00:31:59 · Parag Agrawal
[00:31:59] S2: Yeah, no, I think today there's a lot of people doing a lot of bot detection. The word bot detection, I think is a few different things. It is crawlers. It is actual agents working on behalf of a human or a business. It is malicious bots, which are trying to do harm. So I think we have to use that word now a little bit more carefully than we have in the past because historically people have mostly had negative connotations with that word in that, like it's not serving anyone. Yeah. Today there's a category which is actually highly useful for you or for me. I would like my agents to inherit privileges I have. So I think we should be careful. I do think there is like a bot detection market is doing well right now and doing well right now because we haven't figured out how to do incentive alignment. And there's a a separate kind of incentive alignment needed. Let's call it on the action side. So earlier I spoke about the incentive alignment needed on the content side, which is like, why do I let you read my content? The other side is what happens if I let your agent buy on my store And you're starting to see divergent strategies there. I'm of the belief that we need to figure out how to solve problems so that all agents, not just the ones built by a large lab can both access all content and take all actions that their owner is allowed to. Right? But we're not there right now.
00:33:46 · Molly O'Shea, Parag Agrawal
[00:33:46] S1: One of the biggest topics behind growth is cyber security and active threats. How do you I know you said you're an optimist, but realistically, like, how do you see the biggest concerns around this? [00:34:02] S2: No, I think cybersecurity risks are real. I think agents help both defenders and attackers. Problem always is that the people who rush ahead to adopt agents for defending themselves, I think will be fine because attackers like to find the weakest link and there will always be a weakest link. So broadly, that's what I think. Now, the security risk for something highly sensitive, like the model weights of a frontier model are a whole different conversation. Yeah. But by and large, my intuition is that we will use coding agents aggressively to secure. But along the way, there's going to be a cyber event every few weeks of consequence, which will drive action and urgency in getting ahead on defence. We're going to spend a bunch of influence on both attack and defence. [00:35:21] S1: Do you see any of this analogous to your time at Twitter on consumer internet? Obviously, that's a huge, one of the largest social media platforms, and so, that creates a lot of vulnerabilities.
00:35:21 · Molly O'Shea, Parag Agrawal
[00:35:21] S1: Do you see any of this analogous to your time at Twitter on consumer internet? Obviously, that's a huge, one of the largest social media platforms, and so, that creates a lot of vulnerabilities. [00:35:35] S2: No, all of the... There are obviously parallels, right? Not to just keep calling out my company's name again and again. There was a real incentive to attack, whether it were crypto spams. Those were popular for a period, right? Or whether it was some state sponsored influence campaign, which we used to disclose a lot publicly. Those things always happened. It required real energy focus to control. It is, the thing most people don't understand is, it is almost impossible to take it to zero. You have to contain it because it is a knife fight and the tactics change. Anytime you control some vectors, another new vector will emerge and then you will control that. The other thing people don't understand is most of the bad stuff is stuff that doesn't show up on the surface. So if you saw an account and thought this is a bot, it's harmless because you could see that's bot. So the real problem is going to be in the margins when we lose our ability to differentiate. Yeah. Either as people or as the technical people building these systems, when you can't differentiate a human from a bot or an agent, that's when you're like, okay, now I don't know what I'm up against. And I think that is very similar today that we used to deal with back then. [00:37:21] S1: Do you think bots should address themselves?
00:37:21 · Molly O'Shea, Parag Agrawal
[00:37:21] S1: Do you think bots should address themselves? [00:37:24] S2: I have two contradictory feelings about this. I do think good bots should declare themselves, so that others can recognize them as good bots, and treat them differently from bots that might be adversarial. And we keep using the word bots, I would like to use the word agent. [00:37:46] S1: Okay, well, every time we use bots, I really think
00:37:48 · Parag Agrawal
[00:37:48] S2: use bot has a more negative connotation, while an agent is a more appropriate word. I think agents are perhaps the good bots. Mostly. Yeah. The... So, yeah, I think they should declare themselves. I know there's like a bunch of protocols now being done around declaring identity or the world at large. The core challenge there, why the other side is like, there's a conceptual question. Should your agent have the same access as you? Should it inherit your access or should it have only a part of your access? Now, would really, it would be so much better if we could figure out a way to get rid of the bad bots, but live in a world where we could enable our agents to have the access we have and grant that privilege. Today it's strange, right? Like I can't get my instinct or muse to have the access I have on many websites because they got blocked there and they will get even more blocked if they started declaring themselves. Yeah. And that is why I think we're in this moment of transition where these products have these rough edges, partly because we haven't done incentive alignment. But if we started just declaring them everywhere and they've started declaring themselves everywhere, like we don't participate in that space, but we started doing that. A bunch of people today, we haven't done the work to do business model alignment are going to start blocking this. Yeah. And so these products then will look worse and be less useful and that sucks. So to me, the existential problem is how do we drive alignment between agents and everything on the web.
00:39:44 · Molly O'Shea, Parag Agrawal
[00:39:44] S1: Do you think everyone will need alignment and safety on their team? Is that the next higher class? [00:39:49] S2: Yeah, I think how do you not deal with the problem? Like we, if you think about us, we are the web layer. The web is the dangerous place. Anyone can post anything malicious anywhere. Right? A lot of that content will flow into an agent's context. So you can do injection attacks. To the credit of model builders in the last two years, models have made very real strides on not being very vulnerable. But I think this whole world will rely on defense in many, many layers. So almost every layer has to obsess about this problem. You can't say, models will solve it. We have to really focus on it to solve it at our layer before we get into the models. Then models have to really focus on it. Then the app builders who are building with models and with us have to really focus on it at their layer. So we have to deal with adversaries at every layer and that is actually what gives you true defense. Still the occasional thing will leak out.