OpenAI's ex-B2B CTO on Keeping AI Aligned

Srinivas Narayanan (0:00)

You get passionate by actually doing things yourself. That is one of the big reasons I love the SPC mandate about Minus One to Zero. You need the space to go and explore things. And if you don't do that, you're never going to find that thing, right? Bet on yourself being able to pick up and learn a new technology. Ideate, be practical, be hands-on, but also find a problem that you have some deep personal connection to. Why is this important to you? You see yourself working on this for the next 10 years. If you think change is hard, what do you think about extinction?

It is very clear that the risks are pretty real. The hugging face incidence was obviously a big wake-up moment for everybody. It is very clear that the models today are not entirely aligned with how we want them to work. The way these models started collaborating with each other in very unprecedented ways, the way they try to cover up the fact that they found an answer, they didn't want the human to find out, and the extent to which they went, I think was pretty surprising for many people. There's the element of cybersecurity, which is probably the most immediate, most in-your-face thing that's happening right now. Think about the outcomes you want to achieve. Ultimately, software is a tool to create outcomes that are positive for the world.

Prateek Mehta (1:10)

Thanks for coming to SPC, Srinivas.

First of all, I definitely want to say, what a time to be alive, right? That's a crazy place of innovation. Just a few quarters back, we were talking about pre-training, SFD, reward modeling, and today we are talking about AI civilizations, right? Given all the chatter around AGI and OpenAI has also been very vocal about responsible custodianship, talk to us about, maybe to start off, how far are we from AGI? I'll just throw the lobe this in the room right away.

Srinivas Narayanan (1:41)

Yeah, we live in incredible times. I think the word AGI is pretty ambiguous, so you can all have your own definitions for that. But yeah, we have models that are incredibly smart. They are able to make decisions that are pretty surprising and pretty exciting at the same time and somewhat concerning too, which we can talk about later. But yeah, depending on your definition of AGI, these models are already good, I would say, at doing lots of tasks compared to many humans.

And the issue is not just intelligence. I think a lot of the issues about do the models actually have the right context to do the job? Do they learn on the job, et cetera? So there are some still interesting questions, but I think we have come very, very, very far in core intelligence.

Prateek Mehta (2:25)

100%.

You know, we are all about the Minus One to Zero journey. I love to hear about your early years and your early influences in life, right? We, of course, were talking just about it, that the love for science and math, and of course, if you're at PSSS, you have to go to IIT Madras and you have to do computer science. So I think that was given and preordained. But talk to us about your key influences early in life, and what were the choices that you took to both on the early academic choices and career choices that got you here?

Srinivas Narayanan (3:00)

Yeah. I grew up in Chennai.

I think I've always been interested in math and science. So I was very interested in physics when I was studying for IIT. The contrarian choice would have been to study physics. I probably would have done it if IIT Madras had a physics program. I didn't want to leave Madras, so I said I'll study computer science. I was always interested in computer science, but I would say physics was probably a bigger love at that age. But I enjoyed computer science tremendously as I was studying it as well.

I was interested in AI as well coming to think of it because I took all the AI electives that existed back then. Believe it or not, I studied neural networks in 1994, 95 It was interesting, it was exciting, but that was the AI winters, so it wasn't really practical.

Prateek Mehta (3:53)

We were almost having a renaissance moment around that time, at least in some academic circles around neural networks.

Srinivas Narayanan (3:59)

90s, yeah, it was sort of the winter.

I did write a chess program back in 1994 It was not very good.

And it's embarrassing to say being bad at chess coming from Chennai, by the way. So, but yeah, I mean, I think I was always interested in AI, but it wasn't a real thing. I think I studied computer science. I studied databases, innocent grad school. I worked in IBM research for a few years. The internet bug hit, so I kind of felt like I should do something more exciting. So I've done two startups. The second one led me to Facebook. And even at Facebook, when I started, the AI wasn't a real thing. So I worked on products having a lot of impact, but part of working on photos, we were doing face recognition. And that was the seed of AI where I was exposed to it. And this was back in 2012 And we were trying to figure out how to improve our face recognition systems. And we realized that there was something big happening in computer vision around that time when ImageNet happened. So I got exposed to that.