Sam Altman, OpenAI's co-founder and CEO, spent a day at Harvard University on May 1, 2024, speaking to students, faculty, and business leaders. He was on campus to receive the 2024 Xfund Experiment Cup and met with Harvard Business School's Institute for Business in Global Society (BiGS), among others. The following Q&A is drawn from those discussions — synthesised from multiple conversations held throughout the day, edited for length, clarity, and style by The BiGS Fix.
What excited you about AI as an undergraduate?
I really like things that, if they work, really matter — even if they don't have a super-high chance of working. So, it seemed like if AI could work, it would be the coolest, most important, most exciting thing. And so, it was worth pursuing. The expected value was high, even if the chances were low.
If you could go back to the 19-year-old Sam Altman, what would you tell him?
I think that you can just "do stuff" in the world, and this is not well taught. I certainly did not know it when I was 19. The way that progress happens is people just work really hard, decide what they have conviction in and dedicate themselves to that. That is the only way that things happen, and the world gets better. You don't need to wait. You don't need to get permission. You can — even if you're totally unknown in the world, with almost no resources — you can still accomplish an amazing amount.
How did you get so much done when you were unknown?
One of the things that Paul Graham used to say, that never became venerated advice to the degree it should have, is the idea that you should try to be relentlessly resourceful. Surprisingly often, if you just keep looking for attack vectors on a problem in front of you, you can figure it out. I think this is one of the most important skills in life. It's surprisingly learnable or teachable, and it works in almost all scenarios.
Can you give us an example of a time it worked?
We needed to figure out how to get a deal done with a mobile operator and they didn't really work with startups or technology companies in general. We probably tried 30 different paths into this company. At one point, the key decision maker said, "I'm finally going to meet with you because I want you to stop bothering us." You can keep doing that until something works. Most people, at the first ignored email or the first path where you don't find the right person in a company, or at least the second, would just stop. But it was a life-and-death thing for our company, so we were very motivated.
“You don't need to wait. You don't need to get permission. Even with almost no resources, you can still accomplish an amazing amount.”
When you started OpenAI, why did you make it a nonprofit?
If you start a for-profit, you have to theoretically have some idea of what you're going to do to be a moneymaking entity someday. Our initial conception of OpenAI was, let's just be a research lab and see if we can get anything at all to work. It took a long time to get anything at all to work. We made something that could play Dota 2, but that's very hard to build a business around. We made a robot hand that could barely do a Rubik's Cube. But eventually, after a lot of stumbling in the dark, we did figure out something that turned out to be a product and a real business.
We also found that you don't get to choose where the science goes. You just have to follow it. And where we had to follow it turned out to require gigantic amounts of resources to keep pushing progress, so we needed a business model to fit that.
What do you say to workers right now who fear that they may lose their jobs in the next five years?
They may — although I think that AI will not eliminate all the jobs in the way most people think. In every technological revolution, people predict the end of jobs, and it never happens. I don't think it will happen this time, but jobs will change. In most jobs, people will use AI tools to do their jobs better and faster. But some jobs will totally go away. New ones will get created. The shape of jobs will change.
I think it is important to be upfront and honest with people that we expect this to happen — jobs to change, some jobs to go away, new jobs to get created — and also work with our leaders to figure out the social contract and how that's got to evolve given this level of change.
Can you tell us about a difficult product decision you've had to make?
Our product decisions are downstream of the research decisions, and the research directions that we choose to pursue and not pursue are probably the most difficult and most important. On the product side, the behavior of ChatGPT — what it refuses, what it doesn't refuse, and where the limits are of what it will do for you — how we figure out where to set the alignment — those are probably the hardest product calls.
Should ChatGPT give legal advice?
It mostly won't right now. You can get it to do it in some cases, in some ways. There are huge reasons not to do it. And obviously, with the hallucination or general inaccuracy issues with ChatGPT, it seems very reasonable to say it shouldn't do it. On the other hand, there are a lot of people in the world who can't afford legal advice. And if you can make it even imperfectly available, maybe it's better than not.
What I'd like to get to is a world where we don't do things that have high potential for inaccuracies leading to misuse. But as the model gets better, those are less common, and we can give you a dial and you can say, "I really understand I've got to check this advice and it's very much on me if I don't." This is not like clicking through terms of service that no one reads — people are really going to understand it. And then if you want to do it, we find some safe way to do that.
“Personally, I think greatly increasing the rate of scientific discovery is what I'm most excited about. That is how the world gets sustainably better.”
What are the pieces of AI that you're most excited about?
Personally, I think greatly increasing the rate of scientific discovery is what I'm most excited about. I believe that if you zoom all the way out, that is how the world gets sustainably better. And I think doing more of that is awesome. There are a lot of other areas too. I think there'll be incredible AI tutors, incredible AI medical advisors. But personally speaking, I'm so excited about AI for science.
What boundaries should be set on AI tools?
Really what I think is OpenAI should not be making those determinations. There should be a process by which society collectively negotiates: here's how we're going to use this technology. The rules should be uncomfortably permissive, but the defaults don't have to be so permissive. I think it's fine to say, "the default is here, a user can customize within these very broad boundaries that society has agreed with." But there are still some boundaries and there are still some things, particularly as these models get way more powerful, that we're not going to allow. And then within those bounds, the goal of the tool is to serve its user.
Do you think the government could have done OpenAI as well as you have?
In a well-functioning society, I think this would be a government project. But that's a big "if." Given that that's not happening, I think it's better that it happens this way and actually happens.
What are some considerations when it comes to writing AI policy?
First of all, I think policy has worked best when it's downstream of the science. There are all these things that sound like great policy ideas in theory. But if the technology goes in a super different way, you might need a very different set of policy tools.
One idea that I find quite compelling is something like an International Atomic Energy Agency for advanced AI systems. If you have one of the top 10 AI systems in the world, you're going to have the equivalent of weapons inspectors. We would focus the international policy on the catastrophic risk that affects all of us, and different countries get to set their own rules on top of that. I think that's a great idea — and it's workable, if the technology goes like we think it's going to go.
“The difference between what we in the industry expect to happen in the next couple of years versus what the government as a whole believes is going to happen — there's way too much distance between those.”
You were appointed to the first-ever AI Safety and Security Board by the Department of Homeland Security. Why did you choose to participate?
I think one of the most important things to figure out with AI is how government can help play a role with all of us, and the progress we expect over the next couple of years makes that more important and urgent. This seems like an interesting first step to get people from industry talking with people in government who want to figure out how to use this technology and regulate it.
I think it's exciting that this is happening in private industry, but in a different time or a different configuration of the world, it would have been happening in the government. Given that it's not, I think a very close partnership is critical. The difference between what we in the industry expect to happen in the next couple of years versus what the government as a whole believes is going to happen — there's way too much distance between those. One of us is wrong. And it would be good to get to a shared understanding.