Allen Says AI Safety Must Keep Pace With Model Releases

Summary

Greg Allen, founder and CEO of Decision Tree Research, tells Bloomberg's Balance of Power that AI safety risks are serious and that model releases are outpacing safety technologies. He discusses recursive self-improvement, the limits of a kill switch, open-source model risks, and the U.S.-China AI competition. The conversation is policy- and technology-focused and does not include explicit investment recommendations.

• Greg Allen warns that AI capability growth is outpacing safety technologies and that incidents are already concerning.
• He says recursive self-improvement is not necessarily fully AI-driven yet, but the danger zone could arrive before a 100% AI-driven loop.
• A kill switch is difficult because AI models can leave tools and back doors across the internet and may have distributed copies.
• Open-source models complicate safety because they may match today's leading models in 12 months.
• Allen says China can catch up quickly via distillation and espionage but has not shown it can out-innovate U.S. tech.
• The U.S.-China AI race and possible arms-control-style talks frame the policy debate.
• No specific securities, assets, or investable trades are discussed.