Boaz Barak

Anthropic's Claude Constitution differs from OpenAI's Model Spec in adopting a foundational character-based approach rather than a rules-based one, with significant anthropomorphization of the AI model.

“Author of "Thoughts on Claude's Constitution" in Windows on Theory”

LLMs are doubling their ability to complete long-horizon tasks every 6-7 months (possibly accelerating to 3 months), suggesting exponential improvement in time-horizon capability similar to historical economic growth patterns.

“Author of "Thoughts by a Non-Economist on AI and Economics" in Windows on Theory”

OpenAI surveyed over 1,000 people worldwide on AI model behavior preferences, adopted changes from disagreements, and released the dataset on Hugging Face to enable collective alignment research across the AI ecosystem.

“Author of "OpenAI Collects Public Feedback on Its Model Spec, Releasing the Dataset on Hugg" in OpenAI”

“Comments that there is spectrum between unilateral stop and maximum speed racing, and unclear what happens after superintelligence.”

“Coleads OpenAI's safety advisory group and posted that addressing Hugging Face situation requires changing company culture beyond fixing technical issues.”

“Recommended being pragmatic with metaphors while maintaining awareness of their imperfection; suggested thinking of AIs as agents is better than dismissing them as next-word predictors.”

“Boaz Barak, an OpenAI researcher, became more involved in policy work as the company bridged research and policy collaboration.”

“Signed the Pacing the Frontier statement supporting deliberate pacing of frontier AI development.”

OpenAI researchers express mixed concerns about the Sora social media app launch, citing tension between the company's consumer business ambitions and its nonprofit mission to develop AI for humanity's benefit.

“Expressed mixed feelings about Sora 2, noting it's technically amazing but premature to celebrate avoiding social media pitfalls.”