Train the assistant on the answers you actually want

How it works today

In the product, Training is the name you see on thumbs, tags, and flows like Train on a good answer. We still call the capability Amplifier when we talk about the broader loop: structured feedback tied to the conversation you are in, so we can learn what worked, what missed, and what felt slow or unclear—without treating your browsing history as an open-ended data grab.

Today, Training is not anonymous: feedback is associated with your account and the interaction you rate, so we can debug, attribute issues to app version and context over time, and improve outcomes. On the roadmap are explicit anonymous training (rich signal without stable identifiers in the training payload) and personalized training (opt-in paths that attach identity so the assistant can adapt more directly to you or your team). Those modes are not the default UI yet; shipped behavior and policies will always win over marketing copy.

Under the hood we are moving toward a single interaction-shaped record—prompt, reply, tools, performance, and feedback together—so we can answer questions like what broke, which release changed it, and whether fixes actually helped. That engineering direction supports both short-term quality and longer-term optimization; it is not a promise that every rating retrains a foundation model.

Anything cross-user stays governed by aggregation, consent, and minimization as we ship; final privacy text lives in product terms and in-app notices.