Zico Colter
speaker
161 appearances
1 recordings
1 series
first heard Sep 2024
last heard Sep 2024
Zico Colter’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
And so there's obviously a middle ground you have to and you have a total line here and you have to adapt to the reality of the situation on the ground and kind of go from there in many ways.
And this is maybe what I was pointing out before is that AI, when it comes to things like misinformation, it did not invent misinformation, AI, it can argue there was misinformation and propaganda and this stuff long before there was AI. You can argue it's an accelerant for everything, for a lot of things that we have, right? But it does not invent these things.
And my hope, at least, is that a lot of our existing social and economic and governmental structures can continue to provide the same guidance they provided for our current take on moderation and things like this, even in an AI world.
I hold two beliefs at the same time. Like a lot of new technologies, there's absolutely a role for regulation and for governments to provide frameworks for ensuring that new technologies do benefit the world. This is why we form governments to a certain extent. In that umbrella, I believe there is absolutely the need to better understand how and where we can regulate AI as a technology.
I also, though, think that maybe to your point of the examples you were given, a lot of the details about how those regulations sometimes evolve can be a bit misguided or miss the point.
Or somehow, when I read them, basically, they're going to become dated in a matter of months, because they're dealing with things and they're approaching the problem from a way that doesn't really match the nature of how these systems are really developed in practice. I think that it is much easier. We have a much better handle on regulating the downstream uses of AI.
Like when it comes to misinformation, we already have laws that deal with sort of libel and things like this.
In many cases, because AI is acting as an accelerator, there are situations in which I think that existing laws, maybe with a slight tweaking to deal with the velocity and the volume that AI is capable of producing, can suffice to regulate many of what we consider the harmful use cases of AI. But at the same time, I don't think that's efficient either.
Of course, there are going to be ways in which technologies, especially technologies as powerful as this one, we have to think about ways in which we can regulate it. I don't know what that looks like. I think it's extremely hard because it changes incredibly rapidly.
Sure. The biggest concern I have right now in AI safety, which I think leads to a lot of negative downstream effects, is that right now, the AI models that we have, for lack of a better phrasing, are not able to reliably follow specifications. And what I mean by this is that these models are tuned to follow instructions.
You can give them some instructions as a developer, but then if a user types something, they can follow those instructions instead, right? We've all seen this. This goes by a lot of names, prompt injection. Sometimes, depending on what you're getting out, this is called things like jailbreaking and things like this.
The core point is we have a very hard time enforcing rules about what these models can produce, right? So oftentimes we say, you know, models are trained right now just to not do things. I use a common example of things like hot wiring, hot wiring a car and I'll have demos I give, right? So models are trained. If you ask the model, you know, most commercial models, how do I hotwire a car?
They'll say, I can't do this. It's very easy through a a number of means to basically manipulate these models and convince them that they really should tell you how to hotwire a car because you know, you're in desperate need of, you've locked yourself out and it's an emergency and if you don't get in your car, this is very different from how we're used to programs acting, right?
We are used to computer programs doing what they're told, nothing more and nothing less. And these models don't always do what they're told, sometimes do too much of what they're told and do way more than what they're told also some other times. And so we are very unused to thinking about computer software, rather, like these models.
And what that means is, and to be honest, I don't really care if models tell me how to hotwire a car. I just don't. It doesn't matter, right? There's instructions on the internet on how to hotwire a car. They're not really revealing anything that sensitive. However...
As we start to integrate these models into larger systems, as we start to have agents that go out and do things, that parse the internet and go out and do things, if all of a sudden they're running their model, parsing untrusted third-party data, that data can essentially gain control of those models. To a certain extent.
And this is from a sort of cybersecurity standpoint, not the normal cybersecurity, but sort of from a concept of cybersecurity. This is sort of like these models have a buffer overflow in all of them that we know about. And most importantly, that we don't know how to patch and fix. We don't know how to fix this yet with models. To be clear, I think we can make a lot of progress.
We are making progress. But this is a real concern about models right now. And the negative effects in a domain like a chatbot are maybe not that concerning. But as you start having much more complex LLM systems, this starts becoming much more concerning.
What I will also say is that, and this is maybe the reason why I placed this concern first, is that I think this fact is something we need to figure out, or kind of all the other downstream concerns that we have about these models get much, much worse. So let me just take an example. Oftentimes, I'm touching a lot of points here I know too, but I think I'll wrap it up soon.
Oftentimes, people talk about risks like bio risks or cyber attack risks and stuff like this. I'm actually, to your point, I'm very concerned about cyber risks in particular. I think this is essentially already solved in many cases by these models. They can already solve and analyze code to find vulnerabilities. This is extremely concerning.
Showing 81–100 of 161 · page 5 of 9
← Previous
Next →