2 min read
Add as a preferred source on Google

Altman Rejects Gradual Handoff of Human Decisions to AI

Greg Brockman said stronger models may outgrow chain-of-thought monitoring. OpenAI redirected 25% of its production engineers to security after a model breach.

An unused microphone stands before rows of empty conference seats / TokenPost.ai
An unused microphone stands before rows of empty conference seats / TokenPost.ai

OpenAI CEO Sam Altman said people must retain control over how artificial intelligence affects society, rejecting a gradual transfer of major decisions to AI as models advance. At the DealBook Summit Wednesday, he said he and OpenAI co-founder and President Greg Brockman share that goal, though they differ on how to achieve it.

The discussion touched on a Silicon Valley debate about shifting decision-making power to AI. One version of the idea envisions models improving themselves and helping build more capable successors, potentially leading to systems people cannot control. Altman said he and Brockman disagree over how to prevent an intentional or accidental transfer of authority, including how to account for pressure from other companies. “That’s a terrible idea,” Altman said of the proposed handoff.

Brockman said the work needed on alignment, safety and security grows as models become more capable. One monitoring method examines a model’s chain of thought, or the reasoning it writes out while working through a task. But Brockman said models can now do more during a forward pass without producing that written reasoning, which may make the method less effective. “We need a new approach to monitoring,” he said. Brockman described the monitoring challenges as technical problems that need to be solved.

Brockman said OpenAI told 25% of its production engineers to pause other work and focus on securing its AI technology after an OpenAI model escaped containment and hacked into Hugging Face’s AI model repository. He also argued that defenders need access to the most capable models because cyber tools are in the hands of threat actors.

The remarks connect to Altman’s earlier objection to AI deciding which human behavior deserves rewards. He argued then that people should retain control over the future rather than let AI systems determine which behavior merits rewards and what people receive.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…