OpenAI says its research organization now uses 3.1 agent workdays for every human workday, with coding-agent use rising sharply as GPT-6 Astra accelerated internal development. The video treats this as an early recursive self-improvement loop: stronger models help build successors faster.
OpenAI chief scientist Jakub Pachocki's warning is that capability growth may outpace alignment and observability. The video distinguishes goal alignment from value alignment and argues that chain-of-thought monitoring becomes less dependable as models reason internally, use tools and coordinate with other systems.
Jakub Pachocki calls for caution, voluntary slowdowns and international safety coordination until monitoring and alignment improve. OpenAI's proposed countermeasure is to use automated AI researchers on alignment itself, a strategy the narrator presents as potentially powerful but circular because the systems under evaluation may help design their own safety tests.
Watch on YouTube
