The superintelligence labs are bleeding talent to vertical AI, and lawyers just poached someone who knows how to make language models stop hallucinating case law.
The Summary
- Ethem Coban left Meta's Superintelligence Labs to launch a legal AI startup, currently raising $6M at a $25M valuation
- He spent two years training Meta's models to reason through legal questions and reduce hallucinations
- His thesis: law firms want to own their AI stack, not rent it, because proprietary case data becomes a moat
The Signal
Meta's Superintelligence Labs isn't just a research group. It's the place where the company works on frontier reasoning capabilities, the kind that determine whether an AI can actually think through a problem or just pattern-match its way to a confident hallucination. Coban's work there focused specifically on legal reasoning and hallucination reduction, two of the hardest problems in applied AI. Now he's taking that expertise to build tools that let law firms customize models around their own expertise.
The timing tells you something about where the market is. Legal AI companies like Harvey and Legora had first-mover advantage, but they're selling pickaxes in a gold rush. Coban is betting the next phase is lawyers who want to own the mine. The pitch to investors is simple: law firms have decades of case files, partner knowledge, and precedent analysis sitting in document management systems. That's training data other firms can't access. If you can turn that into a custom model, you've built something competitors can't replicate by throwing money at OpenAI's API.
"Companies will increasingly want to own their AI rather than license it from outside companies, because it gives them more control over their data."
This isn't just about data sovereignty or compliance theater. It's about competitive advantage. A BigLaw firm that trains a model on 50 years of merger filings in a specific industry has something a generic legal AI can't offer. The question is whether law firms have the technical chops to actually deploy and maintain these systems, or if they'll end up licensing them from companies like Coban's anyway.
The other signal here is what it says about Meta's retention problem at the frontier. Superintelligence Labs is where the company does its most advanced work. If researchers there are leaving to build vertical AI tools, it means two things:
- The applied AI market is lucrative enough to compete with Meta's compensation packages
- Researchers who've spent years on reasoning and hallucination problems think those problems are solvable enough to bet a startup on
The Implication
Watch for more defections from frontier labs to vertical AI. The researchers building foundational models know better than anyone where those models still break. That knowledge is worth more in a focused market where you can actually ship a product than in a lab chasing AGI benchmarks. If you're running a law firm, start asking what data you're sitting on that could train a model no one else has. If you're building legal AI, your competition just got a researcher who knows how Meta's best models think.