Dave Willner

We're shipping CoPE-B-S1, an open-weight decision model tuned for safety use cases. It's a significant step forward for our CoPE family of models in a few ways. A thread: blog.zentropi.ai/meet-cope-b-...

When we released the first CoPE almost two years ago, we bet that small, steerable classifiers would become essential AI infrastructure. That bet's held up: CoPE is already the real-time safety layer at Runway ML and some of the largest websites in the world.

We're shipping CoPE-B-S1, an open-weight decision model tuned for safety use cases. It's a significant step forward for our CoPE family of models in a few ways. A thread: blog.zentropi.ai/meet-cope-b-...

So it shines when you need a model to faithfully apply the long, nuanced criteria Trust & Safety policies are built on. It's really, really good at following where you want it to go, no matter how persnickety you need to be.

We're shipping CoPE-B-S1, an open-weight decision model tuned for safety use cases. It's a significant step forward for our CoPE family of models in a few ways. A thread: blog.zentropi.ai/meet-cope-b-...

When every category costs another model call, teams make uncomfortable trade-offs: lumping distinct harms into one blunt classifier, checking only a sample, or quietly leaving some rules unenforced. With CoPE-B-S1 those trade-offs more or less disappear.

We're shipping CoPE-B-S1, an open-weight decision model tuned for safety use cases. It's a significant step forward for our CoPE family of models in a few ways. A thread: blog.zentropi.ai/meet-cope-b-...

First, it's multi-label. Hand it one piece of content and up to 26 individual policies, and it returns every classification in a single call, at close to the speed of a single policy (controlling for overall prompt length).

Runway is doing some very cool stuff using CoPE/Zentropi: "We have designed and tested a new fast moderation system that will run synchronously once a user’s input clears moderation.... The synchronous moderation through Zentropi narrows the exposure window to less than half a second."

This sort of thing is really only possible when you have a multi-modal model that combines accuracy with speed in the way that CoPE does. Big foundation models can do the classification well, but aren't fast enough. Unspecialized small models are fast enough, but can't do the classification well.

At TrustCon this year I talked about a technique we’ve developed for automatically optimizing content-moderation policies, using an inversion of the binocular labeling approach Zentropi had already pioneered. Today we're shipping the tool that technique became. blog.zentropi.ai/optimizing-o...

The system is available now on zentropi.ai usable directly on the website or through the Zentropi Agent Skill for coding agents. Normally, full use is subscriber-only because of how compute-intensive it is, but for launch we’re giving anyone a limited quota of free so folks can try it.

At TrustCon this year I talked about a technique we’ve developed for automatically optimizing content-moderation policies, using an inversion of the binocular labeling approach Zentropi had already pioneered. Today we're shipping the tool that technique became. blog.zentropi.ai/optimizing-o...

To summarize the point - a policy and a set of examples labeled against it aren't independent things you can fix one at a time, because a given label can only be said to be meaningfully right or wrong relative to a specific policy text. So, to improve either you have to work on both simultaneously.

I gave a talk at TrustCon this year that seemed well received, so we have written the argument up as a post on the Zentropi blog and are sharing the slides with it: blog.zentropi.ai/machines-edi... 🧵 1/5

Closing them is a loop: run the policy against examples you trust, look where the model disagrees with your reviewers, and let a second agent rewrite the passages behind the disagreement in the policy's own voice. Partners have gone from 65% to over 90% agreement with their own labels. 🧵 4/5

I gave a talk at TrustCon this year that seemed well received, so we have written the argument up as a post on the Zentropi blog and are sharing the slides with it: blog.zentropi.ai/machines-edi... 🧵 1/5

The argument, briefly: once a small model (CoPE) reads a content policy as written and labels against it, editing the policy is editing the classifier, with no retraining cycle in between. 🧵 2/5

Making impossible safety challenges tractable is a lot of why we built Zentropi. Today @vaishnavi.bsky.social shipped a great example of what that means in practice - a set of open-source self-harm and suicide policies for teens: quire.substack.com/p/what-it-takes-for-llms-to-make-age 🧵 1/6

Because these policies are open source, none of the specific lines are take-it-or-leave-it. If your community needs a different rule, you can fork the text and move the line, and your enforcement follows. 🧵 5/6

Making impossible safety challenges tractable is a lot of why we built Zentropi. Today @vaishnavi.bsky.social shipped a great example of what that means in practice - a set of open-source self-harm and suicide policies for teens: quire.substack.com/p/what-it-takes-for-llms-to-make-age 🧵 1/6

Platforms have mostly handled this content with one, often blunt, policy because keeping different rules for different ages was operationally out of reach. That's something policy-steerable classifiers change: they let you write the rules per age band, and have enforcement follow the text. 🧵 3/6

Making impossible safety challenges tractable is a lot of why we built Zentropi. Today @vaishnavi.bsky.social shipped a great example of what that means in practice - a set of open-source self-harm and suicide policies for teens: quire.substack.com/p/what-it-takes-for-llms-to-make-age 🧵 1/6