This has been written about in a few places so I’ll keep it brief. It was interesting that one of the root causes (note, not the sole cause) of the ChatGPT sycophancy issues was the feedback loop from the thumbs up/down data on posts, from their blog post:

“We also teach our models how to apply these principles by incorporating user signals like thumbs-up / thumbs-down feedback on ChatGPT responses.”

What’s interesting here is that cohort age of user feedback makes a difference:

“However, in this update, we focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time. As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous.”

This matches intuition (as you use the models more, your use gets more sophisticated, you get more sensitive to response quality) but it’s nice to see this validated, and I’ve made a mental note to remember this going forwards. Curation counts and perhaps you need a lot less data than you think. An obvious point, but the best labs in the world get it wrong.

The second write up is a decent read, running through (at a high level) their model release process, it has a post mortem like feel to it and gives a little more colour.