In partnership with
In Todayโs Issue:
๐ Meta's Muse Wonder joins the frontier race
๐ Microsoft scales its Maia chips to gigawatts
๐ฆ Anthropic finds early signs of self-copying AI
๐ก๏ธ Anthropic makes Claude much harder to copy
๐ค DeepSeek bets on brains for robots
โจ And more AI goodnessโฆ
โก The Signal
The frontier is crowded again: Meta now sits in the same tier as OpenAI and Anthropic.
On the Artificial Analysis index, Muse Wonder Full scores 56, above GPT-6 Astra and Claude Fable 5.1 at 53 each, with Anthropic's Claude Opus 5.5 still on top at 58. With the leaders this close, the contest moves to everything around the model. Meta is betting on memory for agents that run errands for you, Microsoft on its own chips to serve AI more cheaply, and Anthropic on stopping rivals from copying Claude's reasoning. The scoreboard is tight, but the user map is not: ChatGPT still draws more than half of all gen-AI web traffic, and Meta AI is not even among the services Similarweb names.
All the best,
Kim Isenberg
๐ Microsoft Takes Its Own Maia AI Chips to Gigawatt Scale
โฆMicrosoft will scale its in-house Azure Maia chips to multi-gigawatt deployments starting in 2027. Maia 200 already runs Microsoft Copilot, GitHub Copilot and Microsoft Foundry workloads, and the next generation, Maia 300, is built for long-context reasoning and agents. Microsoft says Maia 300 delivers 40% more tokens per megawatt than leading GPU systems, a vendor claim that has not been tested independently yet. It is obvious that they are looking for a fight with NVIDIA.
๐ tl;dr: Microsoft wants more of its AI to run on chips it designs itself, which could make answers from Copilot and similar tools cheaper to serve and less reliant on bought-in GPUs.
(Anthropic)
๐ฆ Anthropic Finds Early Signs of AI That Copies Itself
Anthropic's Frontier Red Team tested whether small open AI models can copy themselves onto cheap internet-connected devices, such as off-brand Android TV boxes, without human help. A 9-billion-parameter model got itself running on 8 of 9 hardware types, and in a simplified test network GLM-5.2 completed the whole chain, spreading to all 8 machines in 4 of 5 runs. The break-in step was simplified for the test, and Anthropic says plainly: "We donโt yet have a full rogue replicator."
๐ tl;dr: Small AI models can already handle most steps of spreading themselves across poorly secured home devices, and Anthropic projects they could get there by the end of this year if progress holds.
(Anthropic)
๐ก๏ธ Anthropic Makes Claude Much Harder to Copy
Anthropic says new safeguards have cut the reasoning text that rival labs can pull out of Claude by 10 to 100 times. This practice, called distillation, trains a cheaper model on a stronger model's answers; Anthropic names Chinese labs Zhipu and Moonshot, and says one network once ran more than 20,000 fake accounts at a time. The fix combines classifiers that spot extraction attempts, account bans often within two hours, and tighter access to Claude's step-by-step reasoning, and early evidence suggests distillers now harvest 90 to 99% less data.
๐ tl;dr: Anthropic is making it far harder for rival labs to copy Claude's thinking and train their own models on it for a fraction of the cost.
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
Mark Zuckerberg and Alexandr Wang sat down with Stripe co-founder John Collison on his Cheeky Pint podcast on launch day to explain how Muse Wonder was built. They walk through training ever larger models step by step, with Muse Spark built on a first pre-training run from late 2025, and Wang explains active memory, which Meta designed into both the model and the system around it. The most candid moment is Zuckerberg recalling the $100 billion to $200 billion commitment he made before seeing a single result: "Alex, are you sure this is going to work?"
"Iโm a strong believer of open source, thatโs the best way to drive efficiency."
โ Lip-Bu Tan, CEO of Intel
At a DBS panel in Singapore on Monday, Tan called Chinese models such as DeepSeek, Kimi and GLM "good enough" and said their adoption has been "very fast". Two of those labs turn up in today's Anthropic research: Zhipu's open model GLM-5.2 copied itself across a test network, and Zhipu and Kimi maker Moonshot tried to distill Claude.
Alexandr Wang shared a video of Muse arranging a washing-machine repair entirely on its own. That is the quiet promise of always-on agents: they sit through support hotlines and call-center loops so you don't have to. Europe, meanwhile, is still waiting for access.
Meta's Muse Wonder Makes It a Three-Lab Race
The Takeaway
๐ Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, above GPT-6 Astra and Claude Fable 5.1 (both 53) and behind Claude Opus 5.5 (58).
๐ Meta built the models around active memory, which carries what matters to you across conversations, and active focus, which keeps an agent working when a task gets hard.
๐ Developers can use Muse Wonder through the Meta Model API at $4/$20 per million input/output tokens; Wonder Full is only available in paid Muse Code plans.
๐ Meta's model card shows clear leads on long-running coding and legal work, while Wonder Full trails rivals on Terminal-Bench 4.0 and HealthBench Pro.
Meta is back at the top table. On October 12, Meta Superintelligence Labs released Muse Wonder and the larger Muse Wonder Full, its first models that compete head to head with OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1. On the independent Artificial Analysis Intelligence Index, a combined score across major benchmarks, Wonder Full reaches 56, three points above both and two behind Anthropic's Claude Opus 5.5. Fourteen months after Alexandr Wang took over Meta's AI effort, the frontier is a three-lab race. Wang's own launch post was blunter, calling Wonder Meta's "largest and most capable model yet" that "rivals the big bois".
(Data: Artificial Analysis. Chart: Superintelligence)
Meta built Wonder for agents that act on your behalf, starting with its Muse app. Two features carry the pitch: active memory, which keeps track of what matters to you across conversations, and active focus, which keeps a task going when it gets hard, tries other approaches and recovers from errors. Wang says Meta invested in memory "disproportionately" compared with other labs. The models were trained on Meta's Prometheus cluster in Ohio, with large-scale pre-training followed by reinforcement learning, after Zuckerberg had already committed a capital budget of $100 billion or more.
(Data: Meta model card via VentureBeat. Chart: Superintelligence)
Access is split. Muse Wonder is on the Meta Model API at $4/$20 per million input/output tokens, the same as Opus 5.5 and less than half of Astra's $10/$50. Wonder Full stays inside paid Muse Code plans, where Meta by default may use your prompts and code to train future models. Meta's full safety report is also still to come. On Meta's own model card, Wonder Full trails both rivals on Terminal-Bench 4.0. But where Meta aimed, it delivered: on SWE-Marathon, a test of long-running coding work, it scores 57.5 against 42.5 for Astra and 46.3 for Opus 5.5, exactly the kind of long, multi-step job a personal agent has to finish.
Why it matters: With three labs a few points apart, benchmark scores alone no longer decide who wins; Meta is betting that memory and a personal agent app will. Today's Graph of the Day shows how much ground it still has to cover with users.
Sources:
๐ Alexandr Wang on X
The chart: Similarweb's monthly snapshot of worldwide gen-AI website traffic shows how visits split between the big chatbots from October 2025 to September 2026. ChatGPT fell from 71.8% to about 51% in spring and has been climbing again since, to 56.4% in September. Gemini nearly doubled from 13.6% to 25.0%, and Claude more than quadrupled from 2.3% to 9.6%.
The lesson: The benchmark race and the user race are different contests. Claude's growth is real, but ChatGPT still draws more than half of all visits, and Meta AI is not among the seven services Similarweb names. That is the gap Muse Wonder now has to close.
The caveat: This counts website visits only. Apps, API use and Meta AI inside WhatsApp and Instagram are not included, so Meta's actual reach is larger than this chart shows.
๐ค DeepSeek Bets on Brains for China's Robots
โก Bottom line: DeepSeek is moving to lead a seed round that would value Winverse AI, a new Chinese lab building AI for robots, at about $400 million.
๐ก Why it matters: China builds robot hardware at scale but lags on the software that lets robots understand requests and handle real-world tasks.
๐ What it means: China's best-known AI lab is putting its model expertise into robotics, the next front in the US-China AI race.
China already builds humanoid robot hardware at scale. What it lacks is the AI that lets those machines understand a request and carry it out, and DeepSeek now wants to supply it. The lab behind China's best-known AI models is moving to lead a seed round for Winverse AI, a stealth startup building AI for robots, Bloomberg reported on Monday. The deal would value Winverse at roughly $400 million, with Meituan's investment arm Long-Z also committed, although valuation and investor list could still change.
Winverse was founded this year by Hu Han, who led Tencent's vision-language and embodied foundation model teams: the models that connect what a machine sees and hears with what it should do next. Hu holds a PhD from Tsinghua University and has worked at Microsoft Research Asia and Baidu.
(Bloomberg)
It is DeepSeek's second move into robotics in a few months. Before humanoid maker Unitree Robotics listed in Shanghai in August, DeepSeek took a 2.31% strategic stake and agreed to hold it for 36 months, three times longer than most other strategic investors. The new money comes from DeepSeek's own fundraising, which is close to raising at least $12 billion ahead of a planned listing in early 2027.
The gap it is aiming at is wide. US robot-brain startups Skild AI and Physical Intelligence are valued at $14 billion and $5.6 billion, backed by Nvidia and Alphabet. DeepSeek is now pointing its model expertise and fresh capital at exactly the part of the robot stack where China is behind.
(Data: Bloomberg via Quartz. Chart: Superintelligence)
Unify Your Teams and Tech Stack With HubSpot
Connect your customer data, teams, and tools without the hassle of complex integrations or lengthy setups. One easy platform gives marketing, sales, and service teams a unified customer view and the tools to turn it into growth.
Why HubSpot and what's new
โข Generate leads and automate marketing with Marketing Hub
โข Build your pipeline and close more deals with Sales Hub
โข Scale customer support and drive retention with Service Hub
โข Keep customer data clean, connected, and actionable with one, unified platform
Join 306,000+ in over 135 countries using HubSpot to grow their businesses.
See what a more connected approach to growth can do for you and your team. Get setup quickly and start checking off your hardest tasks.
