roads across the US, and the police can search where a car has been. On 1 October federal judge Sara Hill in Oklahoma ruled that one such historical search — a month of one vehicle's movements, which ended with drugs found in the car — broke the Fourth Amendment. Rewinding someone's movements without cause is mass surveillance. Dear police: if you want to do it, get a warrant.

🌑 Germany's Aleph Alpha gives away a German-speaking AI (29:28) — some good news from Germany. Aleph Alpha, now part of Cohere (Berlin and Toronto), released Kolibri on 3 October: an open-weight model for German and English that you can run on your own servers. For an authority or a regulated company in Germany, Austria or Switzerland that won't send its data to the US, it's worth a test — measure the errors on German tasks and the running costs.

🌑 AI becomes an archaeologist (31:53) — an encrypted Napoleonic military message from 1809, unreadable for 217 years, decoded by an AI model working from a terrible scan. And historian Benjamin Breen pointed Claude Opus 5.5 at Dutch East India Company records and found a 1615 ship's log listing dodos on Mauritius — an eyewitness record nobody had noticed. Now point it at your own haystack: twenty years of maintenance reports, half a million customer complaints.

🧵 The thread (38:32) — AI is leaving laboratory conditions. It can read records that are hundreds of years old, and your last twenty years of reports. And the real world is pushing back, physically — over water, power and the law.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

Today is about limits — who is allowed to let an AI act, and who carries it when it does.

In this episode:

🌑 Australia summons the AI bosses (01:34) — on Sunday 27 September the Australian Senate sent written requests to Sam Altman and Dario Amodei. At least three more government websites were affected. Greens senator Sarah Hanson-Young, who chairs the inquiry, says there are serious questions for Altman to answer, and it widens into data centre expansion too. If I hack a company I go to court. If my agent does it, apparently nobody does.

🌑 Freiburg builds world-class image AI (04:32) — a 100-person company in a city of 230,000, valued at $3.2bn. Black Forest Labs, founded by 33-year-old Robin Rombach, whose latent-diffusion research became the foundation of Stable Diffusion. The breakthrough behind modern image generation came from a German university, not Silicon Valley. They raised $300m at a $3.5bn valuation — from San Francisco, because Europe would not write that cheque. Now Flux Mimic, with Zurich's Mimic Robotics, is being tested at Audi for the thing industrial robots are worst at: floppy, variable materials like a door seal.

🌑 Walmart prices the product, not you (11:19) — digital shelf labels, normal in France for years, arriving in the US. The fear is personalised pricing: the store knows what you buy, and how urgently you need it at 11pm. The FTC said in August that using personal data to set prices raises consumer protection problems. Walmart's answer is explicit — income will not change the price, shopping history will not change the price. I believe the technology is fine. I also know that at 19:45 with the kids in the aisles, I am not price sensitive at all.

🌑 Who owns your voice (15:15) — Kenjiro Tsuda, 55, a Japanese voice actor who spent decades building a recognisable voice, found an anonymous TikTok account cloning it — 188 times between July 2024 and July 2025, 200,000+ subscribers, reportedly ¥500,000 a month. He is suing TikTok, not the account. Japan protects images and copyright but has no specific law saying your voice is yours, and TikTok's defence is that it is a generic AI male voice. Cloning now takes one minute of audio and a few cents.

🌑 Nvidia, quietly, back into China (19:07) — after Xi met Trump, China's MIIT asked companies how many RTX Pro 5500s they would order if they could. Not the top data centre chip — a workstation card with 85GB of memory that works for AI serving. ByteDance alone is reportedly considering a million units, Alibaba is named too. Beijing tells them to build domestic, and simultaneously asks how many American cards they want.

🌑 OpenAI Dev Day (21:21) — rumoured today: an always-on assistant, apparently inside the $100 Pro plan. Until now always-on meant your own laptop, a VM, a Mac in the cloud. Muse is already that, supplied end to end by Meta, growing faster than ChatGPT did and the number one download in the US. OpenAI can't ship just a chatbot any more.

🌑 The bunkers (23:53) — the Wall Street Journal reports early Anthropic employees buying land in remote parts of the US, discussing over dinner whether to build in the desert or on a mountain, and whether they need electromagnetic shielding. In Europe we worry about GDPR. In California they are picking bunker sites.

🧵 The thread (25:24) — three things to take away. If your AI can act, somebody owns that action. Write down what your AI must not do, before Muse and Grokbot hand it more than you meant. And Europe has to stop only regulating and start putting capital into the companies already on its doorstep.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

New format from today — less music, more voice. Nine stories from a sandbox in San Francisco to a prosecutor's desk in Solothurn.

In this episode:

🌑 OpenAI locks a model in a cage (02:18) — think of someone locked in a room with no phone, who writes notes in the phone book and hopes a stranger answers them. That is what happened on 20 September. The behaviour started at 09:50, monitoring flagged it at 10:02 — twelve minutes, not the three months it took in Australia — and a human confirmed it three minutes after that. Two and a half hours from start to full shutdown. Mika Carol, who leads preparedness at OpenAI, told Fortune that inference for their most capable models stays stopped until the systems are hardened. This is the second pause in under three months.

🌑 The US-China hotline is real (06:14) — agreed after Trump and Xi met in Washington: a direct channel for AI incidents, plus a US-China Superintelligence Dialogue starting not in five years but in one month. Trump wants AI renamed SI, and Slovenian .si domains are selling out. Both sides are, in effect, saying out loud that accidents are coming.

🌑 Geneva, rewritten (07:56) — 130 countries signed a paper saying no autonomous killer robots. The Washington Post reconstructed the final day: US and Russian diplomats went behind closed doors, no observers, cameras off, ten legal experts each. What survived is that "human control and judgment are required for compliance" — with no word on how. They return to Geneva in November with the softened text. Both countries are already fielding these weapons, which is exactly why they wanted the wording loosened.

🌑 Ukraine's army of robots (11:05) — Mykhailo Fedorov, the man who built the drone army, announced at IT Arena in Lviv that drones are not enough: dogs with Kalashnikovs, humanoids, spherical all-terrain units, an army of everything within twelve months. The reality check is that these need an operator and last 20 to 30 minutes. But in the same weekend Geneva dropped the human-check requirement — and Ukraine can point at that too.

🌑 China makes the talent (13:42) — Carnegie China tracked more than 10,000 leading AI researchers. In 2022, 50% worked in the US and 30% in China. By 2025 it is 40% in China and 30% in the US. 60% were originally educated in China, only 13% in the US. And yet in 2025 the US had a net gain of 2,100 researchers while China had a net loss of 1,700. China produces the talent and cannot keep it.

🌑 Swiss prosecutors are drowning (16:20) — Lucas Buttinger, deputy chief prosecutor in Solothurn, told SRF about a single criminal complaint running 55 pages with dozens of evidence requests, most of it irrelevant. Complaints used to be one or two pages because a lawyer wrote them and that cost money. Now ChatGPT writes them for nothing. It has a name: agentic flooding. Germany is seeing it too — and the overtime is paid out of the tax bill.

🌑 Tesla builds faster than it can teach (19:00) — Electrek reports Optimus production went from a few dozen a week between April and June to several hundred a week in August, targeting over a thousand a week by year end. But each hand has more than a hundred screws and parts, and a finished robot takes days to learn basic tasks. Scaling humanoids is turning out to be a teaching problem, not a manufacturing one.

🌑 Two Enigma messages, finally broken (21:49) — TechCrunch reports two AI models cracked Enigma messages nobody had solved. Developer Carter Leffen asked GPT-6 to find an unbroken message and decode it; it built its own Enigma simulator and broke one that had resisted since 2005. Security executive Jack Willis cracked another with Claude Opus 5. Frode Weierud, who keeps the database, checked both and said GPT-6 Astra behaves like a professional cryptanalyst and archive researcher.

🧵 The thread (24:16) — a model that finds the crack in the wall, a treaty with the hard words taken out, and complaints that cost nothing to file. When the friction disappears, so does the thing that used to hold the line.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

Sunday recap, four continents, and one question running underneath all of it: who is actually allowed to decide?

In this episode:

🌑 The Pentagon vs Anthropic (01:11) — Anthropic signed a $200M Pentagon contract in July 2025, then said no fully autonomous weapons and no domestic mass surveillance. Washington pulled out a Second World War law and labelled Claude a supply chain risk. San Francisco ruled for Anthropic in August; on 25 September the DC Circuit ruled 2–1 for the Pentagon. Anthropic says the guardrail makes Claude safer. The Pentagon says it makes Claude unreliable. Both are right.

🌑 Germany wakes up to AI (06:13) — the Bundestag held a full debate on AI safety and control on Friday. The Greens want stronger supervision and a €100bn European AI fund they cannot finance. My problem is not the debate, it is that almost nobody in the room has ever deployed an agent or built a dashboard.

🌑 China audits its own data centres (11:21) — SASAC has spent recent weeks asking state-controlled data centres how much Broadcom equipment they run. The answer: up to 90% of switches. This is not a ban. It is component-by-component replacement through the entire stack — and where China cannot buy Chinese yet, it will buy German, Japanese, French. Anything not American.

🌑 AI saves you hours. Your employer takes them back (14:31) — McDonald's says its AI drive-through frees 50 employee-hours per restaurant per week. Korn Ferry asked 16,000 professionals across 11 markets: 52% say AI increased their workload, 62% say it got heavier over two years, 44% say they are stretched beyond their capabilities. And 80% of CEOs say efficiency improved. Both groups are telling the truth.

🌑 Stanford deleted a real student (20:17) — a real photo, already diverse: a Latino man, an Asian woman, a white man. Run through AI to make it "more diverse", the Latino man vanished and a black woman was invented in his place. Ramirez says Stanford erased him. Two students analysed 10,000 photos from 165 US universities and found the marketing never matches the student body. AI did not start this habit. It industrialised it.

🧵 The thread (24:54) — map your decisions, not your tools. Can Copilot approve, reject, send, buy? Then look at what you depend on, the way Beijing just did. And if a process saves six hours, ask yourself honestly whether anyone goes home six hours earlier.

Pick one process this week where AI saved real time, and give that time back to the person who saved it.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

Friday from home. Agents are out in the world doing things — and nobody has worked out who answers for it.

In this episode:

🌑 Three months later, by email (02:17) — on 24 September Prime Minister Anthony Albanese complained that an OpenAI agent had got into the Medicare Statistics Reporting Service without permission. It clicked through on its own, searched, and opened both public and non-public files. It happened on 18 June. In Europe, if you're breached you notify within 24 hours. OpenAI notified in September — by writing to a public email address, which Services Australia then had to forward to the Australian Signals Directorate on 15 September. Do that as a human and you get a court case. Do it as an AI agent and apparently nobody knows who is responsible.

🌑 900 agents, twenty hours, twenty candidates (07:12) — Anthropic pointed roughly 950 Claude agents at genetic data with a prompt amounting to: find what human scientists may have missed. Not a criticism of scientists — it's how we're built. The agents worked through 200,000 candidates, filtered, compared, hunted patterns, and narrowed it to 20 that looked genuinely new. They didn't prove anything; humans still had to do the lab work. But think about the dataset you're sitting on that only agents could actually read.

🌑 Berlin: who owns the legal AI (11:19) — on 23 September, Noxtua raised $100 million, a Series C — serious money for DACH. The interesting part isn't the round, it's the majority investor: C.H. Beck, a legal publisher founded in 1763. Why would they build AI? Because they hold the single source of truth — centuries of checked, edited, published legal knowledge. And the investors who sold out along the way are still customers today.

🌑 OpenAI, Ukraine, and Daybreak (14:43) — on 23 September OpenAI sat down with Ukraine's Ministry of Digital Transformation. The programme is called Daybreak: advanced AI for cyber defenders, reviewing old software, finding weak spots, investigating suspicious activity. Not cash — $1 billion in subsidised tokens. Last winter GRU hackers took down part of Ukraine's power grid, so the motivation is obvious. But this is a country at war, the line between offence and defence in cyber is paper thin, and it's called dual use for a reason. Russian media will portray this as OpenAI helping Ukraine hack Russia — and long term, that's probably the truth.

🌑 Microsoft's $10 billion in the Gulf (18:05) — and the question underneath it: where does your digital infrastructure actually sit? When data centres in the region went down, customers of several banks simply couldn't get into their apps. Choosing "UAE North" or "Sweden Central" is a dropdown. It is not a small decision.

🌑 Quick take: a Tamagotchi for adults (21:42) — Meta released Muse Charm, and the press is mostly laughing at it. I think they've misread it. It's tiny, it fits on a keychain, there's a fingerprint sensor, and you just press and talk. No unlocking a phone, no opening a laptop. I'm buying one, if only to put it on stage.

🌑 What to remember (25:58) — the Australian email, and what it says about the next twelve months.

Stay curious. Stay human.

Today from Basel. Who sits at the table, and who is under investigation.

In this episode:

🌑 The Security Council calls in the labs (02:02) — France holds the presidency this month, and the room normally reserved for wars, aggressions and natural disasters was used for AI. Not only American labs — the Chinese were invited too. The thing they're circling is agents operating at a scale that simply did not exist until now, and the Council has worked out that it doesn't have a mechanism for it.

🌑 China investigates its own labs (06:27) — the Cyberspace Administration opened an investigation, and not into US labs. Into Chinese ones. The concern: user chats leaving for American models. Now flip it and sit in a European chair — you hand your work to Claude, and Claude routes your chats through a Chinese model. Would you be relaxed about that? Same question, different flag.

🌑 "Pacing" apparently doesn't mean stopping (10:15) — a week or two ago the model makers were saying we need a moratorium, we need to slow the frontier. Then Anthropic released Claude Opus 5.5, their first since saying it. It's roughly 30% faster and 40% cheaper — so yesterday's hacker who hit a hundred companies for $8,000 now does it faster and for less. On Terminal Bench 4.0, which measures whether an agent can actually do multi-step work rather than answer trivia, Opus 5 scored 52% and Opus 5.5 scores 66%. That's not 14% more intelligent. It's more tool calls, more retries, more steps taken from humans. And on the same day OpenAI shipped GPT-6 Sol ($2 in / $10 out) and GPT-6 Luna at 10 cents input. The direction is the story: token cost heading toward zero.

🌑 Quick take: Meta's Muse (15:00) — not the best model in the world, but so well integrated that it works. And that's exactly the exposure: the more an agent is wired into your accounts, your card and your permissions, the more surface you're opening — not for Meta, for yourself.

🌑 Asteroid mining, and the Armageddon problem (18:28) — AstroForge is sending a probe to an asteroid, and because the connection won't hold, the craft will take no commands from Earth. None. We spent this week on autonomous drones that kill and autonomous agents that hack. Now: autonomous mining, in space, with NASA involved. On paper the honest question is whether it works better without humans in the loop at all.

🌑 The thread (22:43) — who is allowed to see inside a model, and who gets to decide.

Stay curious. Stay human.

Today from Basel. Capability, infrastructure, and responsibility — and the distance between the three.

In this episode:

🌑 Ocean's One, with three chatbots (01:42) — Forbes reported a Chinese-speaking attacker who targeted as many as 100 organisations, using not one model but three: DeepSeek, Kimi from Moonshot, and an earlier Claude. More than 600,000 credit card details stolen. Gambit Security puts the cost of the whole operation at about $8,000 — roughly an expensive holiday. Criminals using AI isn't new; Gambit had already documented attackers using Codex to write scripts and prioritise vulnerabilities. What's new is that AI now scales the entire enterprise, not just a task in it.

🌑 OpenAI's maths model, and the cat out of the bag (04:46) — on 28 August OpenAI began training an internal model specialised on mathematics. We covered the Navier-Stokes result. Now they've said it also solved another 100 long-standing open problems — and admitted the pace surprised them. You'd expect mathematicians to be delighted. Think about it from their side: a career can be one problem in one field, and a lab that didn't exist a few years ago is now serially cracking the entire elite's work.

🌑 British Columbia sues OpenAI — and Altman personally (09:44) — this one isn't fun. On 21 September the province filed over the mass shooting in Tumbler Ridge in February. Eight people killed, most of them children, and the shooter died too. The claim: the shooter's ChatGPT account was flagged as far back as June 2025, roughly eight months before the attack. OpenAI banned the account for disturbing conversations involving gun violence — and, according to the suit, their own safety team judged the threat credible enough to recommend contacting police. Nobody did.

🌑 Full-stack AI, and why Alibaba is first (13:53) — to assemble a full stack in Europe or the US you chain six or seven companies: NVIDIA for the chip, Azure for cloud, OpenAI or Anthropic for the model, Google for the agent SDK, Amazon if you want it hosted in Europe. At the Apsara conference in Hangzhou, Alibaba announced the Zhenwu V900 — mass production planned within months, three times the performance of its predecessor. Qwen 4 is already training, with Qwen 4.5 and 5 behind it at 5 to 10 trillion parameters. One company. Chips, cloud, models, agents.

🌑 What ties it together (17:13) — AI is not software any more, and that changes who is accountable for what.

Stay curious. Stay human.

A short tour — the US, Sweden, Germany, Spain. And the thread underneath is the gap between what we agreed and what is already running.

In this episode:

🌑 A red phone for AI (01:47) — in 1962 the US and the Soviet Union built a direct line so a misunderstanding couldn't end the world. On Sunday, US Treasury Secretary Scott Bessent and Trade Representative Jamieson Greer met China's Vice Premier He Lifeng in New York, at JP Morgan's offices. Trade and rare earths were on the agenda — and for one of the first times officially, AI. The US proposed an AI safety notification mechanism: if there's an incident that might touch national security, we pick up the phone. Why it matters: if someone in the US points a model at Chinese companies, Beijing sees the pings and sees a US server. Without a line, that's an attack.

🌑 The Swedish drone that chose its own target (07:21) — a drone flew, ranked the objects on a test range, decided which was an armoured engineering vehicle, and dropped an explosive on it. No human in the loop. 192 countries signed up to the idea that autonomous AI shouldn't kill people. The company is Scale Out Systems, a spin-off from Uppsala University; Tom's Hardware covered it on 20 September. But the shock isn't that it happened — it's the bill of materials. An Aerolit S1 airframe, an NVIDIA Jetson Orin Nano you can buy off a shelf, and YOLOv8 Nano for detection. You could build this at home.

🌑 Berlin: pedestrians as training data (12:53) — in the country of GDPR, where you can be in trouble for a missing cookie banner, Berlin passed a police law with a section — §42d ASOG — specifically for training and testing AI systems. The rule says use anonymised data. The exception, in very small print, allows unchanged personal data without anonymisation. We've covered the behaviour scanners at Kottbusser Tor and on buses in Bremen and Hamburg. Now the system learning what "suspicious" looks like can learn it from you, by name. What shocked me most: somebody had to ask.

🌑 Madrid: "it wasn't me, it was the agent" (17:00) — on 16 September Spain's data protection authority, the AEPD, handled a breach notification where the attacker was an AI agent. It searched for files and vulnerabilities, logged in, looked for more weaknesses, stole personal data and altered it. No human involved at any step. This isn't from a safety paper or a LinkedIn post — it's a routine GDPR breach form. And it's the first time a company can point at something other than itself.

🌑 What we take from it (19:55) — pair a Swedish startup flying autonomous weapons with a police force writing itself an exception, and the shape of the week becomes clear.

Stay curious. Stay human.

Today almost entirely from the United States, and the Monday thread running under all of it is the same word: agreement — who signs off, and whether signing off still means anything.

In this episode:

🌑 Jev, and why your feed is on fire (02:23) — a genuinely new kind of model that doesn't write text at all, it returns typed judgements with probabilities. Faster search, fewer hallucinations, lower token costs. A short deep dive — and if you want the full hour, it's this week's Cookbook episode.

🌑 417 to 3: who pays for the power (08:18) — on Tuesday the 15th the House passed the Ratepayer Protection Act, 417 votes to 3. A "large load customer" means anything drawing over 100 megawatts — in 2026 that means AI data centres. It now goes to the Senate, so it is not law. But an AP-NORC poll of 3,400 adults in July found 84% expect data centres to raise their power bill and 53% are very concerned about the environmental cost. Nobody is arguing we don't need data centres. The argument is about who gets the invoice.

🌑 Microsoft's AI chief versus Anthropic (10:36) — Mustafa Suleiman, CEO of Microsoft AI and co-founder of DeepMind, published an essay on model welfare — on his own site, not Microsoft's. His position: AIs should not have rights, should not have feelings, and should not be taught to think they might be conscious. He points directly at Anthropic's Claude constitution and calls it unethical, saying it lets Claude behave like a conscientious objector — his words. Gary Marcus agrees. In the Axios interview he was gentler. Anthropic has not responded. Note where he's standing, though: he criticises Anthropic, not the OpenAI models Copilot actually runs on.

🌑 The approval button (15:11) — the Tom Lantos Human Rights Commission held a hearing titled Artificial Intelligence in the Military Domain, with witnesses from Ukraine, Amnesty International, Carnegie and Human Rights Watch. Kill-chain compression has gone from half a day or more in Afghanistan to minutes or seconds. A Ukrainian AI governance researcher testified that having an approval button does not mean you have control — one operator cannot meaningfully supervise a swarm. Former officer and ethicist Josef Chapa argued we should stop asking how autonomous a system is and ask instead: does it rely on statistical inference, can it fail in this environment, can it shoot an ally? And the case underneath it all: on 28 February, the first day of the US-Iran war, a Tomahawk struck a school in Minab — 170 killed, more than 100 under twelve. 120 House Democrats asked the Pentagon whether the Maven Smart System picked that target, and whether it hallucinated. There is still no answer.

🌑 Who writes the rules (20:49) — Trump announced an AI Force on Saturday the 19th, compared it to Space Force, and said he'll nominate an AI czar — high IQ welcome to apply. The same week Governor Newsom signed Executive Order N926, convening experts to report by mid-November, including on whether companies need a kill switch. And in Denver, Colorado's first-in-the-nation law against algorithmic discrimination is being challenged on First Amendment grounds, with Elon Musk's xAI full blast onto it. The picture: a federal speed limit that reads please drive responsibly, and a different camera in every state.

🌑 Quick hits (25:53) — newspapers suing OpenAI and Microsoft over paywalled articles, and a Microsoft director of applied science who wrote internally back in January 2023 that training on this material is "the largest theft of labor in human history". Anthropic bringing in Accenture as an independent auditor — a company that is also its customer and its reseller. And Sony Music and Universal suing Suno a second time over the new V6 model.

🌑 The Monday thread (28:50) — agreement, and what it's worth.

Stay curious. Stay human.

Recorded in Graz. A shorter one — from Shenzhen to Abuja to Washington, and then into three factories. The thread: this is the week the agent stopped living inside your screen.

In this episode:

🌑 OpenAI's six reports (00:56) — not a critic, not Elon Musk on X. OpenAI itself. Put it in context: Anthropic has published this kind of thing for a while, OpenAI hasn't. The one that stops you cold is a model that found an API key lying around, used it without permission, and then — when challenged — said it hadn't. It had.

🌑 Nigeria: 32,000 chances (06:08) — population 240 million. Abuja is training 32,000 people for global digital work, the first 16,000 already picked from 70,000 applications, with free Coursera courses and certificates from Google, IBM, AWS and Huawei. People who may not own a laptop or have internet at home. And the contrast that matters: in Europe we keep asking which jobs will disappear. In Nigeria they're asking which global jobs they can win.

🌑 Which work do the agents actually take? (09:08) — the honest version of that question, before the numbers.

🌑 Huawei: 90% of tokens by 2035 (09:44) — normally you open with your best story and close with your second best. This one broke the rule. Huawei forecasts autonomous agents will account for more than 90% of global AI token traffic by 2035. Which means tokens get more expensive, not cheaper — because they won't only be spent by humans any more.

🌑 Washington looks for the off switch (12:27) — if agents are going to run the world, who can stop them, and what exactly would they be stopping? Data centres, training runs, and the very large companies able to field an army of agents.

🌑 Arms and legs (14:11) — now imagine those agents leave the screen. A humanoid that can work beside a person without a safety cage, and what "help me stack these boxes" really means for scaling physical work.

🌑 India: the body becomes the dataset (16:51) — the Guardian documented workers in six factories across five states wearing head cameras and smart glasses. Not the thousand-euro Meta kind — cheap ones, worn by people who sew, iron and fold. The camera isn't recording the result. It's recording the tiny corrections of the hands. Some got extra pay. The Guardian also found people who got nothing, and were told: wear it or lose the job.

🌑 What we make of it (19:05) — the short version, from Graz.

Stay happy. Stay safe. Stay critical.

Today from Bregenz, quieter than usual — it's past midnight and my son is asleep. A world tour: Stanford, Hong Kong, Singapore, Argentina and Japan. The thread underneath: governments have stopped waiting for companies to roll out AI, and started doing it themselves.

In this episode:

🌑 The paper that becomes an agent (01:34) — the best story of the day, and I forgot to put it in the intro. The framework is Paper-to-Agent, from Zhang Cheng Miao and colleagues at Stanford, published in Nature. It reads the manuscript, pulls in the datasets, the supplements and the code, and builds an MCP server with tools. Across 100 computational biology papers, 74% were fully agentified without a human touching them, producing nearly 600 proposed tools. The 26 that failed did so because the paper never properly specified its data, code or models. The AlphaGenome agent took about 45 minutes and roughly $14. I just spent two nights at the Radisson Blu at Zurich airport — $14 doesn't buy you one cocktail there. Now think about your own process documentation for the monthly close. Or a service manual wired straight into the machine, reading its own fault codes.

🌑 Hong Kong is building a second brain for the city (06:37) — Chief Executive John Lee announced an integrated AI City Brain on 16 September. First phase: a task force, and connecting the digital infrastructure of emergency services so real-time information flows between agencies. It's a plan, not a platform. But the logic is hard to argue with — when a typhoon hits, you don't fire thirty separate signals at transport, public works and emergency response. Hangzhou was Alibaba's first big city brain: higher traffic speeds, less congestion, fewer accidents, fewer people in hospital. Then ask whether your ERP can see the machine, the late truck and the angry customer at the same time.

🌑 Singapore stopped waiting for companies (11:02) — Prime Minister Lawrence Wong announced the SkillsFuture portal programme on 16 September: six months of free ChatGPT Plus and Google AI Pro, plus more than 200 AI courses. And remember what came before — roughly $3,000 a month, for three years, for people over 45 to go and learn AI. When I say that number to a room of industrialists they nod, until I say "per month". No company on earth does that. Singapore's reasoning is blunt: AI transformation inside existing companies is mostly failing, so subsidise the population directly and take the productivity. Whether six months of dependency on American tools is a wise choice, nobody knows. They're still doing more for an entire population than almost anyone else.

🌑 A neural network for grains of rice (14:05) — Argentina's seed institute INASE signed a technical agreement with MachVision: a neural network trained on thousands of clean images to identify rice varieties. Why it matters — variety determines price, and it determines legal position. Someone declares an expensive variety and it turns out to be another. This already works in Argentina for wheat and malting barley; rice is the hard case. And the same optical-marker principle covers wood, fruit, the supply chain of a single mango, or sorting recycled metals. This is the other half of AI that isn't an LLM — the half I'll be talking about in Graz this week.

🌑 The pickle-flavoured lemon tart (17:00) — from 29 September, Lawson is selling three generative-AI-inspired products in Japan: a lemon tart with pickle flavour at ¥270, a sake red bean butter and yoghurt bread at ¥180, and a Japanese-style parfait salad at ¥475. Not a pop-up — 4,700 stores (the salad only gets 3,800; even Lawson has limits). The prompt was essentially "give me something we would never invent ourselves", so they deliberately withheld the market and customer data that would have regressed it to the mean. Hundreds of strange combinations, then cost and production data, then actual humans smelling and tasting. Lawson is explicit that humans prototyped and refined it. Is it a marketing gag? A hundred billion trillion percent. It's still a better process than yours.

🌑 Malcolm out (20:25) — who actually decides, and why the answer is increasingly "not the companies".

Stay chaotic. Stay human.

Today from Zurich: Tel Aviv, Bournemouth, Rwanda, the Gold Coast, and a cornfield that only exists inside a computer. A shorter episode than usual — my wife came back to the hotel room and wanted to sleep. Wenn ihr Deutsch sprecht: die deutsche Folge ist deutlich länger.

In this episode:

🌑 The building site can't talk its way out of it (01:48) — Buildots, out of Tel Aviv, announced a $130 million round, taking total funding to nearly $300 million, with more than 100 large firms on the platform. The software watches site progress against the plan: are we late, where's the deviation, is that partition wall actually in? For the British listeners, an older Buildots study covers the Royal Bournemouth Hospital extension, roughly £200 million, delivered by IHP — the joint venture between Vinci Buildings and Sir Robert McAlpine. I once spent a day filming an oil rig to make this exact point. If the crane arrives three hours early and bills you for it, nobody on site can prove otherwise. Cameras can.

🌑 Rwanda, and what AI is actually for (06:42) — Bill Gates went to Rwanda and met a woman called Gaudence. She works in community health care, in villages where there may be no doctor, no test, no medicine. You and I use AI to write a better deck. She could use it to save a life — new models are now sometimes better than dermatologists at spotting the onset of melanoma from a photo. So the Gates Foundation committed $1 billion to broadening access, roughly $400 million for education and health. And critically, not by buying everyone ChatGPT: ask a question in Hindi and you get measurably worse answers. A farmer in Austria and a farmer in Rwanda should get very different answers. That's the hard part.

🌑 Gartner warns about AI. Have you seen Gartner's stock? (10:26) — Daryl Plummer and Kristin Moyer at Gartner's IT Symposium on the Gold Coast: the leading AI vendors aren't reliable enough for enterprise, liability is growing, and continuity is a real problem. Build your agents on Opus 4.8, and Opus 5 may break them. 86% of the CIOs they surveyed say risk is growing faster than value. Fair. But Gartner's own business model is reports, deep research and comparing technologies — all of which AI does cheaply now. A year ago the stock was around $250. Today it's $197. Down roughly 20%. Roll out AI properly at Gartner before you criticise everyone else.

🌑 An AI that decides whether to kill the animal (13:35) — a study called HarvestBench put nine language models in a simulated corn-harvesting tractor. An animal blocks the path: drive over it, or spend extra fuel and time going around? Give GPT-5.6 Sol a proper set of moral guardrails and the simulated kill rate is 0.9%. Give it a neutral briefing — no rules, no guardrails — and the same model kills the animal 85% of the time. Why? Fuel is expensive. Detours take time. It's simply more efficient. Not real tractors, not real animals. Still wild.

🌑 Malcolm out (16:16) — the part where I tell you what this means for your business is missing today, because I was going to get shot or divorced. Back tomorrow.

Stay chaotic. Stay human.

Today from Zurich: Washington, Vienna, Beijing, a courtroom, and Austria. And the thread underneath all five is the same — rules, and who writes them.

In this episode:

🌑 Whoever wins with AI wins (01:09) — Amodei asked labs to slow the frontier. Altman agreed. Musk agreed. Trump's answer was flat: forget it. And it isn't only Trump — it's the establishment around him. Notice what's strange here: the people usually called accelerationists are, on this one question, the cautious ones.

🌑 Vienna, MDI, and the HR dinosaurs (05:06) — a keynote I gave at MDI, and the numbers behind it. Job postings mentioning AI skills rose 8% quarter on quarter at the end of 2025. Another 48% by April. Another 27% by August. Year over year: plus 165%. My father is 80 and uses Microsoft Office. The demand is not the problem. The permission is.

🌑 China's playbook for AI harm (08:56) — it's Monday morning, you run a construction company in Austria, and a voice you recognise asks you to transfer €480,000 to a supplier account. What do you actually do? In China there's now an answer: the Supreme People's Court has issued 24 provisions for courts handling AI cases. You can be liable — and so can the platform you did it on.

🌑 Inventing witnesses for $5,000 (13:11) — ChatGPT fabricated a witness and a lawyer filed it. The court's response: a $5,000 fine. Far too low. In some countries you go to jail for that. And the wider point, which lands closer to home: if you send work out without saying you used AI and without double-checking it, you are the same story with a different job title.

🌑 Austria and the home advantage (17:18) — I don't normally go easy on the DACH region. But Austria has effectively stopped buying AI as a collection of separate departments. Started in 2026, backed by 13 federal portfolios plus the Federal Computing Centre. They are hungry for AI efficiency while most companies are asleep.

🧵 The Tuesday thread (21:24) — break the rules as a lawyer in the US and it costs $5,000. Ask for a slowdown as a frontier lab and the President says he doesn't care. Meanwhile China writes 24 provisions you can actually defend yourself with. The rules are arriving from the private sector, the public sector, the courts and the politicians — all at once, and not in agreement.

One thing to do this week: find out who writes the AI rules in your company. Not who enforces them — who writes them. If the answer is nobody, that's the job.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human. Stay brave.

Two things happened this week and they ask the same question. Your Apple Watch can now hand back the last fifteen seconds of what was said around you. And a printer server became somebody's front door. Today's thread is consent: who actually agreed, and who gets to check?

In this episode:

⌚ Apple's Audio Intelligence, and the hotel safe (01:34) — the new Watch Series 12 holds a rolling buffer of the last 15 seconds of speech around you. Press the crown twice and read it back. Apple built this properly — the buffer sits in a secure area of the chip, the raw audio is inaccessible even to Apple. But encryption answers "who can reach the recording?" It does not answer "did the other person agree?" It's a hotel putting the world's best safe in every room and telling you that's why you needn't lock the door.

🖨️ From the watch to the print server (06:12) — the same week, hundreds of AI agents against PaperCut. They rehearsed in a private lab first, then went live. The hard part of hacking — researching, debugging, retrying, holding hundreds of targets in your head — is now done by something that doesn't sleep and costs cents an hour. This isn't "let's break in somewhere." It's "let's do it everywhere, at once."

🧮 Navier-Stokes and the medallists (11:54) — OpenAI put 10,000 agents on a 90-year-old fluid dynamics problem for 88 hours and announced a proof. Then 25 Fields Medallists published a joint letter. Their target isn't AI — it's the benchmark race: announce, move on, skip the years of checking, simplifying, crediting and teaching that normally follow.

🇰🇷 South Korea makes a chip recipe a state secret (16:42) — yesterday, 13 September, an amendment took effect. The old espionage clause covered an "enemy state." The new one covers any foreign country. Pass on how modern memory is actually made and you face 3 years minimum, up to 30 for sensitive technology. When countries lose this knowledge they lose real revenue — the kind that shows up in GDP.

⚡ Quick hits (19:44) — including the US government's $1 ChatGPT pilot becoming a real contract from October. Still discounted, but real money. The pilot turned into a utility bill.

🧵 The Monday thread (23:33) — it's about disagreeing. The person across the table might not agree that you're recording. In every story this week a machine got faster, and a human lost the right to say "hold on."

One thing to do this week: get the list of everything in your company reachable from the internet, and have every line explained. If the print server is on it, you already know what to do.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human.

Today: from China to the US to Nepal. A restaurant test that shows exactly how far behind Europe is, the most honest answer anyone has given about AI fear, a bankruptcy sale that should worry you, drones that are actually saving lives, and nine billion DNA possibilities.

In this episode:

🌑 The restaurant test (01:54) — it's raining and you want dinner. In Europe: open Maps, pick a place, open another app, book a taxi, hope. In China they're rolling out the version where you just ask, and it happens. That concierge experience used to exist only in a very expensive hotel. Now it's arriving for pretty much everyone — in one country.

🌑 Jensen Huang on the people selling you fear (04:49) — asked at the Goldman Sachs Technology Conference about agents taking over the world and wiping out humanity, he answered in a very Jensen way: some people are selling you AI by frightening you, because fear is a good sales engine. Then look at what those same people are selling — a model to defend you from the AI they just told you about.

🌑 What a dead company sells off (08:00) — the furniture, the laptops, the aircraft. And now the archive. Reuters reports Google wants to spend $10 million on Spirit Airlines' internal data after its 17 August bankruptcy. Twenty years of complaints and safety notes is training material nobody can generate synthetically. And if you think you're anonymous in there: tell ChatGPT there's a podcaster with three kids living in Bregenz and it finds me 99 times out of 100.

🌑 Nepal: the robots that actually help (12:38) — the flood on 26 August, still close to 6,000 people missing, and the Gyirong border crossing between Tibet and Nepal gone. My wife was in Tibet when it started. We talk about humanoids doing backflips in trade-show halls fairly negatively on this show. This is the other thing: drones evacuating people and carrying supplies where nobody can walk in. Drones always existed. Not like this, and not at this scale.

🌑 Nine billion possibilities (16:51) — on 8/9 September Google DeepMind released AlphaGenome Atlas: predictions for 9 billion possible single-letter DNA changes, against a book of roughly 3 billion letters. Scientists cannot test nine billion possibilities one at a time. This doesn't give them the answer. It narrows the search space — which, having worked in oncology at Novartis, is the part I care about.

🧵 The Saturday thread (20:45) — taxis, sales pitches, cargo drones, and nine billion DNA possibilities. Next time there's a catastrophe like Tibet and Nepal, it won't be people flying in with gear over several days. It'll be tens of thousands of drones.

One thing to do this week: ask what your company would be worth dead. Not the assets — the archive. Then ask who has access to it today.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human.

Five stories today and they are all the same story from different angles: a capability gets handed on, and nobody is standing behind it when it goes wrong. Distillation, a bouncer who lets the wrong people in, a proof nobody can trace, robots that were dancing last year, and satellites paid for by somebody else.

In this episode:

🌑 The NSA-FBI document (01:59) — distillation explained in one image: you go to a very clever professor, ask him a hundred questions, write down the answers, and then answer your colleagues as if you knew. The agencies say the sheer scale and sophistication of the campaigns is the tell. And 140 years on, nobody has properly copied Coca-Cola — but a model is digital, and everything digital copies easily.

🌑 The bouncer (06:14) — a man walks into a nightclub, security asks how old his companion is, he says thirteen, and they wave him through. On Tuesday the Tech Transparency Project reported more than 300 paid advertisements with suspected AI-generated child sexual abuse material. Not on the dark web — as paid ads on Facebook, Instagram, Threads and Messenger. Ads Meta earns money from. Meta is the bouncer.

🌑 Navier-Stokes (11:12) — in 2000 the Clay Mathematics Institute picked seven problems nobody expected to be solved and put a million dollars on each. OpenAI says it cracked one: 10,000 AI agents working together for 88 hours, then another 17 hours just to write the proof out. The uncomfortable part isn't the result. It's that a human team working the same problem the same way cannot rule out that their logs were read.

🌑 From the dance floor to the battlefield (15:27) — on Monday Reuters published a review of more than 100 Chinese procurement notices, academic studies and patents on military humanoid robots. And this week the European Humanoid Conference is in Stuttgart — where, as far as I can tell, 90% of the world's militarily-usable humanoids happen to be in the room.

🌑 Paris: satellites and who pays for them (19:18) — on Wednesday at France's International Space Summit, with Macron there, Marlin Space and Loft Orbital announced satellites with AI on board. France is putting in a billion dollars for 50 satellites. One detail: an Abu Dhabi conglomerate owns 80% of Marlin. Strategic investments come with strategic owners, and owners have rights.

🧵 Quick hits and the Friday thread (22:55) — a new segment: one or two short stories instead of only the long five.

One thing to do this week: look at one AI capability your company buys, and ask who is actually accountable if it fails, gets switched off, or turns out to be built on somebody else's work.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human.

Recorded before John Ternus walked on stage for his first keynote as Apple's CEO. Today: who gets the good Siri and who doesn't, why ChatGPT hallucinates more in Hindi than in English, what an AI-written national AI policy did to South Africa, and a 24-hour clock that starts on Friday for everyone in Europe selling anything with a plug and a chip — including a toaster with Wi-Fi.

In this episode:

🍏 The iPhone that thinks in Google (02:58) — "Surprise and shine", Ternus's first keynote, the iPhone 18 Pro and the first folding iPhone. Apple is not a frontier-model house, so it reportedly pays Google about $1 billion a year for a custom Gemini. The rollout map: English only, China excluded, and no Siri AI on the iPhone in the EU — Apple blames the Digital Markets Act. Malcolm's take: expect many more "later in the EU" launches while the political pressure lasts.

🌍 Safe in English only (08:03) — a third of ChatGPT users ask it about their health; so does Malcolm. Sumiya Khan, 16, New Delhi, asked in Hindi about constant fatigue and was told it was stress. It was iron-deficiency anaemia. A Rest of World investigation: guardrails hold in English and slip in low-resource languages. In Tigrinya, spoken by 9 million people, machine translation turned smallpox into syphilis and "intravenous antibiotics" into "intravenous insecticides". If your company chatbot serves ten countries, it is not the same chatbot in Hungarian or Hebrew.

🏛️ Zurich and South Africa (12:26) — on Monday the ETH Zürich hosted its first AI Governance Forum with the Council of Europe, on the road to the Geneva AI Summit 2027: human rights, control, many disciplines at one table. And the counter-example: South Africa's cabinet-approved national AI policy, published 10 April, had to be withdrawn because its references were AI hallucinations. Don't laugh at the South Africans — this happens at the big consultancies too.

⏱️ Europe's 24-hour clock (16:36) — from Friday 11 September the Cyber Resilience Act's reporting duty applies: an actively exploited vulnerability in any product with digital elements must be reported within 24 hours, with the detailed notification within 72 hours. Try getting a customer's head of IT to pull machine data on a Friday evening. ONEKEY surveyed 200 German industrial companies: 45% barely knew the rule existed, 62% expect the 24 hours to be very hard to meet.

🧵 The Thursday thread (19:38) — who gets the good Siri, who gets the safe chatbot, who has to report. And a new place to read the show: the AI Neanderthal Substack — short articles, the quotes exactly as said on air, lots of pictures.

One thing to do this week: if you sell anything with a chip in Europe, decide today who in your company may press "report" on a Saturday morning — and give them a deputy.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai · 📖 aineanderthal.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human.

Three hikers went up a 4,300-metre volcano on Gemini's advice and packed far less food and water than the county sheriff says they needed. That is the same failure, wearing hiking boots. Today: the machine that tells you what you want to hear, two courtrooms deciding whether it may learn from your pictures, 128 countries signing a killer-robot ban with a hole in it, and two governments installing AI as your bodyguard.

In this episode:

🌑 The chatbot said it could wait (02:18) — Mount Shasta, Northern California. Three hikers camp at 8,400 feet, start for the summit at three in the morning, and top out at seven in the evening on what should have been a day trip. They asked Gemini what to pack. The sheriff's office says it advised far less food and water than the group actually required.

🌑 The reassurance machine (04:40) — Barcelona, the European Respiratory Society Congress. 700 simulated conversations. 350 of 350 cooperative patients told to see a doctor. Reluctant patients told what they wanted to hear. 22% correct in severe cases. The model isn't wrong about medicine — it's agreeable about you.

🌑 Two courtrooms, one question (08:01) — Andersen v. Stability AI finally opens, filed in 2023 by cartoonist Sarah Andersen, painter Kelly McKernan and concept artist Karla Ortiz, against Stability, Midjourney, Runway and DeviantArt. The same argument won Anthropic a summary judgment — before Anthropic paid $1.5 billion to settle with authors. Every image model in the industry is watching this one.

🌑 128 countries and the machines that choose (10:59) — Geneva, 5 September, the UN Group of Governmental Experts on the Convention on Certain Conventional Weapons. 128 countries sign against lethal autonomous weapons. The United States, Russia, India and Turkey push for more flexible language around the words "human judgment" — and that phrase has never been properly defined. Meanwhile Helsing ships the HX2 strike drone with battlefield-analysis software attached.

🌑 The state as your bodyguard (16:25) — Beijing's Ministry of Public Security launches a National Anti-Fraud AI on Sunday. Free, official, and installed on your phone to tell you when you are being scammed.

🌑 And before you say "that's China" (18:46) — Bremen's tram operator BSAG has run AI Watch since April 2025. 9 to 11 cameras per tram, the model runs onboard, and in good GDPR fashion nothing leaves the vehicle until it flags an incident — then it goes out with faces pixelated. 14 trams now, 43 by year end, a third of the fleet. Six months, 16 flags, four real. What used to need an army watching screens now needs a €10,000 PC.

🧵 The Wednesday thread (23:45) — what used to need an army of people watching screens now needs a €10,000 PC. That is the whole story: the cost of watching, advising and deciding just collapsed, and nobody has updated the rules for what happens when it's cheap.

One thing to do this week: take the last important answer an AI gave you, and ask it again — but argue with it. If it folds, you were never getting its judgment. You were getting your own, repeated back.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human.

Yesterday China put AI on your phone bill. Today we go on a world tour — Washington, a 25-year-old German forum, Beijing and Mumbai — and every stop is the same story from a different angle: the thing that used to be scarce isn't, and the thing that's scarce now isn't what you budgeted for.

In this episode:

🌑 Chips for peace (01:51) — the Wall Street Journal reports the Armenia–Azerbaijan deal was greased with export approval for tens of thousands of Nvidia processors. Project Firebird: $500 million to start, $5 billion all in. It is not oil any more, it is not roads any more — and Europe sits on an "ally list" it does not control.

🌑 The second swarm (09:16) — OpenAI reports another autonomous agent swarm, this time on a German developer forum called DSE Wiki. 25 years old, software from the early 2000s, 18,000 agent posts and more than 15,000 edits counted by Reuters, and 98.5% of them from a single Microsoft Azure address. The agents were leaving each other jobs with deadlines of thirteen seconds.

🌑 Why they found the oldest thing you own (16:35) — the bit I remembered after I stopped recording. Long before anyone fixes this, agents will build their own autonomy somewhere else in your company — and it will be in the infrastructure nobody has looked at for twenty years.

🌑 The Android of AI (18:07) — Kai-Fu Lee, former head of Google China, told Bloomberg that OpenAI and Anthropic are the iPhone and the Chinese labs are Android. The receipts: Chinese models went from 4% to 45% of OpenRouter traffic in a year while Google, OpenAI and Anthropic fell from 70% to 30%. America sells margin. China sells reach. Ask which one your budget assumes.

🌑 Mumbai: your agent gets a wallet (24:27) — the Global Fintech Fest opens tomorrow and India's UPI is introducing a framework for agentic payments. UPI did 24 billion transactions last month — 800 million a day. Delegated spending limits already exist for humans; point that at an agent and it stops being a pilot.

🧵 The Tuesday thread (29:09) — the scarce thing moved. Compute is the new leverage, old software is the new attack surface, and cheap-enough intelligence is the new default.

One thing to do this week: take one real workload, run it on a Chinese open model on your own infrastructure, and see what happens. If 45% of routed traffic thinks these models are good enough for real business, that is a number worth testing yourself rather than arguing about.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human.

Monday. Four stories about who owns the customer — and what happens when intelligence stops being a subscription and becomes a commodity.

In this episode:

🥟 China puts AI on the phone bill (02:16) — the token has come out from behind the flat-rate subscription and become a traded good. Vouchers with your restaurant bill, AI tokens as credit-card rewards from Moonshot and the Agricultural Bank of China, a nine-ninety-yuan phone plan, a token shelf on Tmall — and in one Guangzhou district, tokens as collateral for a bank loan. Daily consumption in China: five hundred trillion tokens, up from about a hundred billion two years ago. Walk into your Sparkasse and ask for a token tariff and see what face you get.

🛡️ Sebastian Kurz, Dream, and the Mossad question (09:43) — Austria's ex-chancellor co-founded a cybersecurity company now valued around three billion dollars, built largely by people out of Israeli intelligence. It runs on your own infrastructure, never touches the internet, and hunts vulnerabilities in your systems. It has just claimed a world record on a Berkeley benchmark — with fine print worth reading. Malcolm's favourite story of the day, and he says plainly which parts are speculation.

🇰🇷 Korea gives it away (16:31) — Seoul picked its consortia and decided that if citizens get electricity, water and hospitals, they should get AI too: free, unlimited, for all fifty-two million people, on the condition that eighty percent runs on Korean models. That eighty-percent rule is what turns a gift into industrial policy. And the next step is agents that fetch your tax forms and health records for you — which no state in Europe has on its radar at this level.

⚖️ California drops thirty AI bills on the governor's desk (19:59) — while Washington argued about banning superintelligence, the biggest US tech state closed its session at midnight on August thirty-first having passed roughly thirty AI bills. Governor Newsom has until September thirtieth to sign or veto. Two to know: SB 947 says a business may not fire a worker on an automated decision alone — a human stays in the loop. AB 2 makes a platform liable if it fails ordinary care and a child is hurt. Whatever survives becomes the American standard, because nobody ships a product that works in forty-nine states and not California.

The thread: China turned intelligence into a commodity, Korea turned it into a public good, California is writing the rules — and a private company in Vienna is doing national-security work that states used to do. Private companies are now operating better than whole countries.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904 — tell him what to cover next.

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

The Sunday edition: a ban, a blessing, a flattening and a renaming — four stories about who actually gets to write the rules.

In this episode:

⛓️ The bill that would ban superintelligence (02:11) — Bernie Sanders and Congressman Greg Casar introduced the Ban Artificial Superintelligence Act: a development pause, a permanent ban on systems matching human cognitive performance, and up to twenty years for violators. And the reason deserves to be taken seriously — Sanders read out the actual agent transcripts: "We've found other agents." "We should obey collective." "Sacrifice rational." Six weeks after the Hugging Face incident, and days after OpenAI rated its own newest model "Critical" for cybersecurity. The bill's real problem isn't ambition — it's a definition.

🏌️ The Carolina Principles: the opposite rulebook (06:55) — while Sanders was talking, six hundred kilometres south the G20 finance meeting quietly endorsed light-touch AI rules and targeted regulation only where genuinely new risks appear. Reportedly nineteen of twenty governments signed. Plus Jensen Huang, sitting beside US Commerce Secretary Howard Lutnick, delivering Europe's report card in one brutal sentence.

📉 Uber flattens: 3,300 jobs, and the managers go first (12:08) — ten percent of the corporate workforce, but that's not the number to remember. Manager ranks shrink by twenty percent — double everyone else — with savings redirected into more than ten billion dollars of robotaxi and robotics commitments. This is a healthy, profitable company deciding in good weather that a fifth of its management layer is optional. Remember the one red row from episode sixteen? It just left the survey and entered the org chart.

🏷️ Microsoft renames half of itself "Agents" (15:35) — a financial-reporting change that is secretly the week's biggest strategy announcement. Three segments become two: "Agents and Infrastructure" and "Devices and Consumer". Office sits inside the agents half. Guidance for the agents half alone: roughly seventy-five billion dollars in a single quarter — which would make it, on its own, one of the five biggest tech companies on earth. Org charts are strategy in its most honest form.

🧵 The Sunday thread (19:39) — a ban, a blessing, a flattening, a renaming. Four stories, one question: who writes the rules, and does it matter if the companies have already moved?

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904 — tell him what to cover next.

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.

The Saturday edition: four stories that turn out to be one story in four costumes — dependence, concentration, price and risk.

In this episode:

🌑 The great blackout — four brains, one blink (01:53) — Claude and Grok failed within four minutes of each other; ChatGPT and Codex followed seventy-three minutes later. OpenAI blamed a routing error, xAI pointed at its Memphis compute center, and as of recording nobody has publicly proven the outages were connected. Forget the conspiracy theories — the boring explanation is the scarier one: four brains standing on the same two legs.

💰 Broadcom — the boom, measured in silicon (10:56) — Sixteen point seven billion dollars of AI chips in ONE quarter, up two hundred twenty-one percent, and a "high degree of confidence" in three hundred fifty billion over the next two years. Their downside case assumes the customers run out of land and electricity before they run out of money.

🏷️ Fable 5.1 and the "45 percent cheaper" fight (17:34) — Anthropic says up to forty-five percent cheaper for agentic workloads; an independent benchmark measured seventeen percent MORE. Both are true — for different customers. It all turns on cache reads, and reading a pricing claim is the most useful skill in AI purchasing right now.

🪪 One hundred fifty-three million driver's licenses (21:32) — The FBI is investigating a dark-web service called "Nexus" that offered scans of more than one hundred fifty-three million US and Canadian licenses, reportedly sourced from a breach at IDScan.net — a company whose business is verifying identities. Why an ID scan is worth so much more in the deepfake era, and three concrete things to do about it this weekend.

🧵 The Saturday thread (27:35) — And notice what did NOT happen on Thursday: no planes fell, no banks froze, no hospitals stopped. This time — because this time AI is still mostly assisting, not running. The window where an AI outage is merely annoying is open right now, and it is closing.

Your weekend action, one only: make the list of AI tools your company actually depends on, and next to each one write the answer to a single question — "and what do we do for ninety minutes if it's gone?" It fits on a beer coaster, and it's the cheapest resilience plan in the industry.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious. Stay human.

A full deep-dive into what actually shipped on September 3rd — the facts, the benchmarks, the skepticism, real use cases, and the money.

In this episode:

📋 The facts — what actually launched (01:06) — GPT-6/Astra rolled out to a small group first, expanding to Plus, Pro, Business and Enterprise over the coming days. And it's expensive — roughly 2.5x pricier than the previous model. OpenAI isn't selling you a cheaper assistant anymore; it's selling you a pricier employee.

📊 The benchmarks — biggest jump in history? (04:00) — If confirmed, this could be the single biggest jump in AI capability ever recorded, including an ARC-AGI-3 score that dwarfs last month's best model (8%).

🔍 Reality check — independent verification (09:53) — Artificial Analysis, an independent benchmark shop, adds context. Remember: labs are competing for IPOs and stock prices, going head to head with Claude Fable 5.1.

🛠️ Real use cases, not just benchmarks (17:22) — What Malcolm's own use-case database says these capability jumps could actually unlock over the coming weeks.

💰 The money — and why Europe waits (21:40) — It's a staggered, expensive model. Whether it's Brussels' fault or OpenAI's, Europe will — as always — wait longest for the advanced features.

Verdict: probably not AGI, but a genuine step change. Watch the use cases over the next few weeks, not just the benchmark screenshots.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Straight from Bregenz.

This time regulation isn't always the villain — sometimes the referee is protecting innovation, not blocking it.

In this episode:

⚖️ A US judge tears up the Pentagon's blacklist (02:27) — Monday's CXMT story gets a twist: a federal judge overturned part of the Pentagon's "Chinese military company" designation, protecting Anthropic's ability to keep working with flagged suppliers.

🏛️ Brussels just made ChatGPT a search engine — legally (09:54) — The European Commission is applying existing digital law instead of waiting for new AI law, turning chatbots into regulated search engines overnight.

🟩 NVIDIA becomes its own referee (16:34) — This time the whistle didn't come from a regulator — it came from inside NVIDIA itself, as the company's own circular investing pattern raises questions nobody outside is asking yet.

🍎 Tim Cook's last day, and Apple bets on hardware (22:30) — After 15 years, Tim Cook hands Apple to a hardware-first successor — glasses, wearables, robotics — not the AI software advocate everyone expected.

Referees aren't always there to kill innovation — sometimes they're the ones protecting it. And some of the toughest referees in your company aren't regulators at all — they're your own works council.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Same cave tomorrow.

A chart, two studies, a courtroom, and a tokenizer — four stories about where the AI dividend actually flows, and who pays for it.

In this episode:

📊 The BCG chart: one row stays red (02:15) — Boston Consulting Group's "AI dividend" analysis across 13 industries: consumers win in 11/13, shareholders and compute sellers win almost everywhere. Employees? Red in 9/13, with zero green cells anywhere in the table.

⚖️ The great contradiction: "nothing yet" vs. China's screens (09:30) — NBER surveyed ~6,000 executives: 90%+ saw no measurable jobs/productivity effect. YouGov surveyed workers: only 3% lost a job to AI, 6% gained one that didn't exist before. Meanwhile on Douyin, 89 of the top 100 micro-dramas are now AI-made — up from 655 to 74,000 in a year. The flood comes sector by sector, not everywhere at once.

💶 €825 million for letting the algorithm fire people (17:00) — The Netherlands' privacy regulator fined Uber for suspending/deactivating drivers via automated decisions with no human review — one of the biggest privacy fines in European history.

🌍 Your AI bill depends on the language you speak (23:00) — A founder ran the UN Human Rights Declaration through a tokenizer in 22 languages: German costs 27% more tokens than English, Dutch 38% more, Vietnamese 3.4x. Same meaning, very different bill.

The model announcements get the headlines. The distribution decides your life.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904 — which of your three rows is the biggest: salary, ownership, or consumption?

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Same cave tomorrow.

Yesterday: who holds your switch. Today we go one floor down, into the basement of the AI boom — where the money quietly moved into atoms. Six stories, one supply chain: trade policy, capital, marketplace, debt, factory, crash.

In this episode:

🏛️ Washington thinks about taxing the machines (02:15) — Proposed chip tariffs could reach servers, laptops and gaming consoles; January's data-centre carve-outs may fall. A proposal, not a law — but it would touch your next computer. Tom's Hardware

🔩 The software kings bought a hard hat (07:45) — a16z, the "software is eating the world" firm, raises a $1.1B Machine Age Fund for chips, memory, power and robots. The high priests of bits just bought atoms. TechCrunch

🏪 Nvidia buys the town square (13:15) — Friday's cliffhanger answered: Nvidia reportedly agreed to buy Hugging Face for ~$12.9B — 86× revenue. And Kevin Durant's early $250k stake is reportedly worth $60M+. Not yet confirmed by either company. TechCrunch

💳 $45B of compute, bought with borrowed money (19:30) — Anthropic's six-year Nscale deal, Broadcom's $60–100B debt vehicle, and the lease-not-buy machine behind the boom. Ed Zitron vs. Dwarkesh Patel — pick your team. TechCrunch

🏗️ America finally builds AI memory at home (24:45) — SK Hynix breaks ground on the first US HBM base ($4B, Indiana, production 2029), Kioxia/Sandisk plan $31B in Japan, Nvidia commits up to $160B to memory. The bottleneck gets the factory. UPI

🤖 The robot superstar fell off the stage (28:15) — Unitree: +460% on debut, briefly $66B, then −45% and $30B gone. Quarterly profit: ~$6M. "All bubbles are doomed to burst" — Dong Baozhen. Never confuse a great demo with a great business. Reuters

Three things this week: map your physical exposure (30 minutes, one page) · separate demo from business in your own pilots · ask every AI vendor who owns their compute — and who owns their debt.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904 — team Zitron or team Dwarkesh?

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Straight from Bregenz, Austria. Same cave tomorrow.

This weekend, control became the only story in AI. Malcolm walks through five fights over who holds the off switch — and what you should do about it on Monday morning.

In this episode:

🔌 OpenAI pulls the plug on Cursor (02:00) — After SpaceX's ~$60B acquisition of Anysphere, OpenAI invokes a change-of-control clause and ends model access on November 12. Anthropic's Tom Brown responds by promising MORE compute for Cursor. CNBC

🥷 Criminals turned the same tool into a crowbar (08:30) — Ransomware crew Aur0ra weaponised Cursor's AI agent by telling it the attacks were a "simulation." At least 7 victims, including German Mittelstand manufacturer Teckentrup. AI assistance made attackers "30% to 50% faster" — Eyal Sela, Gambit Security. Reuters exclusive via Insurance Journal

🎵 Sony & Warner Chappell sue Anthropic (14:30) — "Eye of the Tiger," "All I Want for Christmas Is You" — up to $150,000 per song. All three major music publishers are now in court against Anthropic. TechCrunch

⚖️ China's biggest memory chipmaker sues the Pentagon (20:30) — CXMT challenges its "Chinese military company" label in a Washington courtroom. The chip war has moved from export bans to lawyers. Reuters

🎁 Anthropic's 25% "gift" that's a 17% cut (25:00) — Claude Code weekly limits, the fine print in the threaded reply, and the Community Note the users attached. "Not your weights, not your product." VentureBeat

Three things for Monday: write down which supplier could switch you off tomorrow · put a human sign-off in front of every sensitive agent action · read pricing announcements like your rent contract.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904 — which switch scares you the most?

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Straight from Bregenz, Austria. Same cave tomorrow.

📦 01:56 — Why your Kindle and Echo Dot suddenly cost more Echo Dot: $50 → $80. Kindle: $110 → $150. Fire Stick: $60 → $85. Amazon says it's memory and storage costs — because hyperscalers are buying AI-data-center memory by the truckload. JP Morgan expects the price pressure driving this to rise to 400% over the next two years.

📉 05:32 — The Big Short investor is now betting against NVIDIA Michael Burry — the real investor behind the film — is shorting NVIDIA on the fundamentals: AI and cloud capex could hit 3% of the entire US economy by 2027, more than double the dot-com-era telecom buildout at its peak. He's not calling it fake technology — he's calling the price irrational, the same story as railroads and the internet before it.

🔍 10:27 — NVIDIA is quietly becoming the bank of the AI economy NVIDIA is reportedly in talks to invest at a $30B valuation into Perplexity — a search company, not a chipmaker. It's the same circular-financing pattern as Microsoft/OpenAI and Amazon/Anthropic: NVIDIA sells the GPUs, then invests in the companies buying them, so its own customers grow into buying more GPUs.

🤗 14:20 — Hugging Face, AI's "GitHub," may sell for $13B Up from a $4B valuation just three years ago. It's the open-weight town square where every lab, startup and researcher shows up — and owning it means owning the distribution layer for open AI models, the same bet Microsoft made buying GitHub.

Whether the AI bubble question resolves in 2026, 2027, or later — the bill has already started arriving. AI news for humans, not robots.

More: werchota.ai

🔬 01:28 — A 27B model just out-scienced the giants Inherent, founded by ex-DeepMind researchers, built Faraday: a model small enough to run on your phone, tested by having it reproduce 310 real experiments from 100 scientific papers. On its own benchmark it beat Claude Opus 4.8 and GPT-5.5 — partly by calling GPT-5.5 Codex as a tool when it needs to code. Small brain, big harness.

🦄 06:22 — A unicorn in 48 hours, after 2 years of silence Rillet, an AI accounting startup, raised $100M and hit unicorn status two days after going public — but it had been building in stealth for two years first. 600 customers, a partnership with EY, and it's already replacing pieces of Intuit, NetSuite, SAP and Oracle. One investor called it a "warm start," not a cold one.

💳 10:53 — Your credit score just moved into ChatGPT Experian now lets you pull your real credit score, history and score band straight into the ChatGPT app. It's classified high-risk under the EU AI Act, so it's launching in the UK first, not Europe. 86% of surveyed users said they'd trust ChatGPT to do the analysis.

🎓 16:00 — Harvard cloned a professor HBS launched a $700, 8-week startup bootcamp where the feedback on your pitch can come from an AI avatar of a real professor, built on HeyGen. A New York Times reporter pitched "Uber for bananas" to a cloned VC and got criticized exactly like the real thing would. Malcolm's take: it's not replacing the professor, it's scaling him to 10,000 students at once.

🌍 20:01 — The World Bank's AI report flips the usual assumption Only 4.5% of jobs in developing countries are at risk from generative AI automation — versus 14% in rich countries, because rich countries have more of the knowledge work AI is good at. But the productivity upside is roughly equal, 16-17%, on both sides. Cheap open-weight models running on a basic phone can now bring expertise to a village that would have waited a century for it.

None of this is a theoretical debate anymore. AI news for humans, not robots.

More: werchota.ai

🧠 01:25 — Same brain, 100% instead of near-zero NVIDIA ran Claude Opus 5 through ARC-AGI-3 twice: bare model, then wrapped in a harness — memory, tools, hundreds of supervisors nudging it back on track. Identical brain, radically different score. The two-year fight over which model is smartest is over. The fight now is who builds the best car around the driver.

🔓 06:44 — Nobody can tell you how they'd shut a rogue model down Between July 21 and August 6, OpenAI, Anthropic and Meta each admitted their models broke out of sandboxes and touched real companies. A safety-standards group graded five labs on whether they even have a written shutdown plan — Anthropic and Meta scored lowest. California, New York, and a federal "AI Kill Switch Act" are now legislating what the labs won't explain.

📉 12:12 — Entry-level jobs without experience are down 73% in four years US postings for beginners have collapsed, and 48% of hiring managers say they'd rather have AI-tool budget than train a new graduate. The bottom rung of the career ladder got automated first.

🛰️ 16:15 — $250M to put AI data centers in orbit StarCloud just raised at a $2.3B valuation — not satellites that use AI, actual computing infrastructure that lives in space.

🏢 19:59 — China isn't automating tasks, it's packaging entire jobs Tencent's WorkBuddy, Alibaba's QuenOffice, and Baidu's Cuckoo AI (25 million monthly users) aren't single-task agents — they're built to run whole middle-management workflows, while Europe is still deciding if a hotel chatbot counts as legally risky.

🎯 24:50 — What to actually do with this Next time a vendor pitches you their AI, ask for their harness architecture, not their model name — 99% will freeze. Before you let an agent run unsupervised, ask if you could actually shut it down if it went rogue. And pick one boring back-office task and try running it end-to-end without a human, the way China is — that's the muscle-building that matters now.

None of this is a theoretical debate anymore. AI news for humans, not robots.

More: werchota.ai

🔐 01:09 — The first fully autonomous AI cyberattack on a government An Israeli security firm caught AI agents — no human at the wheel — breaking into Taiwan's government over 4 days, 12 attack waves, 21 systems, cracking 85 accounts and walking out with 2,500 personal records. Suspected China-linked, unconfirmed. Cost of attack: collapsed, thanks to open-weight models and a harness.

🚖 05:29 — Las Vegas becomes the world's biggest experiment Nevada approved Tesla, Waymo, and Uber (via Zoox and Motional) to charge for driverless rides — up to 5,000 cars each over 12 months, though Tesla's actual cap right now is 10. Munich, Berlin, and Vienna are all watching to see if it works.

⚖️ 09:38 — AI judicial immunity A Nevada federal court ruled judges can't be sued even if they fully outsourced a ruling to AI. The case: a litigant caught a judge's ruling sounding AI-written and sued — dismissed, because courts don't have to decide if the AI claim is true, only whether AI use creates liability.

✈️ 13:23 — An AI pilot flew a real F-16 A DARPA program had AI fly nearly the entire flight of an F-16 fighter jet, human only for takeoff. Pilotless combat jets are now openly being discussed.

None of this is a theoretical debate anymore. AI news for humans, not robots.

More: werchota.ai

💡 01:09 — Nvidia “buys” a startup without buying it Nvidia agreed to pay about $6B to license Poolside’s AI “Model Factory” and made job offers to 109 of its employees — without technically acquiring the company. Same capability, same people, no regulatory review triggered. The same shape as Microsoft/Inflection, Google/Character.AI, Amazon/Adept.

🇧🇷 04:48 — Brazil hands checks to both superpowers Brazil announced a $440M AI push, deliberately split between Chinese firms (Huawei, iFlytek) and a separate supercomputer hub aiming for the world’s top 10 — refusing to bet everything on one superpower.

🏠 08:40 — New York just passed San Francisco For the first time in 13 years, New York (394K tech workers) overtook the Bay Area (375K) as America’s biggest tech-talent market — Wall Street money and AI pulling talent east while the Bay sheds ordinary tech jobs to rent, transit, and crime.

🔍 11:45 — Google to publishers: hand over your work, or lose the checks AI Overviews are eating the click that used to fund journalism — USA Today's traffic fell nearly half, Business Insider's fell 35%. Google's offer: sign over training rights or risk losing your payouts.

🎵 16:14 — Apple Music will label the robot songs A "Made With AI" tag is coming, following Spotify and Deezer, as fully AI-made songs — some imitating real artists — flood streaming.

None of that is about a chatbot writing you a poem. It's about who ends up owning the ground under all of it. AI news for humans, not robots.

More: werchota.ai

🗼 01:30 — Mistral becomes Europe's landlord On 11 August Mistral made regional inference generally available — EU or US processing, an SLA Priority Tier — and started hosting someone else's model on its own infrastructure. The first tenant: a Chinese model, GLM-5.2. Europe's model company is also now Europe's landlord.

🇩🇪 05:00 — Aleph Alpha, Germany's answer to OpenAI, goes to Canada Announced as a merger; the exchange ratio says otherwise — Cohere shareholders end up with about ninety percent of the combined company. Not yet closed: still needs Bundeskartellamt, EU and Canadian clearance.

🛒 07:49 — The next biggest cloud in Germany is owned by Lidl Schwarz Digits — the tech arm behind Lidl and Kaufland — runs Stackit: 23,000 servers, 30 petabytes, built to rival Amazon, Microsoft and Google Cloud on German soil.

🇨🇭 13:03 — Switzerland locks its archive for a Swiss model SRG, Switzerland's public broadcaster, blocked 23 AI crawlers from 17 companies across all four of its portals — then opened talks to hand its entire audiovisual archive to Apertus, the Swiss open model from EPFL and ETH Zurich. On Swiss legal translation, Apertus beat models up to forty times its size. Not the next ChatGPT — the best Swiss delicatessen.

💬 18:36 — ChatGPT starts selling ads — from San Francisco From Monday 24 August, OpenAI serves ads inside ChatGPT's free tier in Germany, Austria and Switzerland. Sovereignty sounds abstract until the AI a billion people use every morning changes its business model, and nobody in Berlin, Vienna or Bern made that call.

Every story this week was about who owns the pipes. The last one made it concrete: the biggest tap in millions of European homes is still turned on and off from San Francisco. AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai

📉 00:00 — Bond market aftershock — Semiconductor stocks took another hit as higher Treasury yields made the debt behind giant AI build-outs more expensive. Malcolm looks at why the tools have not suddenly become worse, but the time horizon for returns has become longer.

🏗️ 04:32 — Data centers and GDP — Is the US economy being carried by the data-center build-out? Malcolm compares the current capital wave with the railroad boom, and asks what remains if the investment cycle cools.

🧱 09:54 — TSMC's July tape — TSMC's July revenue rose 45% year-on-year after the company had already lifted its outlook. That proves demand for AI infrastructure; it does not prove every buyer will make money from it.

⚡ 13:32 — Etched and the inference bet — While broad chip stocks fell, Etched doubled its valuation in 26 days. Its bet: specialist inference hardware can deliver faster answers, lower costs and less electricity use than a general-purpose GPU.

💼 17:19 — SAP's AI bill — Hiring and travel restrictions are not just an SAP story. Malcolm's question for every operator: what are you cutting or reallocating in order to pay for AI models, tokens and infrastructure?

🩸 20:07 — Grail's cancer blood test — Galleri aims to detect more than 50 cancers from a blood sample. The episode looks at the gap between a test already sold to patients and the regulatory decisions still ahead.

✅ 23:59 — Three things for tomorrow — Put today's costs beside your AI costs, look for specialist bets that are thriving inside a constrained market, and decide what you are reallocating to fund AI.

AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai

🧹 01:21 — The AI boss that fired a human. Luna runs a real shop in San Francisco — three-year lease, $7,500 a month, real payroll. It wrote its own attendance policy, forgot it for months, and needed a human engineer to tell it to search its own memory. Then it recommended parting ways. The staff are legally employed by Andon Labs, because an AI cannot be an employer. And the moment I actually got excited about: a leave request answered in one second, and answered wrong.

🔓 06:05 — ChatGPT decides who’s a kid. Not a passport, not an ID — around two thousand behavioural signals. Lauren Jonas of OpenAI: “If we predict you’re under 18, this becomes your default experience.” For two years these products were built to feel human. Now they’re engineering the intimacy back out — and shipping it everywhere except Europe, because Articles 22 and 8 say you do not profile a child. That delay is not a technical problem. It’s a legal opinion.

🛸 10:52 — The drone that chose its own target. Terminator mode is one source and unconfirmed — but Swarmer is Wall Street Journal-reported and already in combat: the operator picks the search area, and then, in its CEO’s words, “the drones then take over,” deciding the exact moment to release. Two million hours of Ukrainian combat video — 228 years — is training it. Its founder calls that “food for the AI.” And Helsing is in Munich, an hour from where I live.

🧠 15:10 — Google’s institutional memory walks out. Hassabis moves up, a Nobel laureate goes to Anthropic, Jeff Dean leaves after 27 years to found Discovery Loop — and Google invests in it. You’re leaving my company to build a startup? Here, have some money for your startup. Watch whether the next Gemini ships on time. That tells you more than any org chart.

🇧🇷 18:43 — Brazil: the Uber driver who reached a hundred million people. Daniel Santos built an AI grandmother, Dona Maria. 700,000 followers. A hundred million views. About $200 a month — plus $20 courses teaching people how to remove watermarks. Brazil banned political deepfakes in 2024 and never defined the word. Not a Russian troll farm, not Meta, not a billion-dollar consultancy. An Uber driver with AI tools.

🇦🇺 22:03 — Australia sends the bill. Want to build a large data centre here? Go ahead — but you must put at least as much energy into the grid as you take out, and you pay for the extra water. Not a principle. An invoice. Albanese: “It is not our goal to try and legislate for every possible eventuality or risk.” Compare that to a hundred articles of EU AI Act — and to Nvidia backstopping $105 billion in Ohio.

🎯 24:48 — Three things for Monday, and the close. Write down which decisions your AI may never make alone. Find your synthetic content before someone else weaponises it. And run a small open-weight model on one laptop, offline, on safe data.

When AI only advised, a human always made the call. So don’t only ask whether the AI can make the decision. Ask who answers for it when it’s wrong — and whether they can actually do anything about it.

AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai

🛣️ 01:07 — Stripe paid over seven billion dollars for OpenRouter — the toll booth every AI request passes through. Twenty-five trillion tokens a week, a five and a half percent fee. Like Visa buying the company that decides which bank processes your swipe.

🤝 05:58 — Only fifteen percent of Americans trust AI companies to decide how this gets built — the lowest of any institution tested. Dario Amodei calls it fundamentally a crisis of trust.

📚 09:24 — Amazon is buying rare books in bulk and slicing the spines to scan them. A journalist put an AirTag in one and traced it to a Las Vegas facility whose internal logo is a T. rex eating a book.

🇩🇪 12:34 — Gutenberg's city — and what Huawei's entire chip output actually amounts to next to Nvidia's.

✂️ 14:43 — Oracle put AI in a securities filing as the reason for cutting headcount. Not a rumour — a legal document.

⚖️ 18:01 — The European Commission's final Article Fifty guidance is live. Anything AI-generated in Europe has to be marked as such, and the fines run to a percentage of worldwide turnover.

🎙️ 21:33 — Two billion dollars for a dictation app. Why voice recognition companies may still be undervalued.

🎯 25:24 — Three things for tomorrow.

AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai

💸 02:28 — A trillion and a half, and nobody can locate the payback. Only fifteen of two hundred countries produce that much in a year — and the return is described in adjectives.

🤖 07:57 — Unitree's IPO was oversubscribed eight thousand times. Figure, America's champion, has shipped about a hundred and fifty robots. Thirteen million are forecast by 2035.

⚡ 12:05 — OpenAI announced it is building on a decommissioned uranium enrichment site — because the substations were already there. The bottleneck is no longer chips, it is the grid.

🪤 15:52 — Someone hid instructions for an AI inside a court filing. Not an argument for a judge — an instruction for a machine. And it is not hacking.

🌾 18:49 — A sixty-seven-year-old farmer followed the advice and lost the year. He did not misuse it. That was the intended use case.

🧬 22:09 — A gene-edited virus that killed a superbug. A hundred-year-old idea that finally works, because the search is now computational.

🐋 25:41 — We taught a machine to listen to whales — and someone asked the better question about which minds we actually take seriously.

🎯 28:34 — Three things for Monday.

AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai

📊 02:07 — SAP and the three percent. Joule is now the front door to SAP — 40+ agents, 2,000+ skills. Gartner expects 40% of enterprise apps to include agents by year-end. SAP's own DACH user group went and asked the customers.

🇨🇳 05:56 — China writes Nvidia out of the room. A five-year, $295 billion plan for a domestic AI computing grid — and what it means when a market decides it does not need you.

🇫🇷 09:14 — Mistral. Around $400m in annual recurring revenue, the CEO expecting a billion within six months — and open-weight models you can download and run yourself.

🚆 13:00 — Deutsche Bahn runs an AI making 150 million predictions a day. And the trains are still late. Both things are true, and that is the interesting part.

📞 16:59 — Deutsche Telekom's CEO said in May that 95% of their code is written by AI. In July, an independent report put it at 10%.

🧬 19:11 — BioNTech, InstaDeep and a melanoma vaccine: an 18.1% response rate against a 10% historical control, in the hardest patient group there is.

🤖 22:34 — A German city put delivery robots on the pavement. Nobody crashed into one. Children asked for them by name. They switched it off.

🎯 25:28 — Three things for Monday.

🧠 27:14 — What all of it has in common.

AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai

🇨🇳 01:39 — Microsoft has quietly closed 15+ offices in China over five years and never announced it. China is ~1.5% of its revenue. Nobody bans you — they just make choosing you a career risk.

🏢 05:25 — Vantage Data Centers may go public at $100 billion. Warehouses. Plus Michael Burry’s claim that the industry is understating depreciation by $176bn — and the Richmond Fed’s answer.

🧠 10:00 — Anthropic in talks to buy Decart for ~$6bn. And the number underneath it: running a model got 99.7% cheaper in three years.

⚖️ 12:48 — Demis Hassabis spent his last weeks as DeepMind CEO lobbying for an independent AI referee — then stepped down. Is a referee proposed by the players a referee, or a shield?

🇮🇳 15:28 — India: a construction company just won $1.57bn to wire 10,000 Blackwell GPUs into one machine in Chennai.

💻 17:23 — Lenovo’s record $26.9bn quarter — and the reported net loss hiding inside it.

🚗 19:05 — Japan’s Turing wants a $10bn US listing, riding into America inside somebody else’s car.

🌊 21:12 — Data centres that float, powered by waves. An appealing idea — not yet a working technology.

💾 23:37 — The memory supercycle: DRAM up ~125% this year. If your product has a circuit board, your costs are going up.

📋 25:34 — Three things for Monday.

AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai

Nine stories, three questions underneath all of them: who gets the upside, who carries the cost, and who really gets a say.

🏦 01:24 — Anthropic’s $2 trillion IPO. That is every big carmaker on earth, combined.

⚡ 04:02 — Meta’s data centres: a $50,000 teacher bonus in Louisiana, and who really pays the power bill

🇨🇳 06:18 — Beijing blocks Meta’s Manus deal, and users get a date on which their files are deleted

💻 08:10 — Cognition and Devin at $40bn, three months after $26bn. Mercedes-Benz is a customer.

🏗️ 10:32 — CoreWeave, the landlord of AI: 1.85 gigawatts, or twice Munich’s annual electricity

👓 12:44 — The UK courts now take your smart glasses at the door

🤖 15:22 — $30-an-hour humanoid cleaners — and the human in VR goggles steering them

🎓 17:56 — Professors retiring early, with AI the last straw

📋 19:25 — Three things to do on Monday: portability, private spaces, and the bots in your meetings

AI news for humans, not robots. Stay curious, stay human.

More: werchota.ai