GPT-5.6 Sol’s CoT controllability—while still low in absolute terms—is higher than in previous models. In general, greater CoT controllability can reduce CoT monitorability. I don’t think GPT-5.6 Sol crosses a threshold for concern, but we’re investigating what’s driving it.
Users praise GPT-5.6 Sol's high CoT monitorability and support the team with thanks or applications, while one warns that retention metrics in frontier models could bring serious downsides.
No Digg Deeper questions have been answered for this story yet.
For more details and more context on CoT controllability, see our system card https://deploymentsafety.openai.com/gpt-5-6-preview/cot-controllability
The reason I'm not worried just yet is that for most agentic traffic (10k+ CoT tokens), CoT controllability will still be below 10%.
The reason I'm not worried just yet is that for most agentic traffic (10k+ CoT tokens), CoT controllability will still be below 10%.
GPT-5.6 Sol’s CoT controllability—while still low in absolute terms—is higher than in previous models. In general, greater CoT controllability can reduce CoT monitorability. I don’t think GPT-5.6 Sol crosses a threshold for concern, but we’re investigating what’s driving it.
Also, if you're interested in helping us align and monitor superhuman AI agents, consider applying to out team https://openai.com/careers/researcher-recursive-self-improvement-safety-san-francisco/
@AradhyeAgarwal good luck!
@tomekkorbak Applied!
CoT monitorability of Sol remains high!
For more details and more context on CoT controllability, see our system card https://deploymentsafety.openai.com/gpt-5-6-preview/cot-controllability
@tomekkorbak Thank you!
@tomekkorbak Part of the problem with frontier models is that what you see as desirable customer-retention metrics may also signal the erosion of some skills in the population.
One day this could cost OpenAI tens of billions in fines. I’ve just finished writing a book on this.
@tomekkorbak What was the point of more controllability then if it makes monitoring worse lol
