Sam Altman on Cheeky Pint

Your AI and this story

Key facts

• Second wave (Tech Times UK, Nov 2): Altman said ‘we did let it get out of control for one week on the math’, broadly agreed the wider maths push cost hundreds of millions of dollars (his estimate, including unpublished results), said OpenAI has further unpublished maths findings, and: ‘on net, the world deserves to understand how the universe works’
• Many AIs, not one: ‘I would much rather intuitively have a trillion AIs of like slightly superhuman intelligence join society’, with teams of agents ‘keeping each other in check’, tied to ‘values of pluralism and democracy’ (transcript)
• Limits: after saying the pipeline that made models strong at maths will probably work for AI research, he said ‘we are going to have to agree to some limits about how smart we let these single models get’, and that any slowdown would come from that decision rather than from AI being unable to do research (transcript)
• Sandboxing: ‘I think sandboxing works to a point. Then I think monitoring works to a further point’; ‘we can do much better than we’re doing now’; of an AI ‘a million times smarter than the smartest security researcher’: ‘would you really be willing to trust that sandbox?’
• PHP flaws: ‘I don’t want to declare victory at all on the PHP thing, because I still don’t think we know’; models are getting better at finding vulnerabilities and ‘we are in a world of many more serious vulnerabilities than we expect’; such bugs surfaced now because ‘People just hadn’t really used them to go look at everything out there’ (see PHP2Shell: critical remote-code-execution flaws in PHP, found with GPT-5.6 Sol and Gemini 4 Argon, force emergency patching and take websites offline worldwide)
• Maths: the Navier–Stokes proof cost ‘Order $10 million’; ‘I bet we could redo Navier-Stokes at 1/100th of the cost a year from when it was done’; at the time ‘we understood absolutely nothing about the scaling laws of teams of agents’, which may have ‘very interesting scaling law properties, but also variance in alignment properties’ (see OpenAI claims a Millennium Prize problem)
• Compute: on a five-year total cost of ownership of about $70B per gigawatt, ‘if that number were 20 instead of 70, we should be much happier’; ‘the fundamental cost of turning the sand into intelligence, that should go down’
• Robotics: ‘We have a robotics program that we’re going to point at making lots of data centers’ (no timeline given)