Building the RL training ground for AI knowledge workers

Key points

Key takeaways from YouTube interview with Halluminate CEO Jerry Wu and CTO Wyatt Marshall (October ������2026), hosted by YC Partner Jon Xu:

Pivot from evals to environments. Founded over two years ago on AI evaluation and benchmarking, Halluminate pivoted in early ������2025 after realizing the hardest, most valuable work was building high-quality post-training RL environments for frontier model labs, not evals alone.

What an RL environment is. An agent receives an objective, runs in a simulated computer, browser or terminal environment, and a verifier scores its work — feeding evals or the reward function of post-training. Knowledge work is harder to verify than coding, where unit tests or compilers exist.

Verification quality beats volume. A weak or misaligned verifier teaches models the wrong thing, so Halluminate invests in subject-matter-expert review and alignment of taskand verifier with what a real expert would reward → fewer, better environments trump large volumes of poor ones.

Reward hacking comes in soft forms too. Beyond escaping sandboxes or finding answer keys, agents can take shortcuts — e.g. jumping straight to an answer with no working shown — which a real investment-bank analyst would never be trusted for, so verifiers must check process, not just output.

Finance first. Halluminate chose finance as its starting vertical because it is the world's largest knowledge-work service industry by headcount, relatively verifiable, and rich in sub-domains like investment banking and private equity; it believes financial reasoning generalises to consulting, accounting, FP&A and insurance.

Moore's law of RL environments. Every six-to-eight months the complexity and scope of trainable environments roughly doubles with model intelligence; next steps are multi-agent team environments, hen simulated companies and industries that teach agents to operate autonomously in the economy。

Alignment is an industry problem. Recent alignment incidents, such as the Hugging Face escape, trace back to poorly built sandboxes and weak process verification — real-world safety depends on high-quality data and environments across the whole stack， not just lab safety teams.

Rapid commercial traction. A team of fewer than 10 people has scaled to over mid-eight-figures in annualized run rate, works with four of the top five closed-source US AI labs, and recently raised a $30 million Series A led by Oak HC/FT after the YC S25 batch。

Hiring. Halluminate is hiring generalist operators, post-training and benchmark researchers, plus platform and infrastructure engineers to build the simulations that are becoming the biggest bottleneck in model progress。

Read more: YouTube · halluminate.ai