Sam Altman and Greg Brockman on AI and the Future (Full Audio)

Progress in AI comes from three sources—compute, data, and algorithms—all of which have been progressing at smooth exponential rates across the industry, and technological progress toward AI is not driven by any one company but is a project of all of humanity combining GPUs, algorithmic ideas, and data processing systems.

“The sources of progress, compute data algorithms, those have been constant over many years and they remain so. We've also done studies on how much algorithmic progress there's been: it's a smooth exponential. And so, I think that there's these inputs that you actually develop almost at small scale, or across the whole industry, that together, combined, you put them together – the best engineering, the best systems, the best algorithmic ideas – and then that is what yields systems like GPT-4. And so I think that the one thing we should be cognizant of as we think about this, you know, these rates of progress questions, is really looking at where the progress comes from. And it's not any one company, it's really this, like, if you look at the supply chain of all of the inputs of the GPUs to all of the new algorithmic ideas, to even the large scale data processing systems, like all those things together, it's pretty massive. It's lots of people involved with lots of different companies and it's really a project of, like, all of humanity.”

When Deep Blue beat Garry Kasparov in chess, people predicted chess would become uninteresting and die, but chess is actually more popular now than ever, and humans prefer watching humans play rather than AIs, suggesting pessimistic predictions about AI's impact are often wrong.

“When Deep Blue beat [Garry] Kasparov, which some people talk about is the start of the whole AI revolution, there were a lot of predictions that chess was totally done, that no one was going to bother to play chess anymore, it was just no longer interesting. And it did affect some things and change things, but I believe chess has never been more popular than it is today. And we don't watch AIs play each other, which would be like far more interesting games or far more complex, better games, whatever you want to call it. We seem to be really interested in what other humans can do in this case.”

Scaling matters profoundly for AI capabilities; there are predictable scaling laws showing smooth exponential relationships between compute/data and capability across many different axes of variation, representing a deep scientific discovery about the nature of intelligence.

“I think there's other things we've also sort of discovered, right? That there's all these scaling laws, you can find a lot of these now, these very smooth plots as you, you know, increase the amount of compute in a model or the amount of data, and there's so many different axes you can vary on, and they all give you these incredibly smooth exponentials. And I think there's something really deep going on that we sort of, from a scientific perspective, have been uncovering about, I don't know if it's the nature of our intelligence, but certainly the nature of this artificial intelligence that we are creating.”

GPT-4 enables semantic understanding and capabilities to emerge from syntactic tasks—as demonstrated in the 2017 'Unsupervised Sentiment Neuron' paper where a model trained to predict the next character in Amazon reviews learned state-of-the-art sentiment classification without explicit instruction.

“We trained a model to predict just the next character in Amazon reviews. And you expect it's going to learn where the commas go, where the nouns are, where the verbs are. But the amazing thing was it learned a state-of-the-art sentiment analysis classifier. It could tell you if a review was positive or negative. And so you see that semantics emerged from a syntactic process. Like, where did the meanings come from? We never told the machine the meanings, it just somehow figured it out by pursuing this task.”

Potential technological overhangs—combinations of technologies that could produce step-function societal change if deployed together but haven't been yet—represent the key risk area where coordination and oversight matter most.

“And I think one source of potential risk is overhangs, like, the more that it's possible, if you were to put these things together, but you haven't actually done it, to produce something that's just like totally going to be a step function for society. That's where I start to get nervous.”

AI is fundamentally different from most technologies because every company, every individual, and every business has language flows deeply embedded in their operations, so adding value to language workflows enables universal adoption.

“I think AI might just be different from that, because it's like every company, every individual, every business is a language business. It has language flows deeply baked in. So if you can add a little bit of value in existing language workflows, then it will just be able to be adopted so broadly.”

AI in medicine presents a paradox: you must be careful about overreliance in medical contexts, but humans aren't perfect either, and sometimes AI suggestions could save lives by prompting second opinions or alternative perspectives.

“You have to be so careful about overreliance in areas like medicine, but humans aren't perfect either, right? That first doctor made a terrible call, it could have had really fatal consequences, and so figuring out how to have the right humans have the right oversight and ultimately, you know, as a patient you actually own the outcome, right?”

Helion Energy represents a transformative technology opportunity that could provide abundant, cheap, clean energy globally, which correlates strongly with quality of life historically.

“Abundant energy – like, truly global scale, abundant, cheap clean energy – would not only have all of the obvious benefits of addressing the climate crisis and everything else in that vein, but the cost of energy is so correlated to quality of life, throughout history, probably more than any other single input I could think of off the top of my head, that it seems like a great thing to fund efforts trying to radically change that cost, which Helion is trying to do, and I am hopeful we'll have great news next year.”

OpenAI built a culture of sweating the details, careful engineering, careful science, and letting small improvements compound over long periods, which is a key ingredient in outpacing larger tech companies.

“One big one is, I think we built a culture of – and this is another thing I'll credit Greg for – of really sweating the details and trying to get the details right and bringing people in who want to work that way. And doing, you know, very careful engineering, very careful science, and letting that compound over a long period of time.”

AlphaGo discovered moves that no human had ever played, and humans then learned from the AI and updated their understanding of the game, suggesting that AI can not just solve problems but generate novel insights humans can learn from.

“You go look at AlphaGo, you know, there's a famous move that no human would ever play. And that that was something that humans then got a bunch of insight into how to change the game.”

Larger-scale AI models achieve better alignment and safety properties than smaller models, meaning safety and safety precautions in the future may be better than they are now, so regulation should account for this trajectory.

“One of the things we've learned with larger scale models is we get alignment better. So the questions around safety and safety precautions are better in the future, in some very arguable sense, than now.”

Previous online education platforms like Khan Academy and Coursera were supposed to democratize education but failed to deliver results for billions of people, so the question is whether GPT-4 and AI will actually achieve the step change needed.

“I think a lot of people might say, 'well, we've had that already, we have Khan Academy, we have Coursera, we have all these online classes that were supposed to be this promise, and we didn't see the realization.' Like, why do you both – and Greg, I'm happy to go to you since you brought it up – why do you think this actually will be the step change for the millions and billions of people who don't have the education that they need right now?”

Specialization is a critical problem in fields like medicine where doctors refuse to answer questions outside their specialty, and AGI should help humans cross-specialize and pool knowledge across disciplines.

“If you go to a doctor – like I remember I had a wrist injury at one point and I wasn't using my wrist and I started to have neck pain and I asked the doctor, 'hey, any idea what this might be?' And he looked at me, he's like, 'I'm a wrist doctor.' [laugh] He was not going to answer my neck question. And we need to have the way out, we need to have a way to cross specialize to actually pool knowledge across these disciplines.”

People using GPT-4 with ChatGPT have reported positive outcomes like saving their dog's life by getting medical hypotheses from the AI while retaining human oversight and professional consultation.

“If you look on Twitter, there's someone who saved his dog's life with ChatGPT, and the story there was that he went to a vet and the vet really didn't know what to do and said, let's just observe this dog. And it just kept going downhill and still the vet didn't want to do anything different, so he presented the medical records to ChatGPT, which very correctly said, 'I'm not a veterinarian, you really need to talk to a vet,' but was willing to give him some suggestions, some hypotheses, some interpretations, some brainstorm. And with that he got the confidence to go to a second vet who was then able to run the test to save the dog's life.”

OpenAI is investing through a fund in startups building on top of OpenAI's technology, including companies like Ambience Healthcare that operationalize AI in healthcare, and Microsoft is deploying OpenAI technology through Nuance in Epic and hospitals.

“We also have a fund within OpenAI that invests in startups building on top of our technology, and so there's startups like Ambience [Healthcare] who are actually trying to really operationalize this. And actually even Microsoft is starting to deploy some of our technology through Nuance in Epic and in many hospitals.”

AI technology for legal services like Augmented—which helps tenants understand eviction notices—provides access to legal understanding for people who cannot afford lawyers, directly addressing inequality.

“I think that applying these services to law, I think that there's a lot of benefit to be had there, with giving access to legal services. And my favorite GPT-3 service was a tool called Augrented, which would help tenants who received eviction notices understand what was in there, right? Because, like, many people in that demographic wouldn't necessarily have access to a lawyer and that you can actually help people do things that they wouldn't be able to otherwise.”

The physical world is more resistant to AI transformation than the digital world; OpenAI had a robotics effort with impressive results (robotic hand that could manipulate Rubik's Cube) but shut it down because digital applications were moving faster.

“I think the physical world is maybe the most resistant right now. We ourselves had a robotics effort and a couple years ago we actually shut it down because you know, a lot of people at the time were like – we had some great field-leading results, we had this cool robotics hand that was able to manipulate a Rubik's Cube, all that good stuff. But we've realized that the digital world is moving so much faster. And so that team actually became our Copilot team.”

GPT-4's capabilities demonstrate that machines can now achieve tasks people thought computers could never do, representing a crossing of a threshold that previous AI systems could not cross, and the test of success will be the next couple of years.

“There's no question that the capabilities of something like GPT-4 are just – like, I think everyone who sees it, you really see that, 'I did not think computers could do this, but now they can.' And I think that there's a second step that's also important, which is not just having the raw knowledge, but can we really steer the machine to do the tasks that we want to reflect our intent and to sort of operate according to the values that society chooses to put into that machine?”

People feel anxiety and fear about the rate of AI change because changes OpenAI has been grappling with for years are now happening in the wider world in a few months, making rapid pace of change understandable as a source of anxiety.

“I think there's a lot of anxiety and fear right now, and I always believe people when they're afraid or mad or whatever, even though I don't think people can always explain the reason. I think people feel afraid of the rate of change right now. A lot of the updates that people at OpenAI, who work at OpenAI, have been grappling with for many years a lot of the rest of the world is going through in a few months. And it's very understandable to feel a lot of anxiety in that moment.”

Neural networks were considered a dead-end career path in machine learning when Sam was a student, making the empirical success of deep learning surprising and a form of magic.

“I went to school at a time when they told you if you wanted to study machine learning as a student, that the only way to be guaranteed to have a dead end career was to work on neural networks.”

Sam Altman has publicly agreed that AI should be regulated, reversing what might have been a snarky expectation that tech entrepreneurs resist regulation.

“A few weeks ago on Twitter, I think someone snarkily tweeted at you like, 'well, next thing you're going to say is we should regulate AI!' And you were like, 'yeah, we should regulate AI!'”

Despite ChatGPT appearing to come out of nowhere, the model had been deployed for almost a year before as GPT-3, suggesting people didn't appreciate the continuity of progress because of media amplification.

“Even the fact of seeing the reaction to ChatGPT and how much it felt to people like it came out of nowhere, whereas we've had years of seeing this technology get a little bit better, a little bit better, a little bit better. The model on ChatGPT, that wasn't new, that had been out for almost a year at that point.”

According to physician testimonies, AI-assisted documentation and administrative tasks allow doctors to spend less time on iPad-based record entry and more time on actual patient care, addressing a major source of physician dissatisfaction.

“Everything that I did here today would take me so much longer to do. This will help me be more efficient so I can spend more time actually talking to patients, looking them in the eye… To be able to gather and garner insights into the patient that we may not have been able to before is very exciting to me.”

A key insight about AI's impact on science and expertise is that the world's brightest minds spend enormous time on mundane, intellectually unstimulating work, and freeing them from this drudgery will dramatically accelerate progress across all fields.

“You just realize that the world's brightest minds are spending all their time wading through this not very desirable, not very intellectually stimulating work, and I think that we're just going to see people achieving more across the board.”

A naive pause on AI development is unlikely to help much; what's needed instead is figuring out effective regulatory approaches, safety standards, and governance rules that will actually work in real-world contexts.

“I don't think a pause in the naive sense is likely to help that much. You know, we spent way more than six months, by the way, not way more, somewhat more than six months aligning GPT-4 and safety testing it since we finished training. So, like, taking the time on that stuff is important. But really I think what we need to do is figure out what regulatory approach, what set of rules, what safety standards will actually work, will actually, in the messy context of reality, work.”

Building AGI in secret and deploying all at once has appealing properties—it lets developers focus on technical safety without worrying about society—but it's fundamentally flawed because you need to solve both the technical safety problem and the social coordination problem, and you only get one shot to get it right.

“I think a lot of people look at it and say, 'hey there's a technical safety problem of making sure the AI can even be steered, and there's a society problem. And that second one sounds really hard, but you know, I know technology so I'll just focus on this first one.' And that original plan has the property that you can do that. But that never really sat well with me because I think that you need to solve both of these problems for real, right? How do you even know that your safety process actually worked. You don't want it to be that you get one shot to get this thing right.”

The fundamental purpose of building technology is to improve humanity, improve human lives, and enable people to achieve more.

“why do we build technology in the first place? And fundamentally, it's to improve humanity, to improve human lives, to be able to achieve more.”

Keeping the rate of change constant rather than developing AGI secretly and deploying it suddenly is much better because it allows people and institutions time to adjust, update, provide feedback, identify benefits and harms, and shape the technology's evolution.

“I think what we believe in very strongly is that keeping the rate of change in the world relatively constant, rather than, say, go build AGI in secret and then deploy it all at once when you're done, is much better. This idea that people relatively gradually have time to get used to this incredible new thing that is going to transform so much of the world, get a feel for it, have time to update – you know, institutions and people do not update very well overnight – to be part of its evolution, to provide critical feedback, to tell us when we're doing dumb mistakes, to find the areas of great benefit and potential harm, to make our mistakes and learn our lessons when the stakes are lower than they will be in the future.”