This has been the late summer/fall of personal agents. The bots I run in Grok Bot book my calls and help me research and think through business strategy. Muse books my flights, helps me keep track of family planning, and helped me pick the gift and plan the day for my fourth anniversary with my wife. A few months ago none of these products existed, and now I’ve been using them extensively for both work and home.
I'm clearly not the only one. Meta launched Muse on September 8, and more than 4 million people now use it every week, according to The Information. Grok Bot, which SpaceXAI built with Cursor, had about 418,000 weekly users five weeks after its beta opened, Bloomberg reported. OpenAI put its own version, called dots, into ChatGPT at the end of September, and Instinct, an agent you reach by text message, raised money at a $10 billion valuation in September with about 14 employees.
Using these every day got me thinking about what they learn while they work for me. In a long post about how Meta built Muse, one of its engineers wrote that your conversations with the agent, and "the tool calls and subagent handoffs that result," are "useful data for training new checkpoints" of the model underneath. The post calls that a good default, because every user "gets a better personal agent as we all collectively use the product and help the model understand the intricacies of human life." That last phrase stuck with me, because it names the one kind of data this market still hasn't figured out how to buy.
Work is well documented
Most of what I've written about in Trove has been work data, and there's a lot of it. Companies write down almost everything they do: Slack threads, email chains, support tickets, CRM notes. Over the last few months that record has turned into something you can sell. Google won the auction for Spirit Airlines' internal records, about 100 million emails among them, with a $10 million bid. micro1 said this month it will spend $1 billion over the next year buying records from companies. Startups that shut down are selling their old Slack and Jira archives through wind-down firms like SimpleClosure. (Next week I'll have an interview with one of the people who brokers these deals.)
Life outside work doesn't leave that kind of paper trail, at least not anywhere a company can sell it. When I plan a trip, pick a restaurant, figure out which doctor takes my insurance or remind myself to call someone back, there's no ticketing system logging how I made the decision. It happens on my phone, spread across messages, photos, maps, my calendar and the apps I open twenty times a day. And for most of us, that phone runs on Apple's software.
The iPhone has been a closed door
Apple has spent a decade making that data hard to get. It built its AI to run on the device where it could, and it said it didn't use people's private data to train its models. In 2021 it made apps ask permission before tracking people across other apps, and Meta's finance chief said the next year that the change would cost the company on the order of $10 billion in 2022. So the record of how more than 2.5 billion Apple devices get used day to day mostly stayed where Apple put it, out of reach of ad networks and AI labs alike.
Apple has also been slow to do much with that record itself. The smarter Siri it promised in 2024 slipped for more than a year. In January, Apple and Google announced that Apple's next models would be built on Google's Gemini, and the new Siri finally started rolling out in beta on September 14, in English only. Apple's capital spending in the first nine months of its fiscal year was $6.8 billion, according to its filings, while Alphabet, Amazon, Microsoft and Meta are each planning well over $100 billion this year. I wouldn't call Apple absent, but it's been careful and late, which in this race has mostly meant watching from the sidelines.
Apple seems to have noticed what it's sitting on, though. When you turn on the new Siri in iOS 27, a setup screen asks whether you want to "Share Audio and Text" to help improve Siri and Apple's models. Apple already collected some of that from people who opted in, but now it can train its foundation models on it. Even Apple has decided it needs this data.
The agents go around the phone
The new agents don't wait for Apple's permission. Instinct works over text messages and phone calls. The others live in their own apps and on their own cloud computers, and they reach into your life through the accounts you connect: Gmail, your calendar, your files, your bank, the card you saved at a store. OpenAI's dots can even connect to your laptop's camera, microphone and screen if you allow it. Every errand an agent runs leaves a record of what a person wanted and how it got done, what they searched for, what they compared and what they finally picked. When Muse helped plan my anniversary, it saw what I looked at, what I bought and where we went. That's the "intricacies of human life" Meta's post is talking about, and until this fall it mostly lived inside Apple's walls and the apps on our home screens.
Some of the errands are getting ambitious. Alexandr Wang, Meta's chief AI officer, shared this one on Tuesday:
On most of these products, that record trains the next model unless you go turn it off. I read the privacy policies for Muse, dots, Grok Bot, Hark and Instinct, plus the consumer versions of ChatGPT, Claude and Gemini, and the pattern was close to universal. Personal accounts are trained on by default, paying for a bigger personal plan doesn't change that, and business accounts are left out. Hark, from Figure founder Brett Adcock, is the one exception. It says it doesn't train on your content on any plan.
Meta has even put a price on what that permission is worth. Developers who use its Muse Spark model through the API can pay full price, or they can switch to what Meta calls Contributor mode and get "heavily discounted token prices, in exchange for permission to use your content to train future Meta models." The discount is more than 90%.
Meta's way back in
Which brings me back to that $10 billion. Five years ago Apple closed the door on Meta tracking people across their phones, and by Meta's own estimate it cost about twice the record $5 billion fine the FTC handed it in 2019. Muse is a way back to the same kind of information, and this time people invite it in. Meta says Muse conversations don't go to its ad systems, so for now the payoff is in the models. People connect their inbox and calendar and hand Muse their errands, and every errand teaches Meta's next model a little more about how people live. I don't think that's the only reason Meta built Muse, but it's hard to believe it isn't one of them.
What I'm watching now is what Apple does about it. Bloomberg reported in March that Apple was working on a way for outside AI models to plug into Siri, and The Information reported in May that Apple was weighing how to handle agent apps in the App Store. Neither has turned into a rule yet. If Apple opens the phone to these agents, it gets to set the terms. If it doesn't, the agents will keep working around it, and the most personal data on the phone will keep flowing to whoever runs your agent.
Partnerships & Deals
Compute, Energy & Infrastructure
Modal Labs raised a $750 million Series D led by Accel at a $15.75 billion post-money valuation. The company says its annual run rate has passed $1 billion.
Tenstorrent raised a $1.5 billion Series E at a $6.2 billion post-money valuation for its AI chips, with Fidelity, Morgan Stanley and ARK among the investors.
CoreWeave bonds sold off on October 14 on credit worries, with some yields near 15%. The CEO pointed to a backlog of about $104 billion.
TSMC reported record third-quarter profit, and its chairman said AI demand still exceeds what the company can supply.
EverWind Fuels proposed a C$10 billion AI data center and wind project in Nova Scotia on a 20-year lease with an unnamed "frontier AI developer." The first phase is 95 MW.
Nscale and Anthropic are in talks on a second $45 billion compute deal, according to Bloomberg. Nothing has been announced.
Data, Knowledge & Retrieval
Labelbox raised more than $100 million in a Series E, in equity and credit, led by Infinitum Partners. It says it turned profitable this year, works with over 90% of leading US AI labs, and now sells reinforcement-learning environments through a product called Horizon.
Google now pays more than 400 publications for content used in AI features in Google News, adding CNN, the Financial Times and Axios. It hasn't disclosed prices or whether the deals cover training. This is a separate program from the Search grounding payments we covered on October 5.
Meta updated its business terms to say it could use "any content" from a business account's websites to train its products. Legal experts told Nieman Lab that could let Meta use publisher content for free, and The Verge reports Meta has since removed the provision.
Character.AI filed for Chapter 11 on October 12. Disney bid $20 million for a non-exclusive license to some of its code and patents, and the bid reportedly leaves the user chats in the estate. Attorneys general from 39 states and D.C. have filed conditions for any buyer.
WindBorne won a US Air Force contract worth up to $49 million for weather data from its balloons. The Air Force pays only for the data it orders, and $1.7 million has been ordered so far.
Anthropic says new defenses against distillation have cut the data harvested from Claude by 90% to 99%. The figure is Anthropic's own.
Wiley said its cumulative AI licensing revenue has passed $100 million, according to a Seoul Economic Daily roundup of publisher deals.
Postmedia, Torstar, The Globe and Mail and The Canadian Press sued Anthropic in an Ontario court, saying it copied their journalism to build Claude, broke their sites' terms and keeps scraping their pages to answer user prompts. The suit was filed October 6 and was widely reported this week.
Spotify licensed music from Believe and its TuneCore service for an upcoming AI remix tool, according to Variety. Artists have to opt in, and no terms were disclosed.
Models, Training & Developer Tools
Meta released Muse Wonder, its first frontier model, and Muse Spark Open, an open-weight model that runs on two Mac Studios.
Elorian raised a $55 million seed round at a $300 million valuation, co-led by Striker, Menlo and Altimeter, and released Glimpse, a model that reasons with images.
Anthropic reportedly told investors its third-quarter revenue was $16.8 billion. Separately, it committed 5% of its shares to a public benefit fund.
Enterprise Deployment & Distribution
Workday agreed to acquire C1.ai, which manages permissions and logs for people and AI agents, and partnered with Anthropic so Claude and Workday's Sana can hand tasks to each other.
Anthropic licensed Opaque Systems' confidential-computing technology and hired its CEO and CTO. Opaque is winding down.
Cohere signed a letter of intent to acquire AISLE, which uses AI to find software vulnerabilities, and plans to form a unit called Cohere Cyber.
Industry Applications & Workflows
Ramp raised $1.85 billion at a $60 billion valuation, up from $44 billion in June.
Legora raised a $550 million Series E at an $8.8 billion post-money valuation, led by Atreides. The legal AI company says it has more than $200 million in annual recurring revenue.
Apple bought the assets of Samplab, a Swiss startup whose AI lets musicians rework recorded audio, and hired some of its staff. The deal surfaced on an EU acquisitions list this week.
Physical AI & Robotics
Stryker invested about $400 million in CMR Surgical, with an exclusive option to buy the rest of the company and US sales rights to its Versius robot.
DeepSeek is set to lead a seed round in Winverse AI, a robot-brain startup, at about a $400 million valuation, according to Bloomberg. The deal isn't final.
Mind Robotics, the Rivian-backed startup that trains its models in Rivian's factory, hired former Ford EV chief Doug Field as president.
Eka Robotics raised a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime. It trains its grasping model mostly in a physics simulator instead of on human demonstrations.
Dream Labs, founded this year by former Nvidia robotics researchers, raised a $90 million seed round led by B Capital and Kindred and launched Dream-1, a robot model sold through an API.