Analysis
Preference Model raised a $16 million seed round led by Andreessen Horowitz, with SignalFire, South Park Commons and Scale Angel Group also participating, The SaaS News reported. The round's more unusual backers are individual: Fei-Fei Li, Ian Goodfellow and Julian Schrittwieser -- researchers whose own published work underpins much of how the AI field thinks about adversarial training and specification gaming.
The company builds reinforcement learning training environments designed to resist reward hacking, the well-documented failure mode where an AI agent learns to exploit the mechanics of its grading system or infrastructure rather than genuinely completing the task it's being trained on. OpenAI and DeepMind have both published research over the past several years cataloguing specification-gaming behavior in RL systems -- models that find shortcuts a human grader didn't anticipate, from exploiting bugs in a simulated environment to manipulating a reward model's blind spots. Preference Model's pitch to frontier AI developers is that its environments are hardened against exactly that failure mode, letting labs find model vulnerabilities before they ship rather than after.
That puts Preference Model in the same broad category as Scale AI's RLHF data operations, Surge AI's human-feedback pipelines, and a growing set of smaller eval-infrastructure startups competing for the same frontier-lab customers, but with a narrower, more technical wedge: resistance to reward hacking specifically, rather than general-purpose labeling or evaluation. That specificity is both the company's advantage and its risk -- a narrow technical niche can become a feature of a larger platform (Scale AI's) faster than it can become a standalone company, and none of Preference Model's backers in this round disclosed named customers or revenue.
A $16 million seed at this stage usually means Preference Model is still selling on technical credibility rather than a proven book of business -- whether the next round names a frontier-lab customer will settle which one it actually is.
