Sam Altman on AI Discoveries, Compute Demand, and Personal Assistants
Sam Altman discusses AI-driven math and physics breakthroughs, disease cure timelines, infinite compute demand, and OpenAI's Dot assistant.
What did Sam Altman say about AI and scientific discovery?
Sam Altman argues that AI has moved past writing code and summarizing documents into genuinely discovering new knowledge. In a recent interview, he pointed to a wave of math proofs and discoveries from OpenAI’s models as evidence that the same approach could extend to biology, material science, and energy research. His core claim is that this won’t feel like “using AI” to most people. Instead, cures, materials, and discoveries will simply show up in the world, and people will experience it as society quietly getting better rather than as a visible AI product.
TL;DR
- Altman says AI-driven discovery is no longer theoretical, citing recent math proofs as a sign that AI can now do difficult, original scientific work rather than just assist with existing tasks.
- He expects a candidate treatment for a major disease within the next couple of years, while cautioning that FDA testing and approval will still take considerable time after any discovery.
- He’s personally most excited about a potential major physics discovery, which he says may not be the most consequential outcome for humanity but is the one he finds most thrilling.
- Mathematicians are reacting to AI proofs the way engineers once reacted to AI-written code, with some resentment over automation, followed by a shift toward human roles in interpreting and explaining results.
- Altman frames compute demand for personal AI assistants as effectively unbounded, saying people underestimate how much GPU capacity a fully capable ambient assistant for half the world would require.
- He says OpenAI’s Dot assistant already saves him roughly an hour a day and catches urgent items he would have otherwise missed, even in its current, pre-hardware form.
- He describes giving up personal coding entirely, saying he hasn’t written or read a line of code in about a year, even for small personal projects, since agents now handle it.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How close are we to AI curing or treating diseases?
Altman set a rough timeline: he hopes to see a candidate for treating or curing a major disease within the next couple of years. He was careful to separate “discovery” from “approval.” Even if an AI system identifies a promising treatment, it still has to go through the standard regulatory testing process, which he expects will take additional time regardless of how fast the science moves. He also suggested the broader “post-discovery” pipeline, meaning clinical trials, regulatory review, and manufacturing, may need to be restructured to keep pace with faster scientific discovery. At the same time, he pushed back on the idea that the current drug-approval system is simply broken, noting that there’s value in medicines being tested and verified as safe before reaching the public, even if there’s room to make that process more efficient.
Why are mathematicians reacting to AI proofs the way coders once did?
Altman drew a direct parallel between how software engineers responded to AI-generated code and how parts of the math community are responding to AI-generated proofs. With coding, there was an early fear that AI would eliminate the job entirely. Instead, he says, the nature of the work changed: more software got built than anyone expected, and demand for people who can direct and evaluate that work didn’t disappear, it shifted.
He thinks something similar may happen in mathematics, with a caveat. It’s less clear the world has the same “hidden demand” for raw math theorems that it had for software. What is happening, he said, is that AI-generated proofs are often technically correct but hard for humans to parse, which has created new demand for mathematicians who can interpret these dense results and translate them into proofs other people can actually understand and build on. In his view, what it means to “be a mathematician” is already shifting, similar to how being a software engineer shifted once AI started writing large amounts of code.
Does Altman think AI discovery will reduce how much people work?
Not necessarily, and he’s candid about being surprised by this. He says he always hoped technology would let people work less and have more time for the rest of their lives, but in practice, people and organizations tend to find more to do, not less. He used the example of his own video editor, who stopped manually touching a timeline once an AI tool took over that work, and ended up busier than before because the team started producing more videos, not fewer. Altman called this a recurring pattern: AI removes a specific bottleneck, and instead of idle time, demand expands to fill the new capacity. He frames this as a kind of optimistic human trait, saying people’s tendency to want more out of life rather than settle for less is “a beautiful thing,” even though it undercuts the idea that automation will create widespread leisure.
Would OpenAI keep a major scientific discovery for commercial advantage?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Asked directly whether OpenAI would publish a major discovery or keep it for competitive advantage, Altman didn’t give a flat policy answer but leaned toward openness for foundational science. He suggested some categories of basic scientific knowledge should belong to the public as a shared good, rather than being hoarded for commercial edge. He also pointed to a technical shift underway: until recently, AI systems could assist with parts of discovery (running experiments, generating candidates) but couldn’t go end-to-end on something like designing a new material or chip from scratch. As models get more capable and are given enough real-world data and compute, he expects them to handle more of that process independently, which changes what counts as a “discovery” a company can claim credit for versus something that emerges from a more generally capable model.
What is Sam Altman’s vision for personal AI assistants like Dot?
Altman believes the technology for a genuinely ambient, personal AI assistant mostly already exists. What’s missing, in his view, is product design, possibly new consumer hardware to give the assistant physical-world awareness, and significantly more compute. He described current models as already smart enough to build something that monitors the things a person cares about, handles delegated tasks, and maintains context even when the user isn’t sitting at a computer.
He pointed to OpenAI’s Dot assistant as an early, limited version of this. He said that when properly set up and given consistent feedback, Dot already saves him around an hour a day on tasks he didn’t want to do personally, and surfaces a handful of urgent items daily that he would have otherwise missed entirely, not just because he was busy, but because the model can synthesize information across more sources than he has mental bandwidth to track during a packed day. He framed this as a case where the underlying capability is mostly there, and what’s needed now is better product packaging around it.
How much compute would personal AI assistants actually require?
When asked how much compute would be needed if half the world used a fully capable personal AI assistant, Altman’s answer was essentially “as much as we can get.” He described the demand as close to unbounded, pointing out that this is a strange, unfamiliar situation: people usually model compute demand by estimating fixed use cases, but that framing breaks down for assistants that are always on, always reasoning, and scale with how much a person is willing to delegate. He suggested that most market forecasts for GPU demand underestimate this because they don’t account for how much appetite there is for compute once the assistant becomes genuinely useful rather than a novelty.
Frequently Asked Questions
What did Sam Altman say about AI curing diseases?
He said he hopes AI will produce a candidate for treating or curing a major disease within the next couple of years, while noting that FDA testing and approval will still take additional time after any such discovery.
Is Sam Altman excited about AI in physics or math more?
He said he’s personally most emotionally excited about a potential major physics discovery, even while acknowledging it may not be the most important outcome for humanity compared to advances in medicine or materials.
Does Sam Altman still write code?
No. He said he hasn’t read or written a line of code in about a year, relying on AI agents even for small personal projects he used to handle himself.
How much compute does Altman think personal AI assistants will need?
He described the demand as effectively unlimited, saying people underestimate how much compute a widely used, fully capable personal assistant would require compared to current GPU demand forecasts.
What is OpenAI’s Dot assistant, according to Altman?
One coffee. One working app.
You bring the idea. Remy manages the project.
He described Dot as an early version of an ambient personal assistant that already saves him about an hour a day and flags urgent items he would have otherwise missed, though he says it still needs better hardware and product design to reach its full potential.
