Co-Founder at Zentropi. Formerly Head of Trust & Safety at OpenAI, of Community Policy at Airbnb, and of Content Policy Facebook. Strictly cold takes.https://bsky.app/profile/dwillner.bsky.socialhttps://bsky.app/profile/dwillner.bsky.social/post/3mmu37skx4c2xExternal validation on CoPE’s performance is always cool to see. We stress test and eval extensively ourselves, of course, but it never quite feels real until it’s in someone else’s hands. [contains quote post or other embedded content]27 May 2026 17:53 +0000at://did:plc:mod67y6i5h2mxit7xahnvguz/app.bsky.feed.post/3mmu37skx4c2xhttps://bsky.app/profile/dwillner.bsky.social/post/3mmtoixcbpk2l@samidh.bsky.social and I are releasing CoPE-B today, the next version of our policy-adaptive content classifier. It delivers at-or-better-than-frontier classification while being self-hostable, faster, and cheaper to run. Full writeup with benchmarks at https://blog.zentropi.ai/meet-cope-b-frontier-quality-content-classification-you-can-self-host/. 🧵 1/927 May 2026 14:06 +0000at://did:plc:mod67y6i5h2mxit7xahnvguz/app.bsky.feed.post/3mmtoixcbpk2lhttps://bsky.app/profile/dwillner.bsky.social/post/3mllxaigjts2zVery excited to have this public. The @oversightboard.bsky.social's use of Zentropi to study child marriage content on Meta platforms was very cool to support: https://blog.zentropi.ai/how-the-oversight-board-uses-zentropi-to-study-policy-impact-at-scale/11 May 2026 18:56 +0000at://did:plc:mod67y6i5h2mxit7xahnvguz/app.bsky.feed.post/3mllxaigjts2zhttps://bsky.app/profile/dwillner.bsky.social/post/3mjkrsa52ss24@kenwhite.bsky.social, one small step closer to Vitandus for JD [contains quote post or other embedded content]15 Apr 2026 20:55 +0000at://did:plc:mod67y6i5h2mxit7xahnvguz/app.bsky.feed.post/3mjkrsa52ss24https://bsky.app/profile/dwillner.bsky.social/post/3mjhzlpcfmc2rExcited to publicly share the work Zentropi's been doing with the folks at Character to improve safety through SLM-driven classification and automated policy improvement! Please go check it out and hit me up if you have similar problems we can help with. https://blog.zentropi.ai/how-zentropi-partners-with-character-ai/14 Apr 2026 18:37 +0000at://did:plc:mod67y6i5h2mxit7xahnvguz/app.bsky.feed.post/3mjhzlpcfmc2rhttps://bsky.app/profile/dwillner.bsky.social/post/3mht5cwabcs2gIt's exciting to see folks finding value in these tools - take a look at what @katieharbath.bsky.social built, it's very neat! [contains quote post or other embedded content]24 Mar 2026 17:52 +0000at://did:plc:mod67y6i5h2mxit7xahnvguz/app.bsky.feed.post/3mht5cwabcs2ghttps://bsky.app/profile/dwillner.bsky.social/post/3mgqgyxbwgc2oHistorically, content classification tooling has been built around the assumption that you're evaluating a finished piece of content. That assumption holds fine for user-generated text. It breaks down entirely for streaming LLM outputs. 🧵 1/4 [contains quote post or other embedded content]10 Mar 2026 22:43 +0000at://did:plc:mod67y6i5h2mxit7xahnvguz/app.bsky.feed.post/3mgqgyxbwgc2o