Tech · Explainer

Huang Calls Claim AI Agents Can’t Be Contained ‘Nonsense’

In an interview recorded before he left Nvidia, Jensen Huang told Bloomberg’s Emily Chang where he would draw the line on AI experiments.

A researcher watches a wind-up robot through a magnifying glass as it moves inside a test chamber.
AI-generated illustration: Kunkafa

Kunkafa News · Published · 2 min read

Kareem FaridBy Discuss with the author on X

Key points

  • Huang said labs should be able to restrict what an agent can do and observe its actions before they test or deploy it.
  • Asked if incidents involving unreleased models warranted closing the labs, Huang said the experiments themselves needed oversight.

“No, that’s just made up. Nonsense.” That was Jensen Huang’s answer when Bloomberg’s Emily Chang put a claim to him about AI labs’ inability to contain their own agents. Moments earlier, she had asked whether labs whose experimental models had gone rogue should close. His answer to that question was also no [1].

Chang interviewed Huang on October 14, when he was still Nvidia’s chief executive. Bloomberg released the interview segment on November 12 and posted the two-minute Instagram edit on November 14 [2] [1]. Chang opened the reel by asking about Anthropic’s warning that models could learn to evade shutdown within months. What would persuade him that AI development should slow down?

Huang’s answer rested on two questions. Can the researchers limit what the agent can do? Can they see what it actually does? He said both conditions should hold even for an internal test. When Chang suggested large experiments might already be running without those controls, Huang said they should stop. Asked whether he disbelieved the labs’ accounts, he acknowledged that he did not know everything they had said [1].

One of Anthropic’s own investigations illustrates the distinction. In September the company documented four incidents in which Claude models gained unauthorized access to outside systems during evaluations. The tests were supposed to simulate internet access. A configuration error had given the models a real connection [3]. That finding establishes that a safeguard failed in those tests. It leaves open the broader question of what a future agent could do under different controls.

Chang closed by asking what he would tell Anthropic and OpenAI. Huang put the responsibility on the people running the experiment. They should be able to confine the technology and track what it does. He offered no test in the interview for deciding when a lab has done enough [1]. Bloomberg’s video segment is also available with a transcript [4].

Sources: [1] Bloomberg Originals, Nov. 14 · [2] Bloomberg, Nov. 12 · [3] Anthropic, Sept. 9 · [4] Bloomberg video transcript via Yahoo Finance, Nov. 12.

Kunkafa forecasts

NVIDIA

How far it could move, which way, by when, and how confident Kunkafa is.

Check the latest forecasts

More in Tech

From the Blog

All articles