A model is only as good as its data, and we’ve long since exhausted the internet. From here on out, model progress is gated by data production. ’s Brendan (can/do) joined us at our Sovereign AI event to talk about how RL environments get built, and why your data... might be your real moat: 00:00 Introduction 00:47 A short history of the data market: crowdsourcing to agentic data 02:29 What an RL environment is: worlds, apps, tasks 03:57 Why only humans can measure the frontier 05:35 Building verifiers is the hard part 06:44 Walkthrough: a real legal RL environment 08:18 Leaderboards — and what open weights change 09:45 Post-training results on Apex Agents 11:17 Three ways companies buy data 12:49 Q&A: How do you price data? 14:17 Q&A: What "data quality" actually means 16:42 Q&A: The misunderstanding about synthetic data 18:17 Q&A: Why RL environments now — and what comes after 21:20 Q&A: Can you scale rubric generation with models? 23:00 Q&A: RL environments for cyber defense 25:33 Q&A: Build data in-house or partner?show more

@mercor @BrendanFoody Brendan crushed

@mercor @BrendanFoody DATA FACTORIES!!!! thank you @BrendanFoody for joining us to demystify RL environments and more ☺️

Model training is a four-dimensional problem: Data Complexity, Boundary Conditions, Training Dynamics, and Training Stability. Data and RL environments largely define the first two: what the model must learn and how the learning signal is constructed. The other half is how learning is applied over time and whether manifold homology is preserved.

@mercor @BrendanFoody Sovereign AI has multiple definitions. Here's how I think of them.

@mercor @BrendanFoody This was very interesting thanks for sharing, great watch

@mercor @BrendanFoody 'we exhausted the internet' just means the data team gave up scraping before finishing the job. rebrand the shortcut as a wall.

@mercor_ai @BrendanFoody Do you think this is true for all typew of models? Or is this really just data for the frontier labs?

@mercor @BrendanFoody The legal RL environment chapter is the interesting one. The data that matters in legal was never on the internet: what got conceded at 11pm, which redline the counterparty accepted, why a clause got dropped. It only exists in the negotiation, where we sit at @GenieAI.

@mercor @BrendanFoody i saw this article that preceded this on RL environments as the 3rd big demand for gen AI consumption (next to training/infernece) -

@mercor @BrendanFoody Building robust, automated verifiers for complex non-deterministic outputs (like legal briefs or financial models) remains the single hardest bottleneck in RL post-training.

@mercor @BrendanFoody @grok tldr ?

Pat, Founders, and Sequoia, you’ll have to watch this speech. Sequoia and Steve Hilton @SteveHiltonx should have a talk about the future of California. He’s got all the right ideas and he’s calling for a decade of building for California in this great speech.

@mercor @BrendanFoody turning data scarcity into a lever with RL worlds is clever

@mercor @BrendanFoody interesting.

@mercor @BrendanFoody Data feels very input heavy in this framing. Where do evals come in for outputs against real work? That tells me if the data delivered value in context.

@mercor @BrendanFoody An ignored but more powerful input into AI models is expertise, tacit knowledge that lives in peoples heads. Ie cognitive data. @CogFlowAI

@BrendanFoody @mercor Are you guys the growth lead for the next round? Would be cool if so 🤩

@mercor @BrendanFoody Data really is becoming the moat as model progress gets harder

@mercor @BrendanFoody The timestamps answer the headline: verifiers are the hard part. Production is a supply constraint; it scales with spend. Verification scales with scarce judgment. In regulated data the gate was never volume, it was two systems holding rival, defensible definitions of one entity.