🥺 Sam Altman says OpenAI is willing to let some things go wrong

Sam Altman is offering an unusually blunt look at how OpenAI is thinking about AI risk and the trade-offs are becoming harder to hide.

Altman says OpenAI is willing to accept “some things going wrong” rather than allow enormous AI capabilities to become concentrated in a handful of companies, people or countries. He specifically points to open-source models, while warning that they could trigger “a coming tidal wave of cybersecurity problems.”

On President Trump’s new AI accord, Altman says it currently has “no real power” and argues that more governance will eventually be needed. His dividing line is clear: ordinary failures can be tolerated, but losing meaningful human control over AI cannot.

That philosophy also puts an enormous amount of judgment on OpenAI itself. Altman admits: “I’m going to make many individual decisions that you’re clearly not going to like.”

His response to critics? “Come try making them all; they’re not so easy.”

And OpenAI is already making those calls. Altman says Astra 6.1 was “ready to go,” but the version they had prepared would have “tipped too much towards the accident risk.” OpenAI therefore delayed it, choosing to work toward a version it believes can be distributed more broadly with greater confidence in its alignment.

@aipost 🏴