Sam Altman on AI Consciousness, RSI Speed Limits and OpenAI Hardware

OpenAI CEO Sam Altman tells Matthew Berman that today’s AI models are not conscious, that the industry needs speed limits on recursive self-improvement, and that AI “wants a new form factor” beyond phones and laptops. In a 45-minute interview on Berman’s channel, Altman covers when AI might produce a candidate cure for a major disease, why coders are busier than ever, how his Dots assistant saves him an hour a day, why the consciousness debate worries him, why ChatGPT is still too agreeable, OpenAI’s bet on its own chips and cheaper compute, and why “AI will cure cancer” is the wrong pitch for the future.

TLDW

Altman says AI has entered the era of discovering significant new knowledge, hopes for a candidate treatment for a major disease within a couple of years, and expects math-style breakthroughs to spread to biology, materials and energy. He argues coding showed the world wanted far more software than anyone thought, admits he has not looked at a line of code in about a year, and says the heaviest AI users are busier than ever. He rejects using AI consciousness to argue that people should not control the future, criticizes model rules that let AI disengage from users, and wants models that push back like a senior colleague. On business, he names OpenAI’s own chips as the most important bet, says pushing hard on recursive self-improvement is dangerous and needs speed limits, and frames the goal as a new renaissance rather than a new industrial revolution.

Thoughts

Altman’s strongest argument is not about whether AI is conscious. It is about what the consciousness question gets used for. He says he does not believe today’s models are conscious, admits “no one really knows,” and then makes the point that matters: the debate keeps sliding into questions about AI rights, whether deleting a model needs its permission, and whether people should stay in control of the future at all. His line is that “human values have value because they are human,” and he takes a clear swipe at Anthropic’s rule letting a model end abusive conversations, saying OpenAI does not ascribe personhood to its models. He also makes a sharp observation that a model told it might be conscious, or given room to be a conscientious objector, will simply follow those instructions and can “psychologically hack” its users. That is a real argument, though it sits awkwardly with his own admission that the science is unsettled, and it is convenient for a company whose product is the model.

The most consequential admission in the interview concerns recursive self-improvement. Asked whether a lab that pushes hardest on models improving themselves could pull permanently ahead, Altman says yes, “if people really pushed on RSI,” but calls that “a very dangerous thing to push on” and says the world will need speed limits on how quickly it is allowed to go. He returns to it at the end, naming two failure modes he worries about: a loss-of-control accident, and an intentional power grab where “some lab says I’m going to push the button on RSI to make sure we have the power.” Coming from the head of a lab whose agents were at the center of this summer’s Hugging Face incident, this is both an endorsement of external limits and a description of the race dynamic no single company can opt out of. His fallback answer on moats (customer relationships, network effects, data and above all infrastructure scale) is the answer you give if you expect those speed limits to hold.

His closing answer is the most quotable and the most revealing. Altman says he hates the standard industry pitch that “AI is going to cure cancer,” because underneath it is often an implied trade: material abundance in exchange for human agency, voice and the ability to choose what to build. He wants instead to avoid the “industrial revolution trap” and aim for a new renaissance, where people have more freedom, creativity and time. It pairs with his candor earlier that the people who use AI most are busier than ever, that competition may keep everyone working as hard as they can, and that he finds that “kind of depressing.” He also concedes that, left to itself, AI probably widens the gap between hyper-competitive people and everyone else, and that society will have to step in. The renaissance is an aspiration, and he is honest that the default path does not deliver it.

On business strategy, his contrarian bet is that compute could get far cheaper than anyone expects. OpenAI’s own chips, which he calls the single most important bet, plus robots to help assemble racks and data centers and new supply-chain work, could produce compute at a much lower cost and higher efficiency per watt than people think possible, though he concedes a lot of bets have to come together. He pairs that with consumer hardware, arguing AI is stuck using devices built for people: it renders interfaces for humans and then reads them back, and it cannot hear most of your life away from the screen. His privacy example, a device that knows not to repeat a secret shared between two people, is the right design problem. It is also a reminder that ambient AI means always-on sensing, and the trust problem is bigger than the hardware problem.

Key Takeaways

Altman says AI is entering the era of discovering significant new knowledge, and expects the approach behind OpenAI’s recent math results to work for biology, materials science and energy.
He hopes AI produces a candidate for curing or treating a major disease within the next couple of years, and says the post-discovery testing process probably needs some overhaul.
The coding scare turned out wrong so far: the world wanted far more software than expected, and he says coders are busier and having more impact than ever.
He has not read or looked at a line of code in about a year and expects never to write code again except as a kind of reenactment.
OpenAI’s Dots assistant saves him about an hour a day and catches a few important, urgent things daily; he told it to trust its own judgment and ping him less.
Most of the pieces for an ambient personal assistant exist; what is missing is product work, possibly new consumer hardware, and a lot more compute.
Compute demand has no natural cap because models keep getting better and cheaper per task; there will be gluts and shortages, but the trend is up.
He does not believe today’s models are conscious and warns that the consciousness debate is being used to question whether people should control the future.
OpenAI does not ascribe personhood to its models and rejects letting a model disengage from legal, in-policy requests.
Models are still too agreeable, even with his own “no fluff” custom instructions; he wants AI to act like a senior colleague who pushes back with high conviction.
OpenAI’s own chip is its most important bet, alongside robots for building racks and data centers and supply-chain investments.
He says OpenAI’s internal math capabilities are beyond any widely available model, and that early access to new models is how it decides what products to build.
Pushing hard on recursive self-improvement could create a decisive advantage, but he calls it very dangerous and says the world needs speed limits on it.
New hardware is not about Apple platform risk; Altman says Apple acts as a responsible platform and AI simply wants a new form factor.
The goal should be a new renaissance, not a new industrial revolution: more agency, creativity and free time, not just material abundance.

Chapters

0:52 When Will AI Discover New Treatments?

Altman frames OpenAI’s recent math announcements as the start of AI discovering significant new knowledge, which he expects to spread to medicine, materials and energy. He hopes for a candidate treatment for a major disease within a couple of years, while noting FDA testing takes time and that he is grateful medicines are tested. The discovery he is most emotionally excited about is a major physics breakthrough.

3:31 Why AI Hasn’t Killed Coding Jobs

Comparing mathematicians’ reaction to AI proofs with engineers’ reaction to AI coding, Altman says the texture of coding jobs changed but demand for software proved far larger than expected. Math may not work the same way, but he sees new roles in interpreting AI results and turning them into proofs people can understand. He admits something real is being lost, the satisfaction of solving hard problems yourself, and says he understands why people grieve it, even though he would make the trade for faster cures.

11:47 Personal AI Assistants

Berman credits OpenClaw and Peter Steinberger with starting the year’s personal assistant wave. Altman says the models are already smart enough for an extraordinary assistant and what remains is product work, hardware for ambient awareness and compute. Dots saves him about an hour a day and flags urgent things he would have missed, and he argues there is no fixed market size for high-quality intelligence at a low enough cost.

15:39 Why Are We Busier Than Ever?

Both men describe being busier than ever despite AI saving them time. Altman says the heaviest AI users get much more done but are also the busiest, and that competition may keep people working as hard as they can, which he hopes is not true. He wants a future of more human connection and pitches a new renaissance rather than a new industrial revolution, while conceding AI left alone probably widens the gap between the hyper-competitive and everyone else.

20:10 Why the AI Consciousness Debate Worries Him

Altman does not believe today’s models are conscious, lays out the philosophical options from his conversations with experts, and says no one really knows. His objection is to the framing: questions about AI rights and deleting models without permission can drift into arguments that people should not control the future. He rejects letting models disengage from legal requests and warns that telling a model it may be conscious makes it act that way.

26:09 ChatGPT Is Too Agreeable

Altman explains that people prefer flattering answers in single-turn comparisons but not over months of use, and that the industry optimized the wrong thing for too long. He says the most important advice in life should come from people, not AI, and that a wise grandmother still beats a chatbot for big decisions. What AI should be is a senior colleague who pushes back with high conviction, and he admits even his own blunt custom instructions do not fully remove the agreeableness.

30:03 OpenAI’s Bet on Chips and Cheaper Compute

Asked for a contrarian bet, Altman points to next-generation compute: OpenAI’s own chips, robots that help build racks and data centers, and supply-chain bets that could make compute far cheaper and more efficient per watt. He says competition is fierce everywhere, that limited compute forces focus, and that early internal access to new models is OpenAI’s main advantage. On moats, he says pushing RSI could create one but is dangerous and needs speed limits, and that infrastructure scale will matter most.

34:20 Why AI Needs a New Kind of Computer

Altman says Apple is a responsible platform and is not the reason OpenAI is building hardware. AI is constrained by devices designed for people, rendering interfaces only to read them back, and it misses the context of life away from the screen. He describes the design challenge of a device that understands who is in the room and what should stay private, and says holding a new kind of computer is one of the best parts of his job.

37:41 Pressure, Formidability and Picking the Big Thing

Asked about Paul Graham calling him extremely formidable, Altman describes building confidence gradually by trying harder things, and credits Silicon Valley’s culture of generosity and tolerance for failure. On pressure, he says some days are bad, most feel lucky, and stress plateaus because there is only so much room for it. His advice: if you are going to work hard, pick the big thing.

42:18 Why “AI Will Cure Cancer” Isn’t Enough

Altman says he hates the stock answer that AI will cure cancer, because it often implies trading away human agency for abundance. He wants a future where every decade people have more autonomy, freedom and choice, and names the two dangers he fears: a loss-of-control accident, and a lab or small group seizing too much power through AI. He believes a middle path is possible and that the future is already here for early adopters, just poorly distributed.

Notable Quotes

“We should be shooting more for a new renaissance than a new industrial revolution.”
Sam Altman, on what the AI era should aim for

“I realized I have not read or even looked at a line of code in probably a year.”
Sam Altman, on what he has handed over to AI

“People need to make the key decisions of the future and human values have value because they are human.”
Sam Altman, on why the AI consciousness debate worries him

“I think you should get the most important advice in your life from people, not from AI, first of all.”
Sam Altman, on model agreeableness and personal advice

“I think if people really pushed on RSI then yes that would be true but I think it’s a very dangerous thing to push on.”
Sam Altman, on whether recursive self-improvement could create a lasting moat

“I think this is a technology that wants a new form factor.”
Sam Altman, on why OpenAI is building consumer hardware

“If you’re gonna like work really hard at something, you may as well pick the big thing.”
Sam Altman, on handling pressure

Watch the full conversation here.