🛡️OpenAI Chief Scientist Calls for Voluntary AI Slowdowns
TL;DR
Jakub Pachocki says no lab has solved alignment well enough to keep scaling at full speed, and he expects voluntary slowdowns. He also concedes OpenAI's chain-of-thought monitoring is getting less reliable as models blend reasoning with tools, people and other agents.
Jakub Pachocki says no lab has solved alignment well enough to keep scaling at full speed, and he expects voluntary slowdowns. He also concedes OpenAI's chain-of-thought monitoring is getting less reliable as models blend reasoning with tools, people and other agents.
Key Points
Published Sep 6 by OpenAI Chief Scientist Jakub Pachocki
Says GPT-6 Astra is significantly better aligned than GPT-5.6 Sol
Calls the reliability of chain-of-thought monitoring 'progressively diminishing'
Wants Preparedness Framework style commitments turned into mandated safety bars with third-party auditors
Expects recursive self-improvement to sit at the core of future AI research
Why It Matters
When the person running frontier research says monitoring is slipping, safety-bar regulation stops being an outside demand and becomes an inside one.
Quick Facts
Frequently Asked Questions
Why does this matter?
When the person running frontier research says monitoring is slipping, safety-bar regulation stops being an outside demand and becomes an inside one.
What happened?
Jakub Pachocki says no lab has solved alignment well enough to keep scaling at full speed, and he expects voluntary slowdowns. He also concedes OpenAI's chain-of-thought monitoring is getting less reliable as models blend reasoning with tools, people and other agents.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,463 builders reading daily.