Most of the world’s most sophisticated artificial intelligence programs are written by for-profits, like Facebook, Google, and Google sister company DeepMind. But in 2015, Elon Musk founded an exception: OpenAI, a nonprofit with the mission to do cutting-edge AI research and bring the benefits to everybody.
Since then, the young organization has racked up some impressive accomplishments. It recently built a language-generating system called GPT-2. It writes news articles that are, at a glance, difficult to tell apart from real ones, and this weekend their AI became the first to beat an esports world champion team with its sweeping Dota 2 victory.
It has also shifted directions, in a way that’s left some outside observers nervous. Musk left the board. OpenAI’s safety team concluded that open-sourcing all of their work might, rather than advancing humanity’s common interests, invite trouble; when they developed GPT-2, they didn’t release it publicly, expressing worries that it’d be easy to misuse for plagiarism, bots, fake Amazon reviews, and spam. And last month, they announced a significant restructuring: Instead of a nonprofit, they’ll operate from now on as a new kind of company called OpenAI LP (the LP stands for “limited partnership”.)
The team wanted to raise billions of dollars to stay on the frontiers of AI research. But taking investment money would be a slippery slope towards abandoning their mission: Once you have investors, you have obligations to maximize their profits, which is incompatible with ensuring that the benefits of AI are widely distributed.
The solution? What they’re calling a “hybrid of a for-profit and nonprofit”. The company promises to pay shareholders a return on their investment: up to 100 times what they put in. Everything beyond that goes to the public. The OpenAI nonprofit board still oversees everything.
That sounds a bit ridiculous — after all, how much can possibly be left over after paying investors 100 times what they paid in? But early investors in many tech companies have made far more than 100 times what they invested. Jeff Bezos reportedly invested $250,000 in Google back in 1998; if he held onto those shares, they’d be worth more than $3 billion today. If Google had adopted OpenAI LP’s cap on returns, Bezos would’ve gotten $25 million dollars — a handsome return on his investment — and the rest would go to humankind.
So that’s the idea. The devil, of course, is in the details. OpenAI’s mission is “discovering and enacting the path to safe artificial general intelligence” (AGI) — that is, an artificial intelligence that has human-like problem-solving abilities across many different domains. That’s a huge ambition — one that is likened to the invention of electricity, the internet, or the industrial revolution. Is there really any chance OpenAI can build an AGI? Should they be trying, given the risks? How accountable is OpenAI to the world they’re planning on transforming?
I sat down with OpenAI co-founders Greg Brockman and Ilya Sutskever to discuss these and many other issues. Here’s our conversation, edited for length and clarity:
I read the Open AI LP announcement, where you’ve stated your intent was to raise billions of dollars in order to make progress towards artificial general intelligence. What was the process that led you to that? And to do it in the structure you did, with the LP?
Making advances in AI, towards AGI, is not cheap. You need truly giant amounts of comput[ing power], and you need to attract and retain the best talent. In terms of compute, specifically, we’ve seen that the amount of compute required to get the best results has been growing extremely rapidly each year. And at some point we realized that we’d reached the limits of our fundraising ability as a pure nonprofit.
And so we sought to create a structure that will allow us to raise more money — while simultaneously allowing us to formally adhere to the spirit and the letter of our original OpenAI mission as much as possible.
We think this is a great structure for AGI, but we don’t think it’s just for AGI. There are other possibly transformative technologies coming on the scene — things like CRISPR. It’s important these technologies get developed but you don’t want them subject to pure profit-maximizing for the benefit of one company. So we hope to see this structure be adopted by other people.
We founded OpenAI in 2015. And we spent about two years really trying to design the right structure — one year of figuring out what we think the path is [to being a leading AI organization], and then another year figuring out how you’re supposed to do that while still retaining the mission.
Your understanding of how to achieve your mission has evolved a ton over time. It definitely seems like a lot of people’s understanding of what OpenAI is doing is shaped by some of the early messaging in a direction that doesn’t actually reflect you guys’ understanding of the mission at this point.
A lot of people think you’re about open sourcing progress towards AGI.
OpenAI is about making sure the future is going to be good when you have advanced technologies. The shift for us has been to realize that, as these things get really powerful, everyone having access to everything isn’t actually guaranteed to have a good outcome.
You can look at deepfakes [convincing fake photos and videos made with AI]. Is the world better because deepfakes are out there? It’s not obvious, right?
And so our focus, instead, has really shifted to thinking about the benefits. The way technology has evolved, with the right big idea you can generate huge amounts of value — but then, it also has this wealth-concentrating effect. And if AGI is built, by default, it will be a hundred times, a thousand times, ten thousand times more concentrating than we’ve seen so far.
You’ve gotten some pushback on the move away from transparency. You’ll probably get more as you start to publish less.
If your mission is publish everything, it’s easy for the public to tell whether you guys are still motivated by your mission. I can just check if you’re publishing everything. If your mission is to distribute the benefits, I don’t have a way of evaluating whether you’re still committed to your mission. I have to wait to find out if I can trust you.
I think this is a really good point. And there’s a reason we didn’t just go in and make very quick changes. There’s really a reason we spent a long time thinking about who we are. We looked at every legal structure out there. And in some ways, a nonprofit is just great for having a pure mission that’s very clear how it works. But you know the sad truth is that not enough gets done in a nonprofit, right? And in a for-profit — I think too much gets done there.
So the question was, how do you find something that gets the best of both?
One gap in our current structure that we do want to fill is representative governance. We don’t think that AGI should be just a Silicon Valley thing. We’re talking about world-altering technology. And so how do you get the right representation and governance in there? This is actually a really important focus for us and something we really want broad input on.
One thing that struck me, reading some of the critical reactions to Open AI LP, was that most of your critics don’t believe you that you’re gonna build AGI. So most of the criticism was: “that’s a fairy tale.” And that’s certainly one important angle of critique.
But it seems like there’s maybe a dearth of critics who are like: “All right. I believe you that there’s a significant chance — a chance worth thinking about — that you’re going to build AGI ... and I want to hold you accountable.”
I think I go even one step further and say that it’s not just people who don’t believe that we’re gonna do it but people who don’t even believe that we believe we’re going to do it.
A lot of startups have some language about transforming the whole world on their Web site, which isn’t that sincerely meant. You guys are saying “we’re going to build a general artificial intelligence” —
We’re going to do everything that can be done in that direction while also making sure that we do it in a way that’s safe.
It’s hard to tell how long it will take exactly. But I think it’s no longer possible to be totally confident that this is impossible.
I think it’s interesting to look at the history of technological developments. Have you ever read Arthur C. Clarke’s Profiles of the Future?
Don’t think so.
It’s such a great book. This is [Clarke] trying to say — let me predict what the future’s going to be like. And he starts by looking at the history of inventions.
He goes through flight, space flight, the invention of the atomic bomb, the invention of the incandescent lamp — looking at all of these and saying “what was the climate?” How did people feel at the time these technologies were coming on the scene?
The incandescent bulb was fascinating because Thomas Edison, the year before he created the lamp, had announced what he was doing. He said, “We’re gonna do this, it’s gonna be great” and gas securities in England fell. So the British Parliament put together this committee of distinguished experts who went to go talk to Edison, check out all the stuff.
They came back and they were like this is totally bogus. Never going to happen, everything’s fine. A year later he ships.
And the thing is — the naysaying prediction will be right most of the time. But the question is what happens when that prediction is false.
One way to think about what we are doing is [taking out] a global insurance policy against sooner-than-expected AGI.
But then to add to why it’s even reasonable to talk about AGI in the first place today — if you go back in history, they made a lot of cool demos with little symbolic AI. They could never scale them up, they were never able to get them to solve non-toy problems.
Now with deep learning the situation is reversed. You have this very small set of tools which is general — the same tools solve a huge variety of problems. Not only is it general, it’s also competent — if you want to get the best results on many hard problems, you must use deep learning. And it’s scalable. So then you say, “Okay, well, maybe AGI is not a totally silly thing to begin contemplating.”
So the thing that got me worried was that I talked to some people who weren’t worried, and I said, “All right, what would scare you? What would make you say, ‘Well, you know, maybe we are 10 years away’?” And they said things like “unsupervised learning” — that is, learning from unstructured data. And now GPT-2 is doing unsupervised learning. You know, 10 years ago everybody was saying “AI can’t even look at things.” We’ve basically solved that one.
It’s a very funny psychological phenomenon, because you just adapt so fast and you take it for granted. It’s understandable, because technological advance and development is so difficult to think about.
There’s a great xkcd from 2014:
One word that you used earlier was worried. I think that it’s interesting how when people talk about AGI, often they focus on the negative. With technology generally it’s much easier to focus on those negatives. But the thing with AGI — and this is true of all technologies, but I think AGI in some ways is the most extreme form of technology that we’ve conceived of so far — is that the upsides are going to be so huge.
You think about planetary scale problems that humanity just doesn’t really seem to even have a hope of solving, how about, global health care for everyone? And so that’s what we’re really excited about. We’re excited about the upside. We’re cognizant of the downsides, we think it’s important to navigate those as well.
But if people are just looking at the downsides and not thinking about — every day 150000 people die. It’s possible that with better technology, those are mostly preventable deaths.
That’s right. Exactly.
If you’re not thinking about that, you’re not gonna understand the urgency.
We’ve been talking mostly about the ability of Open AI LP to ensure that benefits get distributed. But distributing the benefits is really far from being the only big problem that people see on the road to AGI. There’s risks of accidents, risks of locking us into bad futures, stuff like that.
We see three main categories of risk from AGI. In some ways they all boil down to one thing which is AGI’s ability to cause rapid change. You can think about the Internet. In a lot of ways, we’ve had 40, 50 years to have the internet play out in society. And honestly that change has still been too fast. You look at recent events and — it’d just be nice if we’d spent more time to understand how this would affect us. With AGI you should view it as almost this more compressed version of what we’ve seen.
The three categories we really see are — one, systems that pursue misspecified goals, so does what no one wants. That’s a “careful what you wish for” scenario.
The second is systems that can be subverted. So you actually build it to do the right thing but someone hacks into it and does bad things. And the third one is: so we get those two problems right. We get the technical stuff right. It can’t be subverted, it does what we intend but — somehow society doesn’t get better. Only a few people benefit. Everyone else’s lives are the same or even worse.
And there is a fourth one, which is misuse.
And if you care about all those risks and you want to navigate all of them, you must think about the human factor and the technological factor. We work to make sure that governments are informed about the state of AI and can reason about it as correctly as possible. And we work on building the AGI itself. Making sure that it is safe, that its goals have been specified and it follows its goals.
And I think that you’ll see a lot of that weirdness with our structure — it doesn’t look like structures that have existed, it’s not really precedented. A lot of that is because we consider all these risks and we think — how does that affect how to do this?
Sign up for the Future Perfect newsletter. Twice a week, you’ll get a roundup of ideas and solutions for tackling our biggest challenges: improving public health, decreasing human and animal suffering, easing catastrophic risks, and — to put it simply — getting better at doing good.