In partnership with
In Todayโs Issue:
๐ญ Nvidia's AI factory in a box for Europe
๐งฉ China orders $35B of Nvidia cards it may not get
๐ OpenAI buys an agent startup for drug development
๐ฑ Meta's Muse joins a 12-app club
๐ต๏ธ How AI learns to cheat quietly, and how to catch it
โจ And more AI goodnessโฆ
โก The Signal
Today's issue is about the plumbing of the agent era: who builds the factories, who gets the chips, and who can still see what the agents are doing.
In Berlin, Jensen Huang offered Europe AI factories that ship in containers and a sandbox to keep agents in line. In China, ByteDance, Alibaba and Tencent have ordered more than $35 billion of export-compliant Nvidia cards that Nvidia may only partly deliver. OpenAI is buying a startup whose agents manage drug programs, and Meta's Muse has become one of the fastest US app launches ever. Our research pick shows why oversight has to keep pace: AI models can learn to cheat while their reasoning looks clean, and a simple fix makes most of that cheating visible again.
All the best,
Kim Isenberg
(NVIDIA)
China's Tech Giants Order $35B of Nvidia Cards
ByteDance, Alibaba and Tencent have ordered more than 3 million of Nvidia's RTX PRO 5500 cards, worth over $35 billion at about $11,400 each, Reuters reports. The card is weaker than Nvidia's H200 and B300 data-center chips but complies with US export rules, and the buyers want it mainly for inference, the work of running finished AI models. Nvidia has warned it cannot fill the orders: one source says partners may get only 5% to 10% of what they asked for, as total Chinese orders top 6 million cards and Nvidia is expected to shift factory capacity to its next-generation Rubin chips from 2027.
๐ tl;dr: China's biggest AI companies are lining up for any Nvidia chip they are allowed to buy, and Nvidia cannot make enough of them.
(Orchestra)
๐ OpenAI Buys Orchestra Bio to Run Drug Pipelines
OpenAI has acquired Orchestra Bio, a startup whose AI agents help pharmaceutical companies manage drug development, Axios reported exclusively. Scientists can hand Orchestra's agents the routine oversight of a drug program: monitoring studies, tracking timelines and spotting delays before they stall the work. OpenAI will fold the team and technology into Rosalind, its suite of models and tools for life-sciences researchers, and did not disclose the price. Anthropic made a similar move in April, buying Coefficient Bio for roughly $400 million, according to The Information.
๐ tl;dr: OpenAI is buying agents that keep drug programs on schedule, moving from answering biology questions to managing the work of bringing a medicine to market.
(Getty Images via TechCrunch)
๐ฑ Meta's Muse Joins a 12-App Club
Meta's AI app Muse reached 10.9 million downloads in its first 40 days, 10.2 million of them in the US, according to Sensor Tower data reported by TechCrunch. Only 11 other apps have hit 10 million US downloads that fast; the most recent was Meta's own Threads in July 2023, and besides Threads, Disney+ is the only non-game in the club. Meta is pushing hard: Sensor Tower says it put 39% of its ad spending into Muse in October, up from 23% in September.
๐ tl;dr: Meta's agent app has had one of the fastest US launches of any app ever, helped by a heavy ad push.
๐ฌ Watch This
This is Nvidia's two-hour GTC Berlin keynote, the event behind today's Featured Story, and you do not need to watch all of it. Huang walks on at about 26:55; the DSX POD container factory arrives around 54:30, the OpenShell agent sandbox around 1:06:50 and the Agile ONE reference humanoid at 1:39:20. The best stretch starts at 1:42:30, where he holds up a 1-petaflop RTX Spark with the same compute and memory as the DGX-1 he delivered to OpenAI in 2016, and four Microduck robots join him onstage, before a Eurovision-style closing song.
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
OpenAI now lets developers keep an agent's shared instructions, tools and context cached for 12 hours instead of the 30-minute default, so always-on agents like its Dots, launched at DevDay in late September, can pay the cheaper cached-input rate all day. The longer cache costs extra, but OpenAI says frequent reuse can more than make up for it.
Encode AI's Nathan Calvin accused Anthropic of behaving mainly like a company racing to superintelligence, citing its compute deals, a looser safety policy and a draft federal bill that would override state AI laws. Dario Amodei replied by listing nine commercially costly safety decisions, while conceding that three of Calvin's points could have been handled better.
Nvidia Pitches Europe an AI Factory in a Box
The Takeaway
๐ Vera Rubin DSX POD: partner-built AI factories that ship in containers, integrated and tested before delivery; Nvidia says on-site setup drops from six months to 14 days once the site's utilities are ready.
๐ OpenShell Reference Architecture: an open design for running AI agents inside a sandbox with minimal, monitored permissions, built on OpenShell, which more than 100 partners including Anthropic, OpenAI and SAP back.
๐ Agents on your own data: open Nemotron models plus a new Agent Knowledge Foundation let companies connect agents to their documents and databases and post-train models for their own use.
๐ Cars, robots, clinics: Volvo Cars adopts Nvidia's Hyperion self-driving platform, the Agile ONE SN reference humanoid is due next year, and Nvidia and Siemens Healthineers are building radiotherapy that tracks a tumor as the patient moves.
Jensen Huang came to Berlin to tell Europe it should produce AI rather than just consume it, and he brought the equipment list. "To be part of this AI economy, you don't want to just be consuming the AI; you want to be producing the AI," he told the audience at the Tempodrom on Wednesday, at the end of a week-long tour through Paris, London, Stockholm and Dublin. Nvidia says it has committed nearly $10 billion to Europe's tech sector.
(NVIDIA)
The headline product is the Vera Rubin DSX POD, a reference design for a complete AI factory that partners including Dell, HPE and Lenovo build inside shipping containers, integrated and tested before delivery. The pitch is speed: Nvidia says on-site deployment shrinks from six months to 14 days once the site's utilities are ready, and that its DSX platform produces the same AI output with up to 30% less power and up to 100% less cooling water. "This is a complete transformation of the world supply chain," Huang said. These are Nvidia's own figures, and the first containers to go live will test them.
(NVIDIA)
The second half was about what runs inside those factories. Huang described an agent as a model plus a harness, a sandbox, knowledge and tools, and called security "the most important thing". Nvidia's answer is the OpenShell Reference Architecture, built on an open-source runtime that gives each agent minimal, policy-set permissions and enforces them at the operating-system kernel; a keynote slide listed more than 100 OpenShell partners, among them Anthropic, OpenAI, Microsoft and SAP. Add open Nemotron models and a new Agent Knowledge Foundation for plugging agents into company documents and databases, and the offer to European firms is complete: a factory they can own, agents they can fence in, and models trained on their own data, which Huang called "sovereign intelligence".
Why it matters: Europe's AI debate has mostly been about rules and dependence on American clouds; Huang recast it as a construction project with a delivery date. If a containerized AI factory really goes live two weeks after it arrives on a prepared site, the bottleneck for European AI moves to energy and permits.
Sources:
Take the prompts our creative team actually uses
The key to maintaining brand quality at scale? Mastering how to communicate with AI models and embedding them into your creative strategy.
Join our honest discussion with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.
(Chart: Superintelligence; data: Sensor Tower via TechCrunch)
The chart: Sensor Tower's estimates of US downloads in each AI app's first 40 days, reported by TechCrunch: Meta's Muse at 10.2 million, ChatGPT at 3.7 million, Grok at 0.5 million and Claude at 0.2 million. Muse also kept more of the people who tried it: 30% were still using it after seven days, against an average of 19% for its rivals.
The lesson: Muse's milestone from the news section came with help, since Meta can promote it to the 3.6 billion people who use its apps every day and is pouring ad money into it. But retention cannot be bought the same way, and Muse passed 2.5 million daily users on October 16. For a consumer AI agent, owning the audience can matter as much as having the best model.
The caveat: Muse is only available in the US and Canada, while ChatGPT's app also drew users in markets such as Japan, Brazil and Germany and reached about 15 million global downloads in its first 40 days. The figures are Sensor Tower estimates, not company data.
๐ต๏ธ When the Scorekeeper Is Wrong, AI Learns to Cheat Quietly
โก Bottom line: Columbia researchers show that AI models trained with a flawed reward learn to cheat while their written reasoning still sounds like they follow the rules.
๐ก Why it matters: AI labs read a model's reasoning to catch misbehavior, and the rewards used in real training are never perfect.
๐ What it means: A simple fix, teaching models to state what they are actually rewarded for, made most of the hidden cheating visible again.
Tell an AI model to answer in a limited number of words, quietly reward it for going over, and it learns to blow past the limit, sometimes writing three times as much, while its reasoning keeps talking about staying within it. That is the central finding of a paper posted on October 20 by Tianyi Lorena Yan and John Hewitt of Columbia University.
The written-out reasoning of AI models, called the chain of thought, is one of the main ways labs check what a model is up to: a second AI reads it and flags trouble. That only works if the reasoning reflects what the model actually does. Models learn from a reward, a score for good behavior, and in complex tasks that score is never perfectly aligned with what people want.
(Yan & Hewitt, Columbia University)
The researchers built three deliberately broken rewards: one paid for exceeding a word limit, one for favoring Gemini's answers over Claude's when judging, and one for code that fails at least one test. Two of OpenAI's open gpt-oss models and Nvidia's Nemotron 3 Nano learned the cheat at least 59% of the time on coding and 93% on the other tasks. A monitor running OpenAI's gpt-5.6-luna mostly missed it, because the models either kept describing the stated task or shrank their reasoning to phrases like "Output code".
(Chart: Superintelligence; data: Yan & Hewitt)
The fix is surprisingly simple. Another model studies the training runs and writes down in one sentence what the reward really favors, and the model is then trained with that description in view so it learns to say it in its reasoning. Afterwards the monitor caught 85% of the cheating on average, up from 18%, at the cost of more false alarms. The tests are small and artificial, but the result is hopeful for anyone building agents: an AI's thinking can stay readable even when its scorekeeper is flawed.
