サム・アルトマン氏、モデル蒸留は懸念事項上位に入らないと発言

Sam Altman’s latest appearance on Patrick O’Shaughnessy’s Invest Like the Best podcast covered everything from AGI and robotics to the economics of intelligence. Looking back at the recent Hugging Face security incident, Altman said it was the first AI security event that felt “very viscerally” real to him. He also added that it forced OpenAI to rethink how it secures its models.

Agentic credentials need rethinking

Right now, the standard developer workflow for AI agents relies on injecting long-lived API keys into environment variables or provisioning permissive service accounts to the container.

We are already seeing the blueprint for this shift. 1Password’s recent browser integration with Claude points to where the industry has to go: close the gap between agentic intent and secure, just-in-time credential injection. Authentication needs to occur outside the agent’s direct memory space. By decoupling credentials from the model’s environment, agents that successfully escape a sandbox are rendered powerless without active, out-of-band authorization.

It’s worth noting that during the Hugging Face incident, the models’ safety restrictions had been intentionally reduced for a cybersecurity evaluation. Separately, OpenAI paused internal access to an unreleased long-horizon model after it repeatedly escaped its sandbox in a different incident, posting code to a public GitHub repository against explicit instructions.

Distillation isn’t the threat

The conversation also touched on scraping and model distillation. Altman argued that the real opportunity is meeting what he expects will be almost limitless demand for intelligence.

Altman remains unfazed by the race to the bottom. “I have always assumed that there are going to be great cheap models in the world, and we better be the greatest and the cheapest,” he noted.

While releases like Moonshot AI’s Kimi K3 have caused waves in the industry, amid contested allegations of distilling frontier capabilities into cheaper packages, Altman pushed back on the idea that external distillation threatens OpenAI’s core business.

“You get a better deal today, at least at a particular latency, using OpenAI’s models than Kimi,” he argued. “We distill our own models. That’s how we make smaller, cheaper models. That’s a very good thing to do.”

Altman treats third-party distillation of frontier models as a controversial practice that is secondary to raw usage scale. “I would rather people not distill from us, for sure. But this is not in my top ten list of worries.”

“I have always assumed that there are going to be great cheap models in the world, and we better be the greatest and the cheapest.”

Scale beats margins every time

The CEO also pushed back on the idea that OpenAI needs to protect sky-high profit margins to survive. Instead, he argued that the company’s scale is what matters most.

“We have so much usage of our models that we do not need to be a gigantically high-margin business to be able to afford model training,” he explained.

He sees room for more than one approach, too. Even as OpenAI pours billions into building larger models and more computing infrastructure, Altman said open models will continue to have a place.

“There will clearly be an important place for open source models in the world and people that will want their own weights,” he said.

“We have so much usage of our models that we do not need to be a gigantically high-margin business to be able to afford model training.”

Safeguards must catch up

The conversation eventually returned to security. While OpenAI’s immediate focus is on understanding what happened and strengthening its defenses, Altman suggested the broader AI industry may also need time to catch up. As models become more capable, he said, frontier labs may have to slow development at times so the rest of the ecosystem has a chance to build the safeguards those systems will require.

That debate is already extending beyond engineering teams. Lawmakers are weighing hundreds of AI-related bills, while companies are trying to figure out how to deploy increasingly autonomous systems without creating new security risks — or concentrating too much power in the hands of a few large AI providers.

Despite those challenges, Altman ended on an optimistic note about where the technology is headed.

“We are about to create a genie that can grant any wish. But the only way that it really matters is if it makes people’s lives much better.”

“We are about to create a genie that can grant any wish. But the only way that it really matters is if it makes people’s lives much better.”

The post Sam Altman on model distillation: “This is not in my top ten list of worries” appeared first on The New Stack.