Resolves YES for a year or year range if there is a frontier AI model released within that range whose chain of thought (CoT) cannot be trusted to reveal when it is misaligned (i.e. intentionally misbehaving).
The assumption for models released that use CoT is that they are CoT monitorable, evidence would need to be provided demonstrating this assumption is false (e.g. a peer reviewed paper or highly endorsed LessWrong post). A frontier model released that always acts without CoT or anything similar would trigger a YES resolution for the time period of its release.
If all frontier models are released in the time period feature CoT and their monitorability is not disproven, then that time period resolves NO.
If in the event no frontier models are released during a time period, that period resolves N/A, as I am interested specifically in when CoT monitoring fails as an oversight / AI control technique.
1,000People are also trading
@traders https://www.lesswrong.com/posts/ntKx9YHWCwxSeGbRB/estimating-gpt-6-astra-s-no-cot-time-horizon
Thoughts? Seems like 2026 could resolve YES? (Though I might wait for a more CoT monitoring specific post demonstrating that it has failed)
Relevant LessWrong post that links to further more specific markets: https://www.lesswrong.com/posts/Em9sihEZmbofZKc2t/a-concrete-roadmap-towards-safety-cases-based-on-chain-of
@Haiku It's not 100% faithful, but it's good enough that we can use it to detect misalignment (Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety ).
Things on this market would resolve NO (edit: YES) if AI safety researchers broadly agreed it was no longer feasible to do this.
