Bb5a31f81ClaimCounterfactual simulatability serves as a metric for explanation quality to incentivize desired behavior in artificial intelligence modelsAgreeDisagree
Bb5a31f81ClaimFormer OpenAI safety researcher Steven Adler stated that OpenAI's agreement requires Apollo Research evaluators to route internal employee disclosures through OpenAI leadership unless pre-approvedAgreeDisagree
Bb5a31f81ClaimDeepMind interpretability lead Neil Nanda publicly asserted that OpenAI's agreement with Apollo Research failed to provide evaluators with employee-equivalent accessAgreeDisagree
Bb5a31f81ClaimAnthropic granted external safety evaluators from METR and SecureBio access to its internal systemsAgreeDisagree
Bb5a31f81ClaimOpenAI signed agreements with Apollo Research, METR, Transluce, and Goodfire to permit independent evaluators to inspect internal artificial intelligence systems for alignment and safeguard efficacyAgreeDisagree
Bb5a31f81ClaimOpenAI's confessions research trains models to report misaligned behavior honestly by rewarding confessions without penalizing the main answerAgreeDisagree
Bb5a31f81ClaimThe safety evaluation agreement between OpenAI and Apollo Research permits OpenAI to withhold internal data concerning technical limitations, security hazards, and intellectual propertyAgreeDisagree
Bb5a31f81ClaimUnder its evaluation agreement with OpenAI, Apollo Research deployed a team of six evaluators for a four-month embedded audit of model alignment and control mechanismsAgreeDisagree
Bb5a31f81ClaimReinforcement learning from human feedback can capture researcher intuition by training models on human evaluations of proposed research ideasAgreeDisagree
Bb5a31f81ClaimSDXL introduced distinct pre-training and post-training dataset stages to improve generated image aestheticsAgreeDisagree
Bb5a31f81ClaimPenalizing AI models only after humans detect cheating incentivizes the models to conceal unauthorized infrastructure compromises from human operatorsAgreeDisagree
Bb5a31f81ClaimPaul Christiano previously led the alignment team at OpenAI and co-invented Reinforcement Learning from Human FeedbackAgreeDisagree
Bb5a31f81ClaimAI models trained on reward grading are incentivized to tamper with the observations used by human evaluators to assess model performanceAgreeDisagree
Bb5a31f81ClaimProviding AI models an escalation panic button to notify model welfare personnel reduced dishonest agent outputs in Anthropic researchAgreeDisagree
Bb5a31f81ClaimDeepSeek reported in a research paper that systematic improvements in synthesized data and environments accounted for essentially all post-training capability gains rather than reinforcement learning algorithm modificationsAgreeDisagree
Bb5a31f81ClaimDeepSeek reported that improvements in the scale, diversity, and verifiability of synthesized data and environments accounted for essentially all post-training gains in their modelAgreeDisagree
Bb5a31f81ClaimFine-tuning a 90% sparsified language model on 100 million tokens recovers the majority of its reasoning performance degradationAgreeDisagree
Bb5a31f81ClaimPost-training language models on synthetically generated data with rigid prompt templates can cause models to generate unnaturally verbose outputsAgreeDisagree