In partnership with
In Todayโs Issue:
๐ Anthropic's drug programs post their first lab results
๐ฅ What it really costs to serve a world model
๐ Apple builds an AI assistant into Pages, Keynote and Numbers
๐ง Why AI memory often loses to a simple search
โจ And more AI goodnessโฆ
โก The Signal
AI is starting to earn its keep where results can be checked: on a lab bench, on a GPU bill and on a test nobody could study for.
Anthropic's drug update is today's clearest case. Claude redesigned part of a cancer therapy, proposed tens of thousands of antibody candidates and caught manufacturing errors that human reviewers had missed, and the companies published technical reports behind the results. Nebius ran the same kind of accounting for world models, measuring how much usable video a GPU server really produces per hour. Artificial Analysis now builds most of its intelligence ranking on private tests that labs cannot train on. And a new study shows that AI memory, a feature most assistants now advertise, often loses to a plain search. The industry is learning to grade itself on results you can measure, and those results are getting good.
All the best,
Kim Isenberg
๐ฅ Nebius Finds the Trade-Off in Serving World Models
Nebius found that running eight single-GPU copies of NVIDIA's Cosmos 3 Super world model on one server, instead of one copy spread across all eight GPUs, produced 42.5% more usable video per hour, but each clip took 5.5 times longer to arrive. World models generate realistic video of how a scene unfolds, which robotics and self-driving teams use as training data. On an HGX B200 node, the single large copy returned a 720p clip in 68.5 seconds on average and the eight small ones in 378 seconds, but at about 30% lower cost per second of video.
๐ tl;dr: Teams generating synthetic training video for robots can get far more footage out of the same hardware, as long as nobody is waiting on any single clip.
(Apple)
๐ Apple Builds an AI Assistant Into Pages, Keynote and Numbers
Apple is adding an AI assistant, built on its own models and OpenAI's, to Keynote, Pages and Numbers. It designs and animates slides, drafts and restructures documents, and suggests charts or builds tables from spreadsheet data. On macOS 27 the three apps also support the Model Context Protocol (MCP), the open standard that lets outside AI agents read and edit your files, and Keynote gets a new agent-readable .keyx format. The beta is for Apple Creator Studio subscribers on Apple Intelligence devices, in US English.
๐ tl;dr: Apple's office apps can now draft and design for you, and third-party AI agents can work directly inside your documents.
(NVIDIA)
๐ค NVIDIA Releases One Motion Model for Game Characters and Robots
NVIDIA has released MotionBricks, a model that generates lifelike movement in real time for both animated characters and a Unitree G1 humanoid robot, together with its code and pretrained weights. Given a character's last four poses and up to four target poses, it fills in the motion between them at 30 frames per second, and reusable "Smart Primitives" turn moves like vaulting, ducking or sitting into blocks that snap together without retraining. The standalone repository, published last week and based on a SIGGRAPH 2026 paper, ships a runtime and demo scenes for NVIDIA GPUs.
๐ tl;dr: Game studios and robot builders get a ready-to-run motion engine, and the same code drives a digital human and a real humanoid.
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
Journalist Kevin Roose interviewed about 150 people across the AI industry for his new book, The AGI Chronicles, and in this Decoder conversation with The Verge's Nilay Patel he explains what that reporting taught him. His sharpest claim is that belief itself is a competitive edge: OpenAI's willingness to overspend on compute a year ago, which even Anthropic's Dario Amodei criticized, now leaves it, in Roose's words, "the only major AI company that is not severely compute constrained." He also describes people in the industry who "lament the fact that the chatbots came before the cancer cures," a fitting companion to today's lead story.
"In short order, we will be officially BANNING the term, 'AI/Artificial Intelligence,' in all relevant documents, and everything else, due to its inaccuracy."
โ President Donald J. Trump, on Truth Social
He wants official documents to say "SI (Super Intelligence)" instead, and asked when The Wall Street Journal would follow. The post did not say how or when the ban would take effect. As a newsletter called Superintelligence, we have no complaints about the name.
OpenAI, one of Nvidia's biggest customers, plans to buy Google's TPU chips, The Information reports, citing one person familiar with the matter. Neither company has confirmed it and no terms are known, but the deal would hand Google a marquee buyer for chips it only began selling directly in April.
Anthropic Puts Claude to Work on New Medicines
The Takeaway
๐ Anthropic scientists partnered with six drug developers, from Generate Biomedicines to Bristol Myers Squibb, and published technical reports for most of the projects.
๐ Anthropic bought Mithril Medicine, a one-person biotech, and now runs its own programs for rare childhood muscular dystrophies.
๐ At BigHat Biosciences, Claude's proposed changes produced antibody fragments that bound their targets more than 10 times as tightly.
๐ Next up are human trials: Anthropic expects its first own drug candidate to be ready for testing in people as soon as late 2028.
Anthropic is now developing medicines with Claude, and it has lab results to show for it. Over the past several months its scientists worked with six drug developers, and Anthropic bought Mithril Medicine, a one-person biotech whose founder, Cody Tranbarger, used Claude to manage contracts, logistics and experiments with 11 outside labs. That setup took a possible treatment for a rare childhood muscular dystrophy from idea to promising mouse results in under 18 months.
(Anthropic technical report: in one week, Tranbarger and Claude turned a 341-page data package on 113 antibody designs into one plan across five contract labs)
Most of this work is still preclinical, meaning it has not been tested in people. But in a small pilot study, mice treated with Mithril's antibody developed about twice the peak muscle force of placebo-treated mice. The partner results point the same way. At Generate Biomedicines, scientists used Claude Code to redesign part of a CAR T-cell cancer therapy; one design was about 70% shorter and let immune cells make roughly 10 times as many cancer-spotting receptors. At Divergence Labs, Claude largely on its own produced about 35,000 antibody designs in a month for a food-allergy drug, and three appeared to pull the allergy molecule IgE off immune cells better than the standard medication.
(Anthropic technical report: in the pilot study, mice given the "2+2" antibody reached about twice the peak force of the placebo group; 7 vs. 4 mice)
Some of the biggest wins came from paperwork. At Enveda, where chemists spent up to 75% of their time on manufacturing records, Claude reviewed 1,061 pages, caught all 11 issues human reviewers had found plus seven they had missed, and cut review time by 89%; one regulatory draft took four hours instead of about three months. Anthropic argues the bottlenecks are spread across the whole pipeline, so it wants Claude in every step, from choosing a drug target to clinical-trial planning, which it plans to expand with Bristol Myers Squibb next year.
Why it matters: Drug development is slow because of thousands of small steps, from planning lab work to writing regulatory filings. If Claude can compress those steps, investors can afford to back riskier ideas, and patients with rare diseases the industry has long overlooked could get treatments sooner.
Sources:
9 Plays. One Mission: Keep More of Your Holiday Profit.
Holiday profit rarely disappears all at once. Deeper discounts, rising ad costs, fulfillment fees, and returns take their share little by little.
Get 9 Plays to Defend Your Margin This Holiday Season, tactical advice from ecommerce experts to uncover where profit gets lost and what you can do to keep more of it this Q4.
The chart: Artificial Analysis rebuilt its Intelligence Index, a combined score across nine tests of coding, science, knowledge and office work, as v5. For the first time, 60% of it runs on private test sets or held-out answers that labs cannot train on. Claude Opus 5.5 (max) leads at 58, ahead of OpenAI's GPT-6 Astra (57) and GPT-6.1 Sol (55); Google's Gemini 4 Argon follows at 50.
The lesson: On the tougher index, the field spread out. Anthropic and OpenAI hold all top five spots, and the gap to the best open-weights model grew from 3 points under v4 to 12. Monday's lead, Kimi K3.1 Pro, which placed fourth with 55 on the old version, now leads that group at 46, but its weights are not out yet; among models you can actually download, DeepSeek V4.1 Flash and GLM-5.3 top out at 32, below Claude Haiku 5.5.
The caveat: v5 scores cannot be compared with v4 numbers, so a model that looks weaker than last month may simply face a harder exam. Price also varies widely: among the ten best models, cost per index task runs from $0.96 (GPT-6.1 Terra) to $13.04 (Claude Fable 5.1).
๐ง Giving AI Agents a Memory Often Makes Them Worse
โก Bottom line
In a large new test, most dedicated AI memory systems answered fewer questions correctly than a plain search through the raw chat history.
๐ก Why it matters
Major assistants now advertise memory, and a system that quietly drops details is least reliable exactly where users trust it most.
๐ What it means
The best designs will keep a path back to the original conversation instead of relying only on their own summaries.
Tell an AI assistant that a taxi from the airport will cost about $60, then ask weeks later how much the train would save you. One memory system tested in a new paper had kept only the note "Prefers cost-effective transportation options due to budget concerns." The number was gone. The example comes from Memory Tax, posted on Oct. 25 by researchers at the University of Illinois Urbana-Champaign, Tsinghua University and other institutions. They pitted 11 memory systems, the add-ons that let AI agents remember past conversations, against the simplest alternative: searching the raw chat history directly.
(Memory Tax paper: on LongMemEval with Qwen3.6-27B, plain search in two rounds, top left, beats every memory system at a fraction of the token cost)
Plain search usually won. On LongMemEval, a standard test of long-term chat memory, with OpenAI's GPT-5.6 Luna answering, simple search scored 82% in one pass and 90% in two; the best memory systems reached 84% and the weakest 63%. The memory systems were also far more expensive: building the search index took about seven minutes, while the memory systems needed roughly 12 to 127 hours of processing for the same 100 conversations.
(Memory Tax paper: blue marks questions only the memory system got right, red those only plain search got right; four of the five systems shown lose more than they gain)
The authors call these losses a memory tax. To stay organized, memory systems rewrite, compress and split up conversations, and details get lost or buried along the way. When the team traced individual facts, one system had fully preserved only about 39% of them, and even facts that survived were often harder to find than the original sentence. Memory still has upsides: the systems did answer some questions plain search missed, but for most of them the losses outweighed the gains.
The fix the authors suggest is simple: keep a way back to the source. When memory systems also searched the raw history and merged both sets of results, scores improved across several of them. For anyone building assistants, the lesson is that memory works best as an index to the original conversation.
