‘Our mistake was definitely just being slow to respo

It’s been four months since OpenAI debuted its chatbot, ChatGPT, and so far the hype train shows no sign of slowing down. These days everyone from teachers to journalists to HR departments are experimenting with generative AI. But there’s also been a wave of criticism over the tech, including some calling out the chatbot for generating answers with political biases.

Things hit a crescendo last week, when Elon Musk announced his plans to build an “anti-woke” rival to ChatGPT. In response, Greg Brockman, OpenAI president and cofounder, told The Information on Thursday that his company “made a mistake” in building in safeguards to prevent responses that many would find inflammatory.

Fast Company caught up with Brockman at South by Southwest in Austin, where he discussed exactly how the company plans to course-correct. The conversation has been edited for length and clarity.

Fast Company: When you spoke to The Information the other day, you said OpenAI had made a mistake in implementing safeguards to prevent the chatbot from producing responses that could be deemed offensive. So what’s the next step? How does OpenAI course-correct?

Greg Brockman: So, I think our mistake was definitely just being slow to respond. And over the past month the team has really made a lot of progress. We’ll be releasing new models soon. We actually wrote a blog post about our overall strategy and plan, which is that, number one, we think that we should have better defaults. Our goal from the beginning has been to have a bot that treats all the mainstream sides equally. It’s kind of neutral and doesn’t really latch on to any individual side. But if you’re a user, and you want to customize, you should be able to within broad bounds. And the important thing about those balances is, well, who decides them? And we think that should be something that is decided through some legitimate sort of collective AI process. And we’re actually working on that, too. 

I don’t think we’re going to have a perfect answer to this yet, but we’re trying to iterate to figure out something that society as a whole can help decide in terms of what should an AI not be able to say? And then to have technological answers: “Well, if you’re within those bounds then you should be able to build an AI that leads in whatever direction you want.” And it’s not just about politics, right? If you’re building a tutor that you want to never tell a student the answer, no matter how much the student asks, you should be able to build that too. So having something that reflects your values, your goals, whatever the task is, and to have that in the system is important.

FC: Is there a way to do that while still controlling for hate speech?

GB: Yeah, for sure. You should try out models that will be coming out. What we have coming over the upcoming weeks and months, I think will make a lot of progress. 

It really should be. If you want to have something that seems a little backwards, that’s on you. If you want to have something that promotes hate speech or violence, you can’t; it will just refuse. I think that is the shape of the future—that’s something we think is not up to one company or one individual to choose. We think that’s something that needs to be society-wide.

FC: Some people are concerned about general intelligence. You’ve spoken to Congress about the importance of developing guardrails and policies before we get to that point of superintelligent AI. Have we, in fact, developed those frameworks and rules?

GB: I think we’ve made a lot of progress on safeguards, both implementation-wise and conceptually. I think there are techniques, like, for example, reinforcement learning from human preferences, that are important not just for having an AI that pursues arbitrary goals but also for being able to communicate to the AI: “No, no, this is what you really should do.” 

There are other conceptual frameworks—for example, iterative amplification. As you start to have AIs that are solving tasks that humans can’t even evaluate—which you want, right? You want them to go off and solve super hard legal problems, or do a bunch of physics or something—it’s hard to even say, “Is this good or not?” How can you have AIs help with that process? If you have an AI that’s trustworthy, and you have another AI that is a little bit smarter but maybe you’re not sure if it’s trustworthy yet, well, you can have your trustworthy AI kind of help with that process of evaluating. 

We’ve made a lot of progress on all of these systems and ideas and processes. But I think that the thing we were totally missing in the early days was the idea of iterative deployment. So the original vision I think a lot of people had was you build the AI, and then kind of do all the safety work, and then push “go,” and it either works or doesn’t. That never felt right to me; it never felt to me like that was the way to a good future. So since 2020, we’ve been iteratively deploying our systems, and that has given us the most insight and confidence about where they go wrong, and where people misuse them. A lot of our ideas for what could have gone wrong with GPT-3 were really focused on misinformation, and in reality, the biggest problem that we ended up with is people generating all sorts of spam for different medical drugs. So I think there’s a mismatch between theory and what happens in cold, hard reality. Closing that gap, I think, has been the most important thing.

FC: Since we’re at South by Southwest, which is sort of futurist in tone, what’s most exciting to you right now about AI?

GB: I’m very flattered that everyone’s so excited about generative AI. I think we still have a little bit to go before it can add the kind of value that I think it frankly will. But I’m glad that everyone’s exploring and trying to find places to integrate it into their products, build new products on top of it, and just use it in their lives. The thing I’m really excited about is amplifying knowledge work; if you have a task that’s really hard—if you’re trying to write some code, or trying to communicate with a coworker—to be able to make that faster and more effective just means that you can be in the manager’s seat and doing much less of all the drudge work. 

FC: We’re seeing a sort of running list of new AI products coming out seemingly every day, and new investments and funds. Are there any red flags to you about this influx of newcomers?

GB: I think it’s important not to do it to be a bandwagon-hopper. That I think is very, very important. The way that I see this being successful for companies, is people who figure out a use case, something within their company or product, that can get value from these systems. You’ve got to start small and build confidence and approach it right. You’ve got to be ready to walk the walk in terms of thinking about safety, in terms of thinking about how to mitigate overreliance on these systems when they’re not perfect yet. And there’s a lot of investment that must be made in order to get to the point that you have something that’s worthy of shipping.