My guest today is Dave Willner, cofounder of Zentropi and a fixture in the Trust and Safety world.
In 2008, he was a young Facebook employee who sat down to replace a one-page list of banned content — it said, roughly, no Hitler and no naked people — with the first real set of community standards the internet had ever seen. The rules he wrote are still the foundation of how Meta governs content shared by billions of people.
He went on to build community policy at Airbnb, and then joined OpenAI in early 2022, before ChatGPT existed, to figure out how to stop an image generator from producing bad things.
Now he's co-founded a company called Zentropi, with Samidh Chakrabarti, which is built on the idea that moderation at scale was never actually impossible — we just didn't have the right tools.
In This Episode:
Human moderation suffers from high turnover, trauma, and inconsistent judgment—problems small, tailored language models solve by strictly enforcing written rules.
Specialized small models can run circles around massive frontier models on speed and cost, making real-time classification financially viable.
Written policies and labeled datasets are two sides of the same coin; when an AI disagrees with a human label, it usually exposes a vague policy rather than a model error. Running moderation models locally on your own servers cuts per-check costs, dodges European privacy paperwork, and keeps user data in-house.
Bad actors are using AI to scale exploits at high speeds, forcing safety teams to adopt AI defenses immediately just to keep up.
A single global moderation rule doesn't work—dating apps, teen platforms, and social feeds require custom enforcement thresholds based on their specific risks.
Using the same LLM to draft a policy, generate test cases, and grade itself creates lazy evaluation loops that hide major safety blind spots.
Giving open-weights AI tools to volunteer moderators on Reddit or Discord lets small teams govern massive communities without burning out.
Chapter Timestamps:
00:00 Intro
3:21 Dave Willner’s trust-and-safety background and the origins of Zentropi
7:04 Zentropi’s specialized moderation models and policy-optimization tools
11:15 Why human moderation is inconsistent and difficult to measure
15:45 The limits of human judgment and the case for automated moderation
21:09 Imperfect language, model precision, and faster adaptation to abuse
24:43 Shared moderation infrastructure versus platform-specific policies
28:46 Why small specialized models can outperform larger general models operationally
32:26 Local deployment, privacy, and trust-and-safety resource constraints
34:16 Customer economics and real-time video moderation
38:09 The danger of “slop” policies and weak AI evaluations
44:10 AI moderation as a tool for decentralized community governance
47:06 The adversarial threat landscape and urgency of AI-enabled defense
Resources & Links:
Zentropi (https://zentropi.ai/)
Dave Willner (https://www.linkedin.com/in/davewillner/)
Rob Leathern (https://www.linkedin.com/in/leathern/)
41 min