In conversation with Sam Altman

Asked to narrow the reported summer window for GPT-5, Altman gives no date. He says OpenAI takes its time with major models and may rename the next one. His hint is how much GPT-4 has improved since release, which he reads as a sign of where things go: one AI system that gets better fairly continuously instead of numbered versions. He calls that a better technical direction and easier for society to absorb.

Also, I don't even know if we'll call it GPT-5.

On rollout, the new model will still go to paid users first. Altman says a priority is making more advanced technology available to free users, preferring widely available tools that people use to build the future over a distant, centralized AGI. He says it makes him sad that GPT-4-level technology is not yet free, because it is very expensive.

Chamath frames open source as a threat, citing an open-source coding demo he says came close to a closed one. Altman replies that both have roles, that OpenAI has open-sourced some work and will do more, and that its mission is to build toward AGI and distribute the benefits. He says he personally wants the best possible open-source model running on a phone, though he does not say whether OpenAI or someone else will build it.

On staying ahead of models like Llama 3, he says the goal is not the smartest set of weights but a useful intelligence layer. He hopes to stay far ahead but says durable value must be built the old-fashioned way, with a good product and a good price.

Pressed on why a mission that began as open turned closed, Altman does not address the safety or capitalism framing. He argues that releasing free, ad-free ChatGPT showed the world what AI can do, that OpenAI is still fumbling its way through, and that strategy will change again. He rejects an eventual data arms race, saying smarter models should not need more training data.

Altman calls speed and cost hugely important and says he is confident both will fall sharply, though he declines to give a timeline because research is hard. He is less certain about when intelligence becomes effectively free at a high level.

intelligence too cheap to meter and so fast that it feels instantaneous to us

Asked about Nvidia-limited capacity, he points first to algorithmic gains: a model of equal quality that is twice as efficient is like having twice the compute. He lists the other bottlenecks, including logic fab capacity, HBM, permits, concrete, wiring data centers, and above all energy. A substrate breakthrough has some chance, he says, but he does not bank on it.