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
AI is an umbrella term for a loosely related set of technologies. You have, on the one hand, generative AI like chat GPT. On the other hand, you have self-driving cars and you have predictive AI, AI that's used in the criminal justice system, for instance, to make enormously consequential decisions about people, AI that's used in healthcare.
These types of AI generally have very little to do with each other. And it's true that some types of AI, notably generative AI, are rapidly advancing.
But we should be careful about, I think, the snake oil salesmen in the AI world who like to just slap the AI label on whatever tech product they're selling to try to get us to think that it is some remarkable technology that's going to solve all our problems for us.
Oh, not at all. And I would say to listeners that if someone tells you it's too complicated, you should be skeptical. They're probably trying to hide something. But in broad strokes, so let's take a couple of different types of AI. So what's happening in ChatGPT is that it's simply a machine, and some of you may have heard this, a machine for predicting the next word in a sequence of words.
What is the most likely next word? And it turns out, and this was largely a surprise to AI researchers as well, that the way for AI to be really good at predicting the next word in a sequence of words is to have some quote unquote understanding of language, of grammatical rules and patterns, and understanding of facts about the world.
Because if you have a sentence like the capital of France is blank, It helps to know what the capital of France is so that you can complete that sentence with a high probability word instead of a low probability word. So it turns out that's really the secret behind it. And it might seem a little bit disappointing to hear that that's all it is. And in a sense, that's all it is.
But I think it is truly remarkable that developers are able to create something useful with this really brute force approach.
Chatbots have been trained on essentially all of the text on the internet, approximately speaking, and a lot of books and so on. So that's what it's pulling from, right?
So what training means is that it has learned the statistical patterns that allow it to say, for example, something you should, the bot would know that no is a likely next word because there are many discussions of this podcast online. And so it has learned that statistical pattern. So that's primarily what it's learning from.
To a lesser extent, these bots learn from their conversations with us, but it's not in the way that a person would learn. It's not automatic. There's a cumbersome process by which companies have to filter these chat conversations and feed that back into the training data. But to a first approximation, it's learning from text on the web.
Sure. Yeah. So predictive AI, on the other hand, is statistics that we've had for a century almost that's been rebranded into quote unquote AI. So this is used, for example, in the criminal justice system to determine if a defendant should be jailed before their trial, you know, which could be months or years away.
It's used in healthcare to detect sepsis in a hospital context, for instance, by looking at various indicators. It's often used in hiring to try to predict which employee might be a fit for the role or who's going to be a good employee and that sort of thing. Now, what's happening in all of these cases is that the system is just picking up crude statistical patterns.
So in criminal justice, the system learns that younger defendants are more likely to re-offend if they're released before their trial, and so recommends treating them more harshly. So these kinds of, again, fairly crude statistical patterns that we've known how to do for a long time, but it's not the same kind of technology behind ChatGPT
It is not something that's advancing quickly, and it is something that I think we should be pretty skeptical about.
Exactly. It's making decisions about the future based on the past. So no matter how accurately it works... I think, you know, kind of on a fundamental philosophical level, we should think about, is this a just way to treat people, right? Should you deny someone their freedom in the criminal justice system because of the behavior of people like them in the past?
So that's, yeah, something that's deeply questionable as well.
Yeah, image generation AI has been advancing very quickly. And over the last year or so, companies have been working on video generation AI. So yes, I think to some extent, the hype around this is real. I think deepfakes are already a problem. Specifically, the thing I'm most concerned about is deepfake nudes.
And this has affected, from what I can tell, hundreds of thousands of people, primarily women, as you can imagine, around the world. And I think we desperately need regulation to curb some of the damage here. Now, in the political sphere, there's also concern that deepfakes can be used to trick voters and that sort of thing. I'm less convinced of that. There has been a lot of alarmism about that.
But I think something we should think about is that In a world where we're online and we have no easy way to tell what's real and what's not, what does that mean for the erosion of trust in the online environment? And how easy that makes it for powerful people, politicians and others to evade accountability by claiming that even real videos are actually deepfakes.
So we see that happening over and over. And that is something I'm worried about.
Showing 81–100 of 255 · page 5 of 13
← Previous
Next →