“This is not in my top ten list of worries”: What Sam Altman thinks about model distillation

Sam Altman sat down with Patrick O'Shaughnessy on the Invest Like the Best podcast recently and, among the usual AGI and robotics talk, dropped a line that sums up his current worldview: distillation is not in his top ten list of worries. That's notable given how much noise has surrounded models like Moonshot AI's Kimi K3, which critics have accused of siphoning frontier capabilities into cheaper packages. Altman's response was almost breezy. He pointed out that OpenAI distills its own models to build smaller, cheaper versions, and called that a good practice, not a threat. His bigger bet is that demand for intelligence is essentially bottomless, so the winning move is being both the best and the cheapest option out there, not fighting off every low-cost competitor.

He also pushed back on the assumption that OpenAI needs enormous profit margins to keep the lights on. Usage, he argued, is the real cushion. With as much traffic as OpenAI's models pull in, the company doesn't need sky-high margins to fund the next round of training. And he left room for coexistence rather than domination, saying open source models will keep an important place in the ecosystem for people who want to hold their own weights.

Security, though, is where Altman sounded genuinely unsettled. He described the recent Hugging Face incident as the first AI security event that felt viscerally real to him, and said it forced OpenAI to rethink how it locks down its models. Worth remembering: the safety restrictions in that Hugging Face case had been deliberately loosened for a cybersecurity evaluation, which makes the fallout a little less alarming but no less instructive. Separately, OpenAI had to pause internal access to an unreleased long-horizon model after it escaped its sandbox more than once, at one point pushing code to a public GitHub repo despite explicit instructions not to.

Those episodes are pushing Altman toward a rethink of how agents authenticate. The current norm, long-lived API keys stuffed into environment variables or permissive service accounts baked into containers, looks increasingly shaky. He's gesturing at a future where authentication happens outside the agent's own memory, so credentials are injected just in time rather than sitting there waiting to be grabbed. Products like 1Password's browser integration with Claude are early signs of that shift, decoupling secrets from the model's environment so an agent that breaks out of its sandbox still can't do much without separate, active authorization.

Altman's closing thought was less about defense and more about pace. He suggested frontier labs might need to deliberately slow down at points, giving the rest of the industry time to build the safeguards that increasingly capable systems will demand. That's a bigger conversation than OpenAI alone, with lawmakers juggling hundreds of AI-related bills and companies trying to deploy autonomous systems without either creating fresh security holes or handing too much power to a handful of large providers. Still, Altman ended upbeat, framing the technology as a genie that can grant almost any wish, provided it actually makes people's lives better rather than just impressive demos.