How can humans evaluate work that is too complex, extensive, or intellectually advanced for them to verify directly?
Scalable oversight is the field's name for this problem. Sandwiching experiments — humans supervising models more capable than themselves on a task — are the main empirical paradigm so far.
View on the map → · Open in Browse →
The oversight agenda moved from debate theory toward buildable systems. Transluce proposed training a dedicated foundation model whose job is answering oversight questions about other models, reframing oversight as executable world-modeling; a complexity-theory result relaxed debate's need for two competing provers to one; and DeepMind's mid-year summary reports the whole amplified-oversight program shifting 'into the midgame' — from proofs to production.
Any oversight question can be formalized as Bayesian inference over the outputs of a Pythonic world model.
Proposes training a dedicated foundation model to answer oversight questions about other models, reframing oversight as world-modeling with executable Python specifications and a label-free data pipeline — the most ambitious concrete vision this window for making oversight scale with capability.
We initiate the study of single-prover interactive proofs for AI safety
Shows interactive verification can work with a single AI prover rather than two competing models, relaxing debate's assumption that an equally capable honest debater exists — advancing the complexity-theoretic foundations of scalable oversight.
We are now fully in the midgame, and focus more on landing things in production.
GDM's consolidated update on amplified oversight (debate variants, prover-estimator debate, human-AI complementarity), CoT monitorability, and alignment evals — a status report on how a frontier lab is operationalizing oversight.