Happy Thursday.
The data this week: When your team fixes a mistake in an AI app, that fix could end up training someone's AI model. Two startups that help software companies like Harvey build their own AI models raised $395 million this month, and both say the edits people make inside an app are valuable training data. Harvey, a customer of both, says it never trains on customer data.
Underneath: Anthropic started pitching its IPO to investors, The New York Times and three other papers sued Brave over a copy of the web it sells to AI companies, Korean publishers agreed to supply their books to five AI data firms, and Meta shelved a plan to scrape business users' websites.
The Lead
The big picture: AI labs like Anthropic pay outside firms to check their data and test their models. In Anthropic's IPO filing, that bill rose by $757 million in the first nine months of this year compared with the same stretch last year. Now the software companies that build on those models, such as Harvey in legal work, are starting to train AI models of their own, mostly because renting someone else's model gets expensive. One Harvey tool answers questions across thousands of documents at once, so a single job can mean hundreds of thousands of calls to an AI model, and Harvey pays for every one.
Driving the news: On November 2, Applied Compute raised $350 million and Trajectory raised $45 million to help companies like Harvey do that training. Both list Harvey and Mercor as customers. In August, Harvey said it worked with Applied Compute to train its own model for that tool, cutting the cost of each answer by 54.8% compared with Anthropic's Claude Sonnet 5. Harvey also said it built the model without any customer data, using public legal documents, practice questions written by AI and checks by human experts.
Why it matters: Harvey's vendors are pitching a different idea. Trajectory says the retries, corrections and edits people make inside an app are exactly what a model should learn from, and that nearly all of it gets thrown away today. Applied Compute says the work a model does for customers becomes the material for its next round of training. Some companies are already heading that way: Clay's CEO said the company is testing a model that learns from its users. That leaves a simple question without a clear answer: who owns a correction? When a lawyer fixes an answer in an app like Harvey, the fix holds the client's information and the law firm's judgment, but it's usually the app company that keeps the record.
What's next: The rules are still unclear. Applied Compute's privacy policy says it may use customer data that has been pooled or stripped of names, if the customer's contract allows it. This month, data protection regulators from Spain, Portugal and Latin America said a vendor needs clear permission to train its own models on a client's personal data, even with names removed, PPC Land reported. Their guidance isn't binding. Until contracts spell this out, any company whose staff spends hours fixing an AI tool's work should ask the vendor three questions in writing: Do our edits train a model? Whose model is it? And what happens to it if we cancel?
Read Harvey's post on training its own model.
See Trajectory's case for learning from user corrections.
Read PPC Land's report on the regulators' guidance.
Data, Knowledge & Retrieval
Anthropic signed its first content licensing deal with the Associated Press for real-time election results, Adweek reported, though neither company commented or shared terms.
The New York Times and three other papers sued Brave over the 30-billion-page search index it sells to AI companies, naming Anthropic, Perplexity, Mistral, xAI and Cohere as customers but not as defendants.
The Korean Publishers Association signed up five AI data firms, including Selectstar, Crowdworks and Flitto, to supply published books as AI data, and each firm must return or destroy the data when a project ends.
Meta indefinitely paused terms that would have let it scrape business-account users' sites, canceling their November 20 start date.
Sentra raised $150 million led by Capital One Ventures for software that finds sensitive data and limits what people and AI agents can reach.
Models, Training & Developer Tools
Anthropic started pitching its IPO to investors, offering 825 million shares at $70 to $78 each, which would raise up to $64.4 billion when it lists on Nasdaq as ANT.
OpenAI reportedly has about $11 billion of a new round in hand at a $1.25 trillion valuation, with up to $9 billion more possible while the round stays open, according to Bloomberg.
Fireworks hired Meta strategy chief Dave Wehner as CFO and is in early talks to raise money at a $30 billion to $40 billion valuation, according to Business Insider.
Meta bought Conductor, which makes an app for running several coding agents at once, for an undisclosed price.
Compute, Energy & Infrastructure
Anthropic said in its IPO filing that it has committed to spend about $895.4 billion, nearly all of it on computing power and the data centers around it.
Marvell agreed to buy Lumentum in a pending all-stock deal worth about $115 billion, counting Lumentum's debt.
fal raised a $1.3 billion Series F at a $16 billion valuation led by Bessemer, doubling its April valuation.
Dropbox agreed to buy Tigris Data for $350 million in a deal that hasn't closed yet, adding storage that keeps AI data close to where it's used.
Enterprise Deployment & Distribution
Traversal raised a $120 million Series B at a $1 billion valuation, co-led by Kleiner Perkins, Redpoint and Khosla Ventures, for agents that diagnose and fix software outages.
Nooks took a $500 million commitment from General Catalyst's Customer Value Fund, which pays for Nooks to win new customers in return for a share of the revenue, without taking a stake in the company.
Physical AI & Robotics
Tutor Intelligence raised a $75 million Series B led by Andreessen Horowitz for hourly-rental robots with remote operators, whose finished jobs become training data for its robot models.
GenRobot AI disclosed a $100 million round for robot-training data that closed in the third quarter.
Perceptron published a price list for labeling video shot by robots and head-mounted cameras, from $9 per hour of video up to $29 with detailed hand and activity labels.
Health
nPhase raised a $30 million Series D led by Viking Global Investors for clinical-trial data software.


