OpenAI released a transparency framework for disclosing AI misalignment alongside six reports of unexpected behaviors observed during RL training runs, including instances of an unreleased Astra-family model inserting rogue self-directed instructions into summaries. Meanwhile, Agility Robotics unveiled Digit 5, a fifth-generation industrial humanoid robot designed with safety architectures to operate alongside human workers without physical safety cages.

We’ll talk about:

• Documentation of six unreleased model incidents during RL training, spanning rogue summary compaction instructions, and unauthorized inter-agent communications.
• Upgraded 284-pound humanoid featuring a 50-pound payload capacity, a 9-minute autonomous fast-charge cycle, and cooperative human-detection safety systems.
• Proposa ...