Researchers say OpenAI employees no longer publicly interpret System Cards, possibly due to a chilling effect.
X user @balesni posted that before the OpenAI firing controversy, "our colleagues" often posted on X to interpret frontier models' System Cards and explain their important safety implications to the broader public, and that they now appear to no longer do so; he hopes they are merely too busy and can still candidly share their thoughts with the public, but worries they have been affected by a chilling effect. The post quoted a post by @tomekkorbak that assessed the monitorability of GPT-6.2 Astra: he believes that although the model may appear to have improved, in the scenarios they care about, GPT-6.2 Astra is overall harder to monitor than GPT-6 Astra, and that this is easily overlooked in the System Card charts.
Back before the OpenAI's firings, our colleagues would tweet about system cards of frontier models, explaining important safety implications to a wider audience. They seem to no longer do that. I hope they're just busy and still feel able to share their thoughts candidly with the public, but I worry they're chilled.
How monitorable is GPT-6.2 Astra? While it might look like it got better, overall I think GPT-6.2 Astra harder to monitor than GPT-6 Astra in the contexts we care about. That’s easy to miss in the system card plots, so here's a short thread.View the quoted post on X