In partnership with
In Todayโs Issue:
๐ช Google breaks open quantum computing's magic-state bottleneck
๐ธ Anthropic's IPO investors talk down the price
๐๏ธ Nvidia's open model learns to see
๐ง Goodfire: AI models are more readable than feared
๐น๐ผ Taiwan makes AI chip smuggling to China a crime
โจ And more AI goodnessโฆ
โก The Signal
This week, big AI claims kept running into hard evidence.
Google showed that a quantum technique its researchers worked out on paper in 2024 works on its Willow chip, cutting errors about 40-fold, if only at small scale so far. At Goodfire, Nick Cammarata argued that AI models are far easier to read from the inside than his own early research suggested. Meta and Nvidia released open models on the same day, and overnight an independent index ranked both. Investors applied the same test to money: Anthropic grew 45% last quarter, yet some now talk about a $1.5 trillion valuation instead of $2 trillion. And Taiwan, after months of charging AI chip smugglers with forged paperwork, will make the smuggling itself a crime.
All the best,
Kim Isenberg
๐ธ Anthropic's IPO Investors Want a Lower Price
Some Anthropic investors are now discussing an IPO valuation of around $1.5 trillion, well below the roughly $2 trillion expected for November's share sale, Reuters reports. The $16.8 billion quarter we covered yesterday was up 45% and profitable on an adjusted basis, but growth slowed after revenue had more than doubled to $11.6 billion in the second quarter. The timing is tense: shares and bonds of AI cloud provider CoreWeave tumbled on Thursday as investors questioned the debt financing the data-center buildout.
๐ tl;dr: Anthropic is still growing fast and turned an adjusted profit, but the AI industry's bellwether IPO may price about a quarter below what investors expected.
(NVIDIA) Nemotron 3.5 Super's coding scores inside six popular agent tools, from Claude Code to OpenCode.
๐๏ธ Nvidia Releases an Open Model That Can See
Nvidia released Nemotron 3.5 Super VL on Thursday, an open model that reads images, documents and video as well as text.
It has 120 billion parameters, of which only 12 billion are active for each token, supports up to 1 million tokens of context, and ships with open weights, training data and recipes; the full-precision version runs on four H100 GPUs. Nvidia's own tests show it working inside the coding agents developers already use, such as Claude Code and Codex.
๐ tl;dr: Developers get an American open model they can download, inspect and retrain for agents that work with screenshots, PDFs and video.
(Goodfire) The essay's own picture of its claim: an orderly city rather than a tangled knot.
๐ง AI Models May Be Far Easier to Read Than Feared
Nick Cammarata, a co-author of the 2020 Zoom In paper that introduced the idea that neural networks fold unrelated concepts into the same neurons, now says the field misread its own key evidence. In an essay published Thursday by interpretability startup Goodfire, he reports that in a preliminary audit of Qwen3-32B, roughly 90% of neuron signals have one main job and only 2-4% are not yet understood; the automated study he started at OpenAI in 2023 judged about 1.7% legible with the tools of the time. It is an essay with a technical paper still to come, not a peer-reviewed result.
๐ tl;dr: If AI models are largely readable neuron by neuron, researchers can check what a model is doing inside instead of judging it only by its answers.
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.
๐ฌ Watch This
Google Quantum AI's three-minute explainer is the quickest way into today's Featured Story. Research scientist Emma Rosenfeld explains why a useful quantum computer will need "somewhere between a million and a trillion magic states and T gates", how cultivation filters noisy states into clean ones, and why the team ended up using the technique to measure itself. Craig Gidney, who co-wrote the original theory, called the video a good summary of the experiment.
Microsoft has long relied on OpenAI for its frontier models. Now its own lab says it has trained base models in the 3-trillion-parameter class and matched DeepSeek-V4-Pro-Base on held-out code with 12x less training compute. The chat-ready versions are still to come.
OpenAI is turning Apple's trade-secret lawsuit around. In a new court filing, it alleges that an employee who returned to Apple after six months sent more than 100 files on an unreleased OpenAI device to personal accounts and later presented its product plans to senior Apple executives; Apple says it found no evidence for the claims.
Google Breaks Open One of Quantum Computing's Biggest Bottlenecks
The Takeaway
๐ Google Quantum AI showed magic state cultivation working on its Willow chip, in a paper published in Science on October 15.
๐ Cultivated magic states had about 40x fewer errors than uncultivated ones, at a record 99.99% logical fidelity.
๐ Magic states are the consumable ingredient of useful quantum programs; some estimates put up to 95% of a machine's capacity into making them.
๐ Next comes scale: if cultivation holds in much larger codes, practical quantum computers could get substantially smaller.
Google has shown on real hardware that one of quantum computing's most expensive overheads can be cut down. In a paper published in Science on Thursday, Google Quantum AI reports that magic state cultivation, a method its researchers Craig Gidney, Noah Shutty and Cody Jones developed on paper in 2024, works on the company's Willow processor. Magic states are what let a quantum computer run the operations that take it beyond an ordinary computer, and each one is used up once spent. Researchers have expected since 2004 that producing them would be a bottleneck: some estimates had up to 95% of a machine's capacity going into making them.
(Google) Willow, the processor that ran the experiment.
Cultivation works like a filter: the chip injects a noisy magic state into a small patch of qubits, checks it in repeated rounds and throws away every attempt that fails, until only very clean states remain. Getting there took months of detective work. Google's Emma Rosenfeld kept seeing data that "seemed to defy the laws of physics" until she traced it to drifting lab equipment and temperature changes that broke the team's usual measurement method; the fixes included a new benchmark that uses cultivation to measure itself. The result was about 40x fewer errors than uncultivated states and a record logical fidelity of 99.99%. The team also moved a cleaned state from the small code it grew in into a larger surface code, the more established error-correction layout.
(Google Quantum AI, arXiv:2512.13908) Red: error without cultivation, lowest at the target T state. Cyan: with cultivation, about 40x lower.
The demonstration is still small. The filter keeps only about 8% of attempts, and Gidney noted on X that the hardest step of his theory, moving states into a code of roughly 500 qubits, does not fit on today's chip:
Still, a fix that existed only on paper is now a working experimental technique, and cheaper magic states could substantially shrink the quantum computer needed for practical jobs.
Why it matters: A large share of a future quantum computer's hardware is expected to go into correcting errors and producing magic states rather than computing. Proving on a real chip that this overhead can shrink brings practical machines closer.
Most AI Tools Weren't Built for This
B2B customer issues move across teams and systems, not through a single chat window. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, breaks down where AI tools fall short in B2B service and what to ask before you invest.
(Chart: Superintelligence; data: Artificial Analysis, October 16)
The chart: Artificial Analysis combines ten tests of coding, agentic work, science and knowledge into one Intelligence Index. Meta's Muse Spark Open, released Thursday, enters at 39, the highest score of any American open-weights model, well ahead of Nvidia's Nemotron 3.5 Super VL from today's News at 26. For scale, the best closed model, Claude Opus 5.5 at maximum effort, scores 58.
The lesson: America's best downloadable model now sits 13 points above Nemotron and within a point of Muse Spark 1.2, Meta's own flagship in August. Companies that want to run a strong model on their own hardware instead of through someone else's API now have a much better American option: Meta says the compressed version runs on four Blackwell Ultra GPUs or on two Mac Studios.
The caveat: China's open models still lead on agent work. GLM-5.3 (max), the closest Chinese model by size, scores 45, and Artificial Analysis says its entire six-point lead comes from the four agentic tests, where Muse Spark Open ranks last among the nine best open-weights models. Its license also requires a separate deal with Meta for model-hosting businesses above $20 million in annual revenue.
๐น๐ผ Taiwan Will Make Smuggling AI Chips to China a Crime
โก Bottom line
Taiwan will make shipping restricted AI chips and servers to China a criminal offense, closing a loophole smugglers relied on.
๐ก Why it matters
Taiwan builds much of the hardware China wants most, and until now diverting it there was not a crime.
๐ What it means
The chip war is moving from Washington's rulebook into allies' criminal law, raising the stakes for every middleman.
Taiwan is about to make it a crime to ship restricted AI hardware to China. Economic Affairs Minister Kung Ming-hsin told Bloomberg that Taipei will add AI servers and cutting-edge chips to its list of controlled items; once that takes effect, sending them to China becomes a criminal offense under the island's Foreign Trade Act. Bloomberg reports the step as part of Taiwan's alignment with US restrictions under its trade deal with Washington.
(Lam Yik Fei/Bloomberg) Nvidia's booth at the Computex 2026 expo in Taipei.
Until now, Taiwanese prosecutors had to work around the law. In August they charged nine people, including an Nvidia employee and two Super Micro staff, over 74 Super Micro servers with Nvidia chips that reached China, 50 of them via Indonesia and eight via Japan; 56 more were seized at the border. The charges were breach of trust and document forgery, because exporting the chips to China was not itself a criminal offense in Taiwan.
The money involved is large. In March, US prosecutors charged Super Micro co-founder Wally Liaw and two others with trying to divert more than $2.5 billion of Nvidia AI technology to China. Chris McGuire of the Council on Foreign Relations called chip smuggling a "really significant problem" in Taiwan and Southeast Asia and said allies need to align with the US "on all of these policies and also legal authorities."
(Coast Guard Administration via Taipei Times) Super Micro AI servers taped off at a warehouse in Keelung on May 20.
Kung gave no start date, and the amendment still has to pass. But Taiwan is turning Washington's chip policy into its own criminal law, covering shipments to China as a whole. For the brokers, forwarders and server builders who route hardware through the island, the risk moves from a forged-paperwork case to a direct charge for the shipment itself.