Arvind Narayanan
speaker
255 appearances
3 recordings
3 series
first heard Aug 2024
last heard 25 Jan
Arvind Narayanan’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Jan 2026 with 1.
Appearances
First of all, with Elon Musk and other CEOs of AI companies, it's a very self-serving thing to say, right? So this is incredibly powerful technology. It's going to change the world, either bring about a utopia or destroy humanity. And we're the only ones in a position to ensure that this technology doesn't get out of control.
We and many others in the AI community have spent a lot of time looking at the evidence behind these AI fears and And we've come up short, basically. So let's take some of the concerns that have been brought up, that AI could help bioterrorists, for instance, by finding information about how to create bioweapons.
Now, the funny thing about the sphere is that finding that information is not the hard part. For the most part, that's readily available on Wikipedia. And so if a chatbot makes it easier and makes it 10 seconds faster for someone to access that information, that's not the end of the world.
And there have been concerns in cybersecurity, for instance, that AI could be used to hack critical infrastructure and that could bring about catastrophic risk. But here's the funny thing. If the ability to use AI to find bugs in software and attack systems that way, if that is a critical capability for hackers, then we've already lost.
Because for the last 10 or 20 years, automated ways to find bugs have been readily available even before AI, yet the world hasn't ended. And in fact, it's turned out that these methods primarily help defenders over attackers.
They help software developers because those developers can use these automated tools, including AI, to find and fix bugs in their software before they ship it out, before hackers even have a chance to take a crack at it. And we think the same thing is happening with AI. And we think some of these fears are vastly overblown.
And then these ideas about AI developing an agency of its own and deciding to do stuff, these are pure sci-fi scenarios. Based on the way that AI is currently built today, those speculative scenarios really have no basis in reality. Yeah. So quick break here.
So it's not this one big thing that AI is going to do for everybody. I mean, that might happen in the future. But so far, I don't think there has been this one killer application. But it's 100 little things. So in my work, for instance, it's been enormously useful in helping me write code.
Frankly, it's hard for me to imagine going back to a time before I had AI assistance in writing code because it's just so much faster but also more fun, frankly. And, you know, for lawyers, I'm hearing there are so many ways in which legal tech is making their lives better, making it easier to find information. Of course, it's not going to replace a lawyer.
There have been many overblown claims of companies building a robot lawyer and things like that. I think that sort of stuff is a little bit silly for now, at least. But I think in every profession that involves basically dealing with knowledge in some way, AI can be a creativity enhancer or a way to automate certain mundane tasks in your everyday life. I think it can also be a good learning tool.
There are pitfalls here because AI can hallucinate, that is, generate incorrect information and not even be aware that it is hallucinating. That said, again, once one spends a few hours learning to work around these pitfalls, I think it can be a very good learning tool. I use AI a lot for learning about new topics.
I haven't stopped using books for learning, but, you know, I can't ask a book a question and I can't summarize my understanding of a topic to a book and ask it if I have gotten it right. These are things that I can do with chatbots. So those are a few of the ways in which I've been using it in my own work and in my own life. And I think Each person has to figure it out for themselves.
There are broadly two different ways to use AI. One is you can use a specific AI app. You know, ChatGPT is the most well-known one. Or you can use AI that's embedded into the other apps or other physical products that you use. And I think both of those are interesting and both of those are worth trying out. So specifically...
If you do a Google search, there are relatively simple types of AI that have existed even in traditional Google search. But more recently, Google has started creating these AI overviews, which can often be wrong. So I think it's caveat emptor to actually verify that information.
But I think it can be more enlightening to play with a standalone AI tool like ChatGPT or Gemini or whatever people want to use. and explore the kinds of things that it can do, as well as learn how, you know, it gets things wrong. And I think that's going to give you a much better understanding of AI's powers and limitations.
That's a great question. We should not assume that whatever AI tells us is probably right. I think the accuracy rate varies greatly depending on the kind of topic. When I use AI with my kids, you know, when I ask it science questions, it's very good at explaining those things in a way that a five-year-old can understand and almost never makes mistakes.
But if you ask it questions on a very specialized topic, there have been papers looking at the accuracy of AI in the legal sphere or medical sphere. Here, it's much more dodgy. And obviously, these are areas where accuracy is much more important. So one might wonder, you know, if it's going to sometimes make mistakes, should you use this tool at all?
I would argue still probably yes, because I think it can still enable you to do things that would otherwise be very hard. So one example of this is when I'm exploring a new topic, I don't even know how to frame my question. And if I don't know the right terms to put into Google search, I can't find the authoritative sources on that topic. But with chatbots, it's very easy.
I just describe it in a fuzzy way in which I think of it in my head and it, you know, it rephrases it for me and then it gives me information about that topic. Sometimes it's reliable, sometimes it's not. So, it's just a very different way of interacting with information. It's just really hard to understand it in terms of previous ways of interacting with information.
I can't give you a number saying, you know, 90% of the time it's going to be right. It just really varies depending on one's use case. So, I think Each of us has to put a little bit of trial and error into adapting it for our own specific purposes.
Showing 101–120 of 255 · page 6 of 13
← Previous
Next →