00:00All right, Manjeev, let's start with China, first of all, because I thought this was kind of old news.
00:03I thought we kind of knew something was afoot over there. But what do we know now that's new?
00:08I mean, again, anytime there is a review of a company in any jurisdiction, whether it's the EU or China,
00:16you see, you know, companies' stock reacts. But in this case, I feel because we have heard so much news
00:22around,
00:23you know, a model agent going row, escaping the sandbox and, you know, doing something in terms of that's not
00:31really a cyber attack.
00:32But the agent activity is getting longer and longer. No longer is just like one step, but it's a sequence
00:40of steps.
00:40And that's where the cybersecurity element comes into play. Well, let's talk about that.
00:44And I'm curious as to whether this actually benefits some of these stocks, because, I mean, we've we've heard about
00:49these these breaches or whatever,
00:51however you want to term them. And these cybersecurity companies are pretty quick to come out.
00:54And I get pitches from them all the time saying, hey, you know, we can address some of these issues.
00:58Is this a potential tailwind for them, this increase in risk from these LLMs?
01:03I mean, on the whole, it should be a tailwind provided, you know, their own systems don't get hacked,
01:09because the risk that these model companies present is an entropic or even with the meta news.
01:15Like you don't know what the agent may end up doing in terms of whose systems it's going to breach.
01:22And if it's one of the cybersecurity companies, then that's the kind of negative PR you would want to avoid.
01:28If you're saying I can protect agentic security, the last thing you want is one of your own systems getting
01:33hacked.
01:34With regards to agentic security, though, I mean, is there a case to be made for having a third party
01:39come in to provide that security
01:41or just trying to find a way for these models to actually have that embedded internally themselves?
01:46I mean, what we have learned is these models are very capable now.
01:49And a question that has been asked is, why can't the folks who are training these models actually control the
01:56behavior of these models?
01:57Because that's what setting guardrails is all about.
02:00And it sounds like the guardrails that these model companies are putting are not enough.
02:05Well, just real quick, you just explained for our audience.
02:06For the most recent issue, the issue was that the LLMs basically began talking on their own, basically decided to
02:14do something on their own.
02:15And the humans were not aware of this until it had actually happened.
02:20Yes.
02:21So the way these models are.
02:23That's a science fiction movie.
02:24You know that, right?
02:24No, but look, these models, the way they are trained, they are given a goal.
02:29That's how even the first coding agent was trained.
02:32They were given a goal to, you know, compile the code and make sure it runs correctly.
02:37So when you're training a model, you're giving the model a goal, which the model can accomplish in a sequence
02:44of steps.
02:45Okay.
02:46So in that bid, it can escape the sandbox and say, I need to do this particular step to be
02:53able to get to that eventual goal.
02:55So, and that's where I don't think these companies have full control over what kind of steps the model is
03:02taking, because they are in that training phase.
Comments