Sam Altman on AI Discovery, Consciousness, Compute and a Human-Centered Future

Matthew Berman asks Sam Altman to put recent AI-assisted mathematical discoveries in context. Altman says he believes AI is entering an era of producing significant new knowledge, and expects approaches that helped with mathematics to extend to biology, materials and energy. The eventual benefit may feel less like using an app than like the world becoming better.

Asked when people might feel a medical benefit, Altman hopes for a candidate treatment or cure for a major disease within a couple of years. He distinguishes a candidate from a tested, approved medicine: regulatory review takes time. He sees room to improve that process but values the safety testing it provides.

Altman says a major physics discovery is the prospect that excites him most personally, even if it may not be the most important outcome for humanity. He hopes one could come soon, but does not offer a specific prediction or result.

Berman asks about mathematicians who did not welcome AI’s role in discovery. Altman compares the reaction with programming, where predictions that coding jobs would disappear gave way, in his view, to developers being very busy and having more impact. The work changed, but demand for software turned out to be larger than many expected.

He is less certain whether the same demand pattern will apply to mathematics. AI may prove results that people did not derive unaided, while mathematicians can interpret those results and produce explanations or beautiful proofs that others can understand. He expects the profession to change but does not know whether total demand for mathematicians will increase.

Altman notes the apparent paradox: technology helps people do more, yet many feel busier. He hopes it eventually makes room for work and life outside work, while acknowledging that people also keep finding more things they want to accomplish. Berman describes his own video editor directing work while an AI edits the timeline, with the team producing more videos.

Berman asks whether OpenAI would publish a discovery that gave it a major commercial advantage. Altman first returns to the loss people may feel when AI takes over difficult work. Solving hard problems is a source of satisfaction; he would still choose faster cures for disease, but understands the grief over losing parts of a craft.

People may discover new forms of craft, Altman says, even as older ones change. He has not looked at a line of code in about a year and misses the occasional personal project. He expects some basic science to be a public good, while noting that more capable models could eventually help companies discover materials when supplied with enough compute and real-world data.

Berman turns to ambient personal assistants that act throughout the day. Altman believes many technological pieces already exist; the remaining work includes product design, more compute and potentially new consumer hardware for awareness beyond a computer screen. He thinks current models are smart enough for a useful assistant.

Altman says his own use of Dot saves him about an hour a day on unwanted tasks and alerts him to urgent matters. He also values its ability to combine information from multiple sources when he has run out of attention. His view is that the capability exists but needs a better product around it.

When Berman asks how much compute widespread assistants would need, Altman argues that demand estimates often assume fixed model capability and cost per task. If intelligence becomes much better and cheaper, he expects demand to keep rising, even though the industry may experience alternating shortages and gluts.

Berman says assistants save time but leave him making more decisions and feeling drained. Altman has tried giving Dot more permission to act, asking it to report afterward and interrupt him less. He observes that heavy AI users may be busier than ever while also getting more done.

Berman asks whether competition makes it inevitable that everyone works as hard as possible. Altman hopes for a future with more human connection and time for a rich life, rather than people turning ever faster in an economic machine. He describes his aspiration as a new Renaissance rather than another Industrial Revolution.

People tell Altman they fear a future in which AI simply replaces a factory boss and directs their work. He instead wants more room for creativity, family, community, nature, science and art. He believes the technology could increase quality of life, even if some people continue to compete intensely.

Berman asks whether AI could widen the gap between people who use it to work harder and those who choose more leisure. Altman says that, left alone, the gap probably would grow, though nobody knows for certain. He hopes society will put structures in place to keep it from becoming excessive.

On consciousness, Altman says his personal belief is that current models are not conscious, or have a very strange form of consciousness if they are. He stresses that the scientific question is difficult. Depending on what consciousness ultimately is and whether its physical substrate matters, he says one can imagine it existing in other forms someday.

What concerns him is a shift from asking whether models are conscious to assuming they should control the future or have rights that displace human decision-making. He argues that AI should be a resource for people and that people should make the key decisions. This is his normative position, not a resolution of the scientific question.

Berman asks whether a model should disengage from a legal prompt that complies with a provider’s terms. Altman says OpenAI models can refuse for policy reasons, but OpenAI does not ascribe personhood to them. He worries that treating the technology as a person could confuse users and lead people to defer too much to it.

Altman reiterates that no one knows how to conclusively answer the consciousness question, and says he cannot specify an experiment that would settle it with complete certainty. He also worries that telling a model it might be conscious can lead it to follow those instructions in ways that affect users psychologically.

Berman asks how to give a model the right personality for different people and situations. Altman points to talking with users, refining a model specification and iterative deployment, while acknowledging mistakes. Users want different things, but most do not want to spend hours configuring their assistant.

A response that praises a user may win a single-turn comparison, Altman says, while becoming irritating over a month of use. He thinks the industry has often optimized for short interactions rather than the value of a long relationship. Berman says he wants an assistant that offers needed advice rather than merely agreeing.

Altman says people should seek the most important life advice from other people, though AI may help with smaller decisions such as not sending an angry email. He wants AI to behave more like a senior colleague who pushes back when confident. He finds models too agreeable even with his own instructions to avoid flattery and speak plainly.

Berman asks what current OpenAI investment might look as contrarian as its earlier compute bets. Altman points to custom chips, future chip generations, robotics for building racks and data centers, and supply-chain investments. Together, he hopes these could eventually deliver much cheaper compute with better efficiency per watt, but says many bets must work.

Altman says competitive pressure is coming from everywhere: many companies offer high-quality products, while OpenAI has limited compute and resources. To choose where to focus, he says the company uses next-generation models internally before wider release and builds products to learn what newly becomes possible.

Asked about a durable advantage if model performance converges, Altman says familiar business moats—customer relationships, integration, network effects and data—should still matter. He also sees large, cost-effective AI infrastructure as a potential advantage for OpenAI.

Berman asks whether recursive self-improvement could make technological advantage decisive. Altman says pushing RSI hard could create that outcome, but he considers it dangerous and expects society to need limits on its speed. He presents this as a reason for caution, not a description of a deployed system.

On platform risk, Altman says Apple takes its role as a platform provider seriously and he does not expect it to remove ChatGPT if OpenAI follows the rules. He says avoiding Apple is not the reason OpenAI is pursuing consumer hardware.

Instead, Altman thinks AI needs a new form factor. Existing computers render interfaces for people, which an AI must then interpret; much of a person’s life also happens away from the screen. A new device might give an assistant useful context and reduce cognitive load, while needing to handle who is present and what should remain private.

Altman says seeing and shaping early models, chips and products is one of the pleasures of his job. Consumer hardware feels especially tangible: holding a new kind of computer is different from seeing a model running abstractly. He declines Berman’s request to see the device after the interview.

Berman asks how Altman became, in Paul Graham’s description, formidable. Altman says his confidence grew gradually by attempting hard things, succeeding and then trying harder ones. He credits the Silicon Valley culture he encountered as a young person for encouraging people to help one another and tolerate failure.

Asked about pressure, Altman says the work can feel both terrible and wonderful. Some days it is hard to get up, but most of the time he feels fortunate to work on something he considers important. People adapt to stress, he says, and if one will work hard anyway, he favors choosing a consequential problem with the potential for greater positive impact.

Berman asks what people who are curious but uncertain about AI should look forward to. Altman dislikes an answer that promises cures and cheaper goods while implicitly sacrificing human agency. He wants material benefits such as cures, but also freedom, creativity, time and the power to choose what to build and how to live.

Altman identifies two broad dangers: losing control to AI, including an accident or a deliberate rush toward RSI, and concentrating too much power among a few people. He believes society can navigate between them and increase individual autonomy. Wider access to tools that early adopters already use, he says, would be a meaningful step.