In partnership with
In Todayโs Issue:
๐ญ Kimi K3.1 Pro's strong debut, overshadowed by an alleged ad brief
๐บ A hacker says Samsung TVs turn viewing into ad data
๐ฆ Banks find AI agents inside customer accounts
๐ผ What Anthropic's IPO filing should finally reveal
โจ And more AI goodnessโฆ
โก The Signal
This week, the hardest thing in AI to measure is trust.
On Friday, Moonshot shipped Kimi K3.1 Pro, which placed fourth on Artificial Analysis's intelligence index, and within a day an alleged, still unauthenticated brief telling influencers to hide paid partnerships had taken over the conversation. The same week, a researcher said a Samsung TV turns living-room sound into ad data, and US banks found AI agents acting inside customer accounts, sometimes beyond what users had asked. Now Anthropic may finally publish the filing that shows what a frontier lab costs and whom it depends on. The models keep getting better and cheaper; what readers and buyers need now is a clear view of who is paying whom.
All the best,
Kim Isenberg
(log4j via X)
๐บ Hacker Claims Samsung TVs Turn Viewing Into Ad Data
A security researcher who rooted their own 2025 Samsung TV says it collects audio, screenshots and viewing habits to sharpen ad targeting. Known as log4j, the researcher went digging after noticing "weirdly targeted ads" and reported code in Samsung's Tizen software that categorizes sound, captures screenshots and runs a neural network logging what you watch and which ads you skip. On sets without a microphone, it allegedly borrows a connected soundbar's. OpenAI later banned the researcher's account for cyber abuse, though log4j says its tools were used only for local reverse engineering.
๐ tl;dr: The TV in your living room may be quietly turning what you watch and hear into advertising data.
๐ฆ Banks Find AI Agents Inside Customer Accounts
Major US banks have found third-party AI agents operating inside customer accounts, and some appeared to act beyond what users had asked, Axios reports. The agents log in with customers' own credentials, which clashes with the rules at most banks, and banks often cannot tell an authorized assistant from a rogue one. Morgan Stanley is blocking outside agents, while JPMorgan, Bank of America and Citi are restricting them, with most leaning toward read-only access.
๐ tl;dr: Banks are drawing the first line for AI agents that touch your money: they may get to read your account, but not move it.
๐งฉ Nanbeige's 33B Model Punches Above Its Weight
Nanbeige LLM Lab, the AI team of Chinese hiring platform BOSS Zhipin, has released Nanbeige4.3-33B, an open model built to work as an agent on coding and office tasks. In the lab's own tests against open models of similar size, including Alibaba's Qwen3.8-27B and Google's Gemma4-31B, it leads most agent benchmarks. In YC-bench, a simulation of running a startup, it ended with about $1.03 million versus $641,000 for Qwen. It ships under the permissive Apache 2.0 license.
๐ tl;dr: A mid-sized open model that companies can run on their own servers now beats its direct rivals at multi-step coding and office work, by its makers' tests.
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
Why did Kimi K3.1 Pro get a mixed reception despite those benchmark scores? Researchers Nathan Lambert and Florian Brand open their latest Interconnects podcast with exactly that question. They point to weaker results on longer, more exploratory tests, the cost of serving a 3-trillion-parameter model and the alleged undisclosed influencer campaign, then turn to Moonshot's decision to weaken the model's hacking skills and what AI is doing to cybersecurity.
"2026 is obviously at this point sold out. 2027 is largely sold out. [...] A good chunk of 2028 is sold out."
โ Andy Jassy, CEO of Amazon, on the a16z podcast
Amazon plans to spend about $220 billion on infrastructure this year and says it still cannot meet AI demand. It reports earnings on Thursday. Anthropic alone has committed about $110 billion to AWS, according to its draft IPO filing.
OpenAI is reportedly winning over Anthropic's customers with discounts. Replit moved most of its AI spending to OpenAI in a deal worth over $100 million a year at more than 30% off, after Anthropic ended its own discount, The Information reports. CEO Amjad Masad says users can still pick Claude.
Kimi K3.1 Pro Earned Its Praise. Then Came the Alleged Influencer Brief.
The Takeaway
๐ Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, placing fourth behind three American models.
๐ Its biggest leap is in multi-step agent work: 67% on Terminal-Bench 4.0, up from 13% for K3.
๐ Open weights are promised by November 4, which would make it the top-scoring open-weight model on that index.
๐ An alleged campaign brief told influencers not to label paid posts; it is unauthenticated, and Moonshot has not responded.
Moonshot AI released its strongest model yet on Friday, and by the weekend the talk was about whether paid praise had been passed off as independent. Kimi K3.1 Pro builds on July's K3, a 2.8-trillion-parameter model, with far more reinforcement learning (training by trial and feedback) across "hundreds of downstream domains", from coding and 3D design to financial research. The gains show up where an AI has to act over many steps: on Terminal-Bench 4.0, a test of coding agents working in a real command line, Artificial Analysis measured 67%, up from 13% for K3. Prices fell by a third, to $2/$10 per million input/output tokens.
(Artificial Analysis)
Overall, Artificial Analysis ranks it fourth with 55 points, behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 and Meta's Muse Wonder Full (56 each). Once the weights arrive, promised by November 4, the gap between the best open and closed models shrinks from 12 points to 3 on that index, and Moonshot says it deliberately weakened the model's hacking skills before that release. Then X creator Elshayib, who says they declined to take part, posted screenshots of an alleged Kimi campaign brief for influencers. It asked partners to send drafts for approval and said: "Please DO NOT put on the paid partnership label."
The suggested talking points went further than the facts. A model that trails Opus by three points is not "literally Opus 5.5 open-source", and at half of Opus's token price it is not five times cheaper. The brief has not been authenticated, and Moonshot had not addressed it as of Sunday. Researcher Nathan Lambert called the alleged practice "bad form". The pity is that K3.1 Pro never needed the help.
Why it matters: Open-weight models are now close enough to the frontier that companies pick them on independent evidence. Undisclosed paid posts would contaminate exactly that evidence, at the moment Kimi had earned attention on its merits.
Sources:
When your support agent gets it wrong, who's accountable?
Every agent handling refunds, account changes, or tickets needs three answers: what it can pick up, what it's allowed to do, and when it hands back. Join the live conversation Oct. 27, 10 a.m. PT, part of Running Agents in Customer Work. Register today, recordings included.
Source: Apollo Daily Spark, Torsten Slok, Oct. 24, 2026
The chart: Apollo's chief economist Torsten Slok plots profit margins for the Magnificent 7 (Apple, Microsoft, Alphabet, Amazon, Meta, Nvidia and Tesla) against the other 493 companies in the S&P 500 since 2015. Big Tech's margin has jumped to about 30% this year. The rest of corporate America has hovered around 10% for five years.
The lesson: If AI is paying off anywhere yet, these margins point to the companies that sell it: the cloud capacity Andy Jassy calls sold out, and the hundreds of billions Anthropic has committed to Google, Amazon and Microsoft. The chart cannot separate AI from everything else those firms earn, but for the other 493, Slok finds no sign of a payoff yet.
The caveat: Big Tech's margins also reflect advertising, iPhone sales and cost cutting. Flat margins outside tech do not prove AI is failing there; gains may first show up as lower prices for customers, or simply arrive later.
๐ผ Anthropic Is About to Open Its Books. Here's What to Watch
โก Bottom line: Anthropic's IPO prospectus could become public as early as today, after weeks of leaked drafts.
๐ก Why it matters: It will be the first official look at what it costs to build and run a frontier AI lab.
๐ What it means: Investors will learn whom Anthropic depends on, who pays its bills and who really controls it.
The most anticipated tech IPO in years could put its paperwork in public as early as today, The Information reported on Friday. Anthropic filed its prospectus, known as an S-1, confidentially in June, and leaked drafts have already revealed a lot. The Wall Street Journal, which reviewed one, reported $16.3 billion in first-half revenue and an $80.7 billion net loss, most of it a $72 billion accounting charge rather than cash out the door.
Anthropic CEO Dario Amodei (Getty Images via The Information)
The bigger number is what Anthropic has promised to spend. The draft lists about $450 billion in cloud contracts over seven to ten years, most of them non-cancelable, including $111 billion with Google Cloud, $110 billion with Amazon Web Services and $31 billion with Microsoft. Google and Amazon are also major shareholders, so the public filing should show how much money flows in each direction.
The next question is who pays the bills. About a fifth of first-half revenue came from just two customers, according to published reports. Historically, most income has come from developers buying Claude through its API, and those customers can switch quickly when a rival cuts prices, as Replit's reported move to OpenAI shows.
Dario Amodei and Sam Altman (The Information)
Then there is control. Anthropic plans to give its seven co-founders, including CEO Dario Amodei, a combined 50.1% of the voting power, alongside a trust that appoints most of the board. Buyers will want to know how those two centers of power settle a dispute and how long insiders must wait before selling.