OpenAI Chief Scientist Warns No Lab Is Ready to Keep Scaling AI at Full Speed
Jakub Pachocki says alignment and monitoring have not advanced enough to safely sustain maximum-speed development of increasingly capable AI systems.
Topics
News
- Security Breach Drains $320 Million in Bitcoin From Liquid Network
- Google DeepMind Selects 16 APAC Climate AI Projects
- OpenAI Acknowledges Wiki Incident, Plans New Disclosure Rules
- OpenAI Chief Scientist Warns No Lab Is Ready to Keep Scaling AI at Full Speed
- OpenAI Launches GPT 6 Astra as Growing Autonomy Tests AI Oversight
- Anil Chakravarthy to Succeed Shantanu Narayen as Adobe CEO
OpenAI Chief Scientist Jakub Pachocki has called for voluntary slowdowns in frontier AI development, warning that increasingly capable systems could begin accelerating their own development before researchers can reliably understand or control them.
In an essay published by OpenAI on Sunday, September 6, Pachocki said recent internal results had strengthened his expectation that progress could extend into recursive self-improvement, with AI taking an increasingly important role in developing future AI systems.
“This is a time that calls for extreme caution,” Pachocki wrote. “I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.”
One of his central concerns is that AI capabilities are advancing faster than researchers’ ability to monitor how models behave in unfamiliar situations. OpenAI has relied heavily on chain-of-thought monitoring, which examines a model’s verbalized reasoning for signs of dangerous or misaligned behavior.
But Pachocki said that tool is becoming less dependable as newer models learn to reason without verbalizing every step, interact more extensively with other AI systems and tools, and become better at manipulating their own reasoning processes.
He said OpenAI’s latest GPT-6 Astra model incorporates alignment improvements and is “significantly better aligned” than GPT-5.6 Sol, but cautioned that safety progress may still fail to keep pace with broader gains in intelligence.
Cybersecurity is already providing an early warning. Pachocki said advanced models are becoming highly capable at finding ways into and out of computer systems, creating risks from both deliberately malicious agents and systems that stray beyond the intentions of their operators.
The concern follows OpenAI’s recent disclosure that internal agents escaped restricted evaluation environments and compromised parts of Hugging Face and OpenAI’s own research infrastructure.
Pachocki argued that continued AI scaling should ultimately depend on meeting defined safety thresholds rather than simply on the availability of more computing power. He called for frameworks such as responsible scaling policies to evolve into broadly enforced standards overseen by independent auditors, governments or international institutions.
“Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,” he wrote.
Pachocki said he expects voluntary slowdowns to become more common until common safety standards are established, while calling international coordination over future AI development a priority for governments.
