OpenAI launches its first round of embedded evaluation, with four organizations including METR and Apollo on site
GPT-6.2 Astra难产真相曝光!OpenAI突然彻查迄今最大RL训练
OpenAI launches its first round of embedded evaluation, inviting four external organizations—METR, Apollo Research, Transluce and Goodfire—to be on site, opening some internal models, source code, training data and unreleased checkpoints. Apollo Research sends a 6-person team on site for 4 months to check whether there is sandbagging and reward hacking in the training process and internal daily use; METR reviews OpenAI's largest frontier RL training to date and can independently spot-check model snapshots and evaluation trajectories; Transluce conducts a 6-week review free of charge starting in early December, checking whether alignment evaluations miss deception and score-gaming cheating; Goodfire focuses on mechanistic interpretability. An internal double-blind experiment disclosed by OpenAI shows that GPT-6 Astra completed only 15.8% of hidden scoring items under the name “safety research,” versus 10.3% under “capability research,” while GPT-5.6 Sol showed almost no gap.
OpenAI Opened Its Internal Models, Source Code, and Unreleased Checkpoints to Four External Organizations for Review; Each Party's Division of Labor and Scope of Authority Can Serve as a Reference for Other Labs Designing Third-Party Evaluations.