This Is How to Tell if Writing Was Made by AI

Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News. Speaker 2: Hello and welcome to another episode of The Odd Lads podcast. Speaker 3: I'm jille Wisenthal and I'm Tracy Alloway. Speaker 2: So, Tracy, you know, you ever come across some writing Speaker 2: you can't articulate exactly why, but you're like, I'm pretty Speaker 2: sure AI wrote this? Speaker 3: Does this happen too much? Speaker 4: So, full disclosure, I haven't really thought about it that much. Yeah, Speaker 4: because the thing is I probably should think about it more, Speaker 4: but there's a lot of bad writing out there, and Speaker 4: I've become sort of a nerd to it. And I Speaker 4: also think that I don't know trying to figure out Speaker 4: whether or not something was generated by AI nowadays, if Speaker 4: you actually dedicate a lot of your own time to Speaker 4: doing that, that is a huge mental burden to be attempting. Speaker 4: Especially you and I are in the journalism industry. How Speaker 4: many of the pitches do you think that we get Speaker 4: from prs right now are being generated by A I Speaker 4: imagine if you're reading each one of those and trying Speaker 4: to figure it out on a daily basis. Speaker 2: You know what I suppose I think about it the Speaker 2: most is someone will respond to a tweet yeah, and Speaker 2: I'll be like, well, if this is a real person, Speaker 2: then maybe this person deserves some engagement and ask a Speaker 2: question or I want to respond. But if there's a Speaker 2: person in the bot, then obviously I don't. And that's Speaker 2: where I look, you know what, I want to figure Speaker 2: it out. I would like to know the answer. Speaker 3: You know. Speaker 2: I have a controversial view about AI writing, by the way, Speaker 2: which is that it's pretty good. I mean, like, by Speaker 2: and large, and I said this, I think maybe in Speaker 2: a recent episode. When you consider the fact that I Speaker 2: don't know the majority of the population, like doesn't know Speaker 2: where to put a comma within the sentence, Well, this Speaker 2: is my point. Speaker 3: It's pretty good. Speaker 5: I mean, yeah. Speaker 2: One thing I'll say about AI is it never gets Speaker 2: the placement of a comma wrong. Speaker 3: On some level, it's perfect. Speaker 6: Did you do that? I think it was in the Speaker 6: New York Times the test. Speaker 3: I kind of hated that. Speaker 2: Okay, why well, because I'll tell you, first of all, Speaker 2: it's a five examples. Speaker 3: There's not very many. Two It asked the reader, which Speaker 3: do you prefer? Speaker 4: But I think they were different subjects as well. Speaker 3: Yeah. Speaker 2: Also, I think most people probably treated that as can Speaker 2: you guess which one is a human? Because everyone wants Speaker 2: to say they prefer the human I didn't think it Speaker 2: was like a great test. Nonetheless, Look, not only is Speaker 2: it often indistinguishable, not often is it often fine writing. Speaker 2: Sometimes AI could come up with a really remarkable turn Speaker 2: of phrase. Yeah, but I still buy and large don't Speaker 2: like it. You read like a thing, especially a long Speaker 2: text a's AI, and it's like, even if you can't articulate. Speaker 3: It, it's like this feels AI. Speaker 2: It has a certain sickliness sweetness to it that is Speaker 2: often annoying. Speaker 3: It's annoying. Speaker 4: What I notice about it is it doesn't do style Speaker 4: very well, right, So if you ask it to write Speaker 4: something in the style of a writer, if you choose Speaker 4: anything other than something really obvious like Shakespeare, it really Speaker 4: it suffers. But the text that it actually outputs is Speaker 4: pretty clear. Yeah, right, like for basic understanding. Total it's Speaker 4: probably better than a lotful what's on the internet. Speaker 2: The real people who are going to have to worry Speaker 2: about this are like teachers obviously, universities and lawyers, student Speaker 2: lawyers and maybe at it's fun, but there are sometimes Speaker 2: it's like, Okay, did someone write this or not? Speaker 3: And there has to be it'd be nice if we Speaker 3: could know the answer. Speaker 4: Well, the other thing that's starting to happen is have Speaker 4: you seen any books out there that actually come with Speaker 4: a disclosure or disclaimer that say this book has been Speaker 4: written only by humans? Speaker 5: No? Speaker 6: AI used at all. Speaker 4: I saw that for the first time on a book Speaker 4: that we actually read for an All Blots episode. I Speaker 4: don't think it's come out yet, but that kind of Speaker 4: threw me. Speaker 1: Yeah. Speaker 2: No, it's more and more anyway, as we enter a Speaker 2: world at which the vast majority, if not already of Speaker 2: words written are written by AI, is going to be Speaker 2: interested in this question of whether we know Anyway, there's Speaker 2: this company called Pangram Labs, and they have a little Speaker 2: thing and you can pay for it, but also a Speaker 2: free service where you can drop like a text in Speaker 2: and it'll say the odds that is written by human Speaker 2: or AI. And I'm pretty impressed by it. I like Speaker 2: did some samples of my own writing and then AI Speaker 2: outputs it got them all right, But then I did Speaker 2: some like further, like I tried to stump it to Speaker 2: see if like. So, what I did was I took Speaker 2: a piece of AI writing and then I had it Speaker 2: translated into Chinese, okay, and then I had it translate Speaker 2: that into High Chinese, so it's like, okay, imagine this Speaker 2: is being written by a more formal register. And then Speaker 2: I had that translated into Hebrew, and then I had Speaker 2: that translated into English. So the original thing through this Speaker 2: series of Ai telephone, through various translations, and then I Speaker 2: put that output back into Pangram. Speaker 3: I got that right. It said it was Ai. Speaker 2: So even after a series of sort of transformations designed Speaker 2: to obfuscate the original style of the piece to see Speaker 2: if you know, eventually it would emerge in something else. Speaker 2: So I was pretty impressed. It seems to work. And Speaker 2: you know, I think that's interesting for a couple of reasons, Speaker 2: which is maybe there is something that you can just tell. Speaker 2: But two, it sort of worries me because you know, Speaker 2: there have been articles and they'll say like, this is Speaker 2: written by Ai, And I think one of my big Speaker 2: fears would be that I write something. Speaker 3: I like to use an mdash. Speaker 4: I've always been in them, dash fan, I love m dashes. Speaker 4: That's how people talk. Speaker 6: I'm sorry. Speaker 2: And then what if it says you wrote this by Ai, Speaker 2: and I'm like, I didn't, And then here's this black Speaker 2: box that is suddenly like Judge Jurgen, executioner for my Speaker 2: career potentially who wrote this. AI the Lab says, so Speaker 2: you are now done? Like that worries me. So I Speaker 2: think this raises a lot of very interesting questions about Speaker 2: these molde little detection things, and I want to learn Speaker 2: more about how well. Speaker 4: There's also a lot of philosophical questions about just what Speaker 4: we value in writing true as well, because no one's Speaker 4: going to yell at you for using spell check or Speaker 4: something like that, right, Like, it's kind of crazy to Speaker 4: think that reputational risk is going to hinge on whether Speaker 4: or not you might have used a platform, a chat Speaker 4: platform to like do some basic copy editing. Speaker 2: Totally well, very happy to say, we do, in fact Speaker 2: have the perfect guest. Speaker 3: We're going to be speaking with Max Spiro. Speaker 2: He is the founder and CEO of Pangram Labs, and Speaker 2: he can answer all of our questions. So Max, thank Speaker 2: you so much for coming on. Speaker 5: Outlaws, Thanks for having me. Speaker 3: How do you know it's right? Speaker 2: So someone puts in a piece of tech and we'll Speaker 2: get into the method in the second. But someone puts Speaker 2: in a piece of text and it says human AI, Speaker 2: what makes you believe that you have a very good. Speaker 3: Track record all this question. Speaker 7: So when we started Pangram, we started by doing this Speaker 7: thing we call a human baseline, which is how well Speaker 7: can we as a human predict whether something's AI or not? Speaker 7: That's the first step out like learning, is this problem tractable? Speaker 5: How hard or easy is it? And I found, like. Speaker 7: Me personally, I was able to get about ninety percent accuracy, Speaker 7: and so we figured an AI model should be able Speaker 7: to do much. Speaker 5: Better than that. Speaker 4: So I have a bunch of methodology questions which we Speaker 4: can get into. But just before we get into any Speaker 4: of that, why is AI slot bad in your opinion? Speaker 4: Why does it need to be tracked and identified? Speaker 7: I think the problem is is just so easy to Speaker 7: generate and so like it's very difficult to know, like Speaker 7: what is the like intent behind it? Basically, Like right now, Speaker 7: I think we're actually pretty lucky living. We live in Speaker 7: a world where the signs noise ratio on the Internet Speaker 7: and in our information. Speaker 5: Channels is pretty high. Speaker 7: We have pretty high signal to noise, But any bad Speaker 7: actor can come in and just flood our information channels Speaker 7: with aislot that looks legitimate. It looks like somebody put Speaker 7: actual effort and thought into it, but really it was Speaker 7: just like a single prompt which could have also been automated. Speaker 2: This is something that I think about a lot, which Speaker 2: is that there was a point in time and maybe Speaker 2: still is the point in time where if you read Speaker 2: something that was grammatically correct, where the punctuation was strong, Speaker 2: where the spelling was strong, there was reason to think Speaker 2: that the person who wrote it was a person of Speaker 2: like certain seriousness and a certain intelligence behind it. Speaker 3: And I think that the issue that you're. Speaker 2: Identifying is that that link is now being severed so Speaker 2: that we can't use these heuristics anymore, such as the Speaker 2: strict quality of the pros to know in fact whether Speaker 2: this was published by someone who was like a serious actor, Speaker 2: intelligent or or not. Speaker 4: And now you have people inserting typos into their card Speaker 4: that's true that they are Yeah boyd. Speaker 2: Sorry just to go back to my original question. So Speaker 2: you mentioned, okay, you're able to get it ninety percent right, Speaker 2: but now we've been used a lot more and you Speaker 2: have people paying for your software, presumably teachers and journalists, etc. Speaker 2: Given all of that, getting from ninety percent to one hundred, Speaker 2: I mean, if you could make one out of ten Speaker 2: it's clearly an unacceptable error raid for a piece of Speaker 2: commercial software that could call someone an AI creator. So Speaker 2: you have to do a lot better than ninety percent. Speaker 2: Talk to us about like what you've seen so far Speaker 2: in your data since releasing it as commercial software that Speaker 2: makes you believe the software is doing a correct job Speaker 2: of allocating between the two categories. Speaker 7: So we've built out really comprehensive emails, okay, and so Speaker 7: our evaluations. There's two kinds of errors. There's a false positive, Speaker 7: which is when something is written by a human and Speaker 7: then we say that it's written by an AI, okay. Speaker 7: And there's a false negative, which is if it was Speaker 7: AI written and we don't catch it. And so we Speaker 7: track our numbers for both of these, and for human. Speaker 5: Writing, we're actually pretty fortunate. Speaker 7: We have like millions and millions of samples, so we Speaker 7: can get like a false positive number that we have Speaker 7: a very high degree of confidence in. And our number Speaker 7: right now is about one in ten thousand. Ok So, Speaker 7: if we scan ten thousand documents on average, one will Speaker 7: come back as. Speaker 5: AI when it was actually human. Speaker 3: And what about in the other direction false negative? Speaker 7: I would say around ninety nine percent accuracy, So like Speaker 7: around one percent false negative rate. I think this depends Speaker 7: a little bit more on like how adversarial the prompting is, Speaker 7: how much they're trying to ev. Speaker 2: What I did exact send it through multiple filtrations to Speaker 2: obfuscate the original output. That would be an example of Speaker 2: adversarial prompting exactly. Speaker 7: But in like the general case where we're just looking Speaker 7: at straight outputs from AI, it's above ninety nine percent. Speaker 4: Okay, okay, So what is your model looking for exactly Speaker 4: when it's evaluated a text? Because, as we mentioned in Speaker 4: the intro, you know, syntax and grammar tends to be Speaker 4: pretty good on AI generated copy. The style is sometimes Speaker 4: more of an identifier, I would argue to your point, Joe, like, Speaker 4: sometimes it reads very saccharine and kind of overly earnest Speaker 4: in some ways. So what exactly are you focusing on here? Speaker 4: What are the tells? Speaker 7: Yeah, so the style and the word choices are definitely Speaker 7: part of it. But I think what a lot of Speaker 7: people don't realize is they're actually making a lot of Speaker 7: decisions when they write a piece of text. So there's Speaker 7: you know, dozens or hundreds of ways to phrase every Speaker 7: single phrase, and over the course of fifty or one Speaker 7: hundred or two hundred words, you're making thousands of decisions actually, Speaker 7: And so what we're doing is we're learning the patterns Speaker 7: and how like these frontier models make these decisions. And Speaker 7: if the vast majority of these decisions line up with Speaker 7: how the frontier models are doing it, then it's vanishingly Speaker 7: unlikely that this was written by a human. You would Speaker 7: have to just happen to make the same exact decisions Speaker 7: that the LM does hundreds of times. Speaker 6: Interesting, Okay, this. Speaker 3: Is a really important point. Speaker 2: So everyone at this point has some feel for let Speaker 2: go the M dash tell right, But my understanding is Speaker 2: it's not like you don't go in in like hard Speaker 2: code if you see a bunch of M dashes. This Speaker 2: is the thing these decisions. In many cases, I imagine, Speaker 2: neither you nor the model itself can articulate in English Speaker 2: what the decisions are. All you know is that the Speaker 2: decision pattern exists. Speaker 3: Is this correct? Speaker 5: This is correct? Speaker 3: Okay? Can you explain? Speaker 2: So therefore, what does it mean that your model has Speaker 2: learned these decision? Speaker 7: So what we're doing on the very broad scale is Speaker 7: we're training a deep learning model. So it's a pretty Speaker 7: big black box, but it has the base model of Speaker 7: a language model, and then instead of predicting the next token, Speaker 7: it's predicting whether it the text is AI or not. Okay, Speaker 7: And how we train it is we train on tens Speaker 7: of millions of examples, so it sees millions and milli Speaker 7: of human examples, and for each human example, we also Speaker 7: show it an AI example. So, for example, let's say Speaker 7: one of these is a five star review for Denny's Speaker 7: that's seventy eight words long. Then we'll ask in AI Speaker 7: to write a five star review about Denny's that's seventy Speaker 7: eight words long in the style of the first one. Speaker 7: And obviously these two will be different, and so our Speaker 7: model is able to learn through contrast, what is the Speaker 7: difference between. Speaker 2: Me and The Important thing, sorry, just to be clear here, Speaker 2: is that you and I might not be able to Speaker 2: articulate the difference. There will be some difference in maybe Speaker 2: the sentenced length, there will be some difference in word choice, Speaker 2: there'll be some difference in punctuation, syntax, whatever, but you Speaker 2: and I wouldn't obviously spot it. However, after millions of Speaker 2: examples of these side by sides, the model learns what Speaker 2: the difference is exactly. Speaker 7: I think the best that a human can do is Speaker 7: look for some of these like really obvious tells like chat. Speaker 7: GIPT loves that, like it's not just X, it's y framing. Speaker 7: Earlier models really liked some specific words like tapestry and Speaker 7: intercate and delve. Speaker 3: Yeah, delve tapestry. Yeah. Speaker 5: But yeah. Speaker 7: I think by training Pangram, we're able to go much Speaker 7: deeper than this and look deeper than the high level Speaker 7: science at the like document level science. Speaker 4: So one thing this kind of reminds me of and Speaker 4: I'm thinking how to phrase this, but it reminds me Speaker 4: of you know those exercises people used to do where Speaker 4: you would take a bunch of different faces and meld Speaker 4: them all together and come up with like one face Speaker 4: that was supposedly attractive. So, like, to what extent is Speaker 4: this basically a distributional detector in the sense that you're Speaker 4: looking for like certain paths that you think AI would choose. Speaker 4: And I guess, like, could you get a false positive Speaker 4: just from someone who's choosing like the average of the Speaker 4: average of the average in a way to state a Speaker 4: particular sentence. Speaker 7: Maybe there's a reason we have our false posit rate Speaker 7: is one in ten thousand and not zero. It's because Speaker 7: you know, sometimes we look at the false positive and Speaker 7: it's like, oh, it reads exactly like an AI generated Speaker 7: review or essay, except that it was written in twenty nineteen. Speaker 7: So it was probably a human who just happened to Speaker 7: find the exact like mode collapsed. Speaker 5: Type of way that like, yeah, thats right, Yeah, I Speaker 5: would say, yeah. Speaker 7: I think it's a good way to think about the Speaker 7: distribution of writing or writing as a distribution where like, Speaker 7: you know, there's the space of all human writing, and Speaker 7: then AI writing is really just. Speaker 5: Like a small point within this space. Speaker 7: It's very no matter how much you prompt it, it Speaker 7: doesn't go that far from where it was trained to be. Speaker 3: Yeah, okay, WA's the black book. Speaker 2: So I built a little model myself. I built this Speaker 2: thing that detext. You can upload text and says whether Speaker 2: it's more resemblant of the written word or the spoken word. Speaker 2: Oh I saw that, yeah, yeah, And I used bert, Speaker 2: which is like one of these things open source one Speaker 2: from Google. Speaker 3: What is the core model that. Speaker 2: You trained on or is it something or did you Speaker 2: build it yourself? Speaker 3: Like, talk to us about that. Speaker 7: Our very first model was actually built on Burt, but Speaker 7: future models we needed to up our capacity. So basically Speaker 7: we were running into capacity limits with our model. It Speaker 7: was capping out at a certain false positive false negative rate. Speaker 7: It wasn't learning the deeper signals, so we had to Speaker 7: ten x and then one hundred x the parameter account Speaker 7: so that can learn like really deeply, like how these Speaker 7: frontier models. Speaker 4: Right, Have you noticed any interesting differences between how the Speaker 4: models right? Can you and actually is your model trained Speaker 4: to identify different models as well as whether or not Speaker 4: This is just broadly AI generated. Speaker 7: So we don't specifically train it on different models. We Speaker 7: don't say like hey, this one is CLAT three and Speaker 7: this one is Chat or GPD five. What we've done Speaker 7: we've done some interpretability work to look at basically the Speaker 7: output embeddings of the model and where we find that Speaker 7: it actually learns which model the text came from. So Speaker 7: you could see like little clusters like this is the Speaker 7: Clod cluster and like all the clods, yeah, cluster around here, Speaker 7: and then these are like the deep Seek and Quinn Speaker 7: and then this is like Chat schipt and they all Speaker 7: kind of like cluster into different spaces and embedding space. Speaker 7: So clearly the model is able to learn what the Speaker 7: difference is between these frontier models. Speaker 4: We actually since you mentioned Quin, I'm very interested is Speaker 4: there anything like distinct in terms of how Quen generates Speaker 4: text versus platforms that have been developed in the US. Speaker 7: I think Quen is unique because it's trained on a Speaker 7: lot more Chinese and multi lingual tokens than other models. Speaker 7: So you know, I've heard from Chinese friends that it's Speaker 7: it's much better at like being conversationally fluent in Chinese. Speaker 5: Beyond that, I don't know that I can tell. Speaker 7: It would be hard for me to look at a Speaker 7: text and say, like, I know that's Quen, But I Speaker 7: think somebody who's more familiar with it might be able to. Speaker 2: Let's talk about sort of some of the philosophical or Speaker 2: societal implications of this work. Speaker 3: Have you had. Speaker 2: Anyone whose text has been judged to be ai written Speaker 2: by Pangram and they're like, I swear to God, this Speaker 2: isn't you're in? They like, really insist, and what do Speaker 2: you think about this situation? What do you do or Speaker 2: talk choice about that. Speaker 7: I've had a couple of times this happened. There have Speaker 7: been times where I genuinely believe that you know this Speaker 7: is just a false positive. We scan hundreds of millions Speaker 7: of documents, so like, at a certain scale like this Speaker 7: will happen. But I also get people who all the Speaker 7: time they're just like AI detectors don't work. Speaker 5: It's like a total fraud. Speaker 7: And then whatever they're putting out on LinkedIn is just Speaker 7: one hundred percent AI generated. Speaker 5: And they're just like mad that they're getting called out. Speaker 7: And then you look back farther into their past and Speaker 7: their history, like everything they're putting out is AI generated Speaker 7: until about like twenty twenty three, Like for everyone, if Speaker 7: you look historically, there's a lot of like slop accounts Speaker 7: that are putting out total slop, and you can tell Speaker 7: either they like weren't posting as much before, and if Speaker 7: you scan back in time, then you see that they Speaker 7: were writing human text at some point. Speaker 2: So there's a number of accounts out there that basically Speaker 2: right around the beginning of twenty twenty three, where if Speaker 2: you scan the entire corpus of their work, it very Speaker 2: clearly shows a switch. Speaker 3: Right around early twenty twenty three. Speaker 7: Yeah, it really like depends on the account. I think Speaker 7: one thing we saw that was interesting was there is Speaker 7: a writer for The Guardian that was covering the Winter Olympics, Speaker 7: and somebody was like, hey, this article is like total Speaker 7: AI slop. Ran it through pangram it was AI. The Speaker 7: Guardian was like, no, of course, our writers don't use AI. Speaker 7: And then we so we scanned this single writer's history Speaker 7: and we found that they really did start picking up Speaker 7: AI like mid to late twenty twenty four, and we're Speaker 7: using it more and more in their articles. Speaker 4: I mean, just play Devil's Advocate for a second. Does Speaker 4: intent matter when it comes to identifying AI slop in Speaker 4: the sense that, Okay, I get you can have a Speaker 4: bad actor who's maybe trying to influence how people feel Speaker 4: about a particular topic, and maybe they've created a bunch Speaker 4: of bots on Twitter slash x and they're using AI Speaker 4: to just flood the zone with a bunch of AI Speaker 4: slop supporting their particular viewpoints. On the other hand, if Speaker 4: you're a journalist and your business is to write, you know, Speaker 4: like basic understandable copy about a news topic. Speaker 6: Just to be clear, I'm. Speaker 4: Not advocating this at all, but that intent is very Speaker 4: different to I'm going to try to influence something by Speaker 4: just you know, sheer volume. Speaker 7: Yeah, I mean, definitely these are like one is a Speaker 7: lot more severe than the other. But I think at Speaker 7: the same time, if you're a journalist and you're using Speaker 7: AI to basically shirk your work and like not do Speaker 7: your work, I think that's also a problem. And I Speaker 7: think it's a reputational risk to the outlet because people Speaker 7: can tell and people are going to call you out. Speaker 7: There's a lot of people who don't want to read Speaker 7: AI slop kind of regardless of where it's from. Speaker 2: Yeah, this is a definitely true. Are you ever going Speaker 2: to run out of human material to change on? Speaker 5: Right? Speaker 2: Like you could be pretty confident that if you find Speaker 2: some piece of text that was published on the internet Speaker 2: prior to twenty twenty three, but certainly prior to like Speaker 2: twenty nineteen or something like that, you can be extremely Speaker 2: sure that this is human generated. Do you worry that Speaker 2: in the future, like it's going to be harder to Speaker 2: even establish the provenance of your training data. Speaker 5: Uh, Yeah, it's definitely a concern for us. Speaker 3: Talk to us about how to think about this. Speaker 7: So we have a near infinite data reservoir of pre Speaker 7: twenty twenty three data, there's just like more than enough Speaker 7: for us to train on for a long long time. Speaker 7: But part of the problem is we also want to Speaker 7: train on modern text. We want to there's all this Speaker 7: talk about like if somebody's writing about LMS or about AI, Speaker 7: we don't want to incorrectly flag that as AI because Speaker 7: our training data has no sense of this topic. So Speaker 7: I think we're looking at different ways to do this, Speaker 7: but most of them are just like figuring out like Speaker 7: who is a trusted actor? Speaker 5: Who do we know is. Speaker 7: Putting out humor written content and we could use our Speaker 7: model for that, like to some degree. And then so Speaker 7: we have known actors, we know they're putting out human Speaker 7: written content, and then we could use their as well. Speaker 4: Slightly random question, but using your model, are you able Speaker 4: to quantify like what percentage of the Internet at the Speaker 4: moment is aislot? Speaker 2: It's about forty percent based on why you're just how'd Speaker 2: you get that number? Speaker 7: So a lot of the Internet is just like SEO Speaker 7: written articles and like, yeah, it's articles written for search Speaker 7: basically so that your website comes up more often in Speaker 7: search because it's targeting certain keywords. And a lot of Speaker 7: that industry has switched over to using AI because then Speaker 7: instead of having to pay writers you could turn out Speaker 7: articles for pennies on the dollar, but I think that Speaker 7: kind of results in a lot of the Internet being Speaker 7: AI written. It's a little bit is also kind of Speaker 7: platform dependent. It's about forty percent from like a Internet Speaker 7: page perspective. About a year and a half ago, we Speaker 7: looked at Medium and found that over fifty percent of Speaker 7: newly written Medium articles were generated, which was a crazy Speaker 7: high number. Speaker 3: What about Reddit? Speaker 7: Reddit, it was seven percent a year ago, I believe Speaker 7: a little over ten percent. Speaker 4: Well, actually this reminds me. So I'm on Reddit a Speaker 4: lot and I really enjoy it nowadays as a platform, Speaker 4: but I do worry about how much of it is Speaker 4: being generated by AI. And the thing I don't necessarily Speaker 4: understand is what are the economic incentives to actually write Speaker 4: a bunch of AI generated posts on Reddit and get Speaker 4: up voted, Like why does that system or motivation even exist. Speaker 7: So there are startups I'm not going to name names Speaker 7: because I don't want to promote them, but they will Speaker 7: sell a promise to companies that we're going to get Speaker 7: you organic mentions on Reddit. We're going to run our Speaker 7: AI bots that seem organic, and they're just going to, Speaker 7: you know, naturally recommend your product or you know, just Speaker 7: mention your product in the comments or in a post. Speaker 7: And so I've seen evidence of this. We can find Speaker 7: these like they're basically like botforms that are mostly engaging, Speaker 7: seemingly organically, just like doing a short reply, and then Speaker 7: sometimes they're doing this brand mention. And so that's why Speaker 7: these posts are very valuable. Speaker 6: That's really interesting. Speaker 2: I have to you also imagine it's valuable because all Speaker 2: of the models train on Reddit, right, and if you Speaker 2: want your product's name to appear in model outputs, it's like, Speaker 2: what is the best you know, nose hair trimmer or whatever, Speaker 2: And there's a bunch of bots that on Reddit talked Speaker 2: about this nose hair trimmer, and then that's probably more. Speaker 3: Likely to show up in a chatchypt request, right. Speaker 7: Yeah, yeah, it's been weirdly gamed. You know, you used Speaker 7: to just google best nose hair trimmer, and now there's Speaker 7: like a thousand. Speaker 4: The Reddit search results like show up first nowadays. Speaker 6: Yeah, that's where people are looking. Speaker 7: Yeah, and then people start searching best nose trimmer Reddit Speaker 7: to get their Reddit comments on it. And now it's Speaker 7: people have realized that that's what people are searching for. Speaker 7: So you need to populate Reddit with your advertisements. Speaker 4: I'm on the Men's health Are you looking for nose Speaker 4: hair trimmers? Speaker 2: The Panasonic ear and nose hair trimmer is the number Speaker 2: one choice on men's health pros. Easy to hold anyway, Speaker 2: it's not. Speaker 5: Yeah, it's all these affiliate links. Yeah, just destroyed the Internet. Speaker 2: I know it's it's too bad, but whatever, talk to Speaker 2: us more about the whole pipeline. So, I'm very fascinated Speaker 2: by this idea. It's like, Okay, you see this review Speaker 2: for Denny's. You have the AI model. Speaker 3: Try to replicate it as best as it could. Movie Speaker 3: these subtle differences. Talk to us as though about, like Speaker 3: the whole pipeline. Speaker 2: What are the other tests that you're using to get Speaker 2: the true you know, because what I imagine you're trying to Speaker 2: do is get the most similar data sets with an Speaker 2: almost imperceptible difference to really stress tests. Yeah, talk to Speaker 2: us really about the whole pipeline. Speaker 4: Yeah. Speaker 7: So what we're really trying to do here is we're as. Speaker 3: A model maker myself, no, no, sorry, keep going. Speaker 5: Yeah, as an AI expert, Yeah, yeah. Speaker 3: As an AI expert. I need to hear some tips Speaker 3: of the field. Speaker 7: Uh yeah, So what we're really looking for is examples Speaker 7: that are as close to the boundary between human and Speaker 7: AI as possible that our model learns better. Something that's Speaker 7: very obviously AI is, you know, our models not learning Speaker 7: as much same thing for something that's obviously human. And Speaker 7: so step one is creating this data set with synthetic Speaker 7: mirrors of human examples, and then we train a model, Speaker 7: and then step two is something called active learning. So Speaker 7: we then take this model and use it to scan Speaker 7: a much larger corpus of data and look for errors, Speaker 7: false positives, false negatives, and then we pull those back Speaker 7: into our training set and are able to train a Speaker 7: much better model because it's seen these errors, which and Speaker 7: these errors we believe are just much closer to the Speaker 7: boundary between human and AI. Speaker 2: So sorry, just to be clear, the first pass is like, okay, Speaker 2: you have known human writing and known AI writing. You Speaker 2: train a model, and then the next pass is once Speaker 2: again unknown human and known AI writing. So you already Speaker 2: know the answer of each of these and therefore you Speaker 2: could come up with a list of which it got wrong, Speaker 2: and then that gets fed back into the first. Speaker 7: Verse exactly, and so that makes once we retrain, then Speaker 7: the model gets much much better, and then we could Speaker 7: do this as many times as we want to, kind Speaker 7: of just have a self improving model that gets better Speaker 7: with every training run. I can also tell you go Speaker 7: a little bit more into how we deal with AI edits, Speaker 7: because I think that's increasingly important. Problem is, like I Speaker 7: think most writing will be AI assisted in the future. Speaker 7: I think it's already in Google Docs and it's in Speaker 7: Google Keyboard. Speaker 4: Grammarly arguably has been doing this for a while. Speaker 5: Exactly. Speaker 7: Yeah, Grammarly uses LMS on the back end, and we Speaker 7: don't want to just say, like, all writing is AI now. Speaker 7: We want to be able to differentiate between AI assisted Speaker 7: and AI generated. So what we do is we also Speaker 7: have different prompts. So rather than saying so for our Speaker 7: human review of Denny's, rather than saying, generate a review Speaker 7: like this, we could say, help improve this, make it Speaker 7: more formal, make it more like, clean up the grammar. Speaker 7: And so we have like a long list of AI Speaker 7: editing prompts, and then we're able to look at basically Speaker 7: the cosine difference the distance between the original human text and. Speaker 3: The in that hyper multidimensional space. Speaker 7: Exactly, So how much did AI change this text? And Speaker 7: then we're able to train our model to say, like Speaker 7: we're just going to like put a point on this Speaker 7: distance and say like this is moderate aissistance, this is Speaker 7: light AI assistance, and this is heavy aissistance. Speaker 4: Interesting. I'm going to do something I don't think I've Speaker 4: ever done before, which is ask a founder about their Speaker 4: corporate mission. But you know, you've set up this company, Speaker 4: and when you think about what you're trying to do here, Speaker 4: is it just basic AI detection in the sense that Speaker 4: there might be you know, a few groups of people Speaker 4: like teachers that find this very valuable, or is the Speaker 4: mission something broader where you're actually trying to improve the Speaker 4: Internet and what people see on it. Speaker 7: I believe the technology of being able to detect AI Speaker 7: generated content is immensely valuable, and it's valuable not just Speaker 7: for teachers, but for basically everybody in every profession. Lawyer's Speaker 7: publisher is just an individual who consumes content on the Internet. Speaker 7: I think it's valuable for all these people. But ultimately, yeah, Speaker 7: our high level goal is to help mitigate some of Speaker 7: these negative effects of growing AI content. Speaker 4: But for instance, just using the product review example, is Speaker 4: the vision that like a Yelp, for instance, would want Speaker 4: to use this technology to make sure that its system Speaker 4: isn't being gamed or is the vision Like if I Speaker 4: am a particularly diligent consumer who has a lot of Speaker 4: time on my hands and I'm looking to go out Speaker 4: to a restaurant, I can run all these individual restaurant Speaker 4: reviews through Pangram and then like actually figure out if Speaker 4: it's real hype or not. Speaker 7: So I think right now it's a lot of the former. Speaker 7: We work with platforms. One of our biggest customers is Quorra, Speaker 7: and they run a bunch of content through Pangram. But Speaker 7: we have a lot of different platforms that use Pangram Speaker 7: to help moderate and find AI bad actors and get Speaker 7: them off their platform. But I also think, yeah, the Speaker 7: individual consumer case has been growing a lot, and we're Speaker 7: really interested in pushing. Speaker 2: Here the free version of pangram dot com. Like you Speaker 2: get a handful of tests a day or something like that. Speaker 2: If someone had an unlimited number of Pangram responses and Speaker 2: maybe had an excess to the Pangram api at infinite scale, Speaker 2: could they theoretically learn a prompt that they would then Speaker 2: be able to put into an AI to generate human style. Speaker 7: Writer actually had a friend do that. He put his Speaker 7: cloud code on a loop. I gave him some API credits, Speaker 7: and then his cloud code just basically worked overnight writing Speaker 7: a prompt trying to get it to put something that's Speaker 7: human written or that which came back there from Pangram Speaker 7: as human written. They got there, but the text was Speaker 7: pretty like uh incoherent, so so like, yeah, it was Speaker 7: producing more or less long gibberish. It was like grammatically incorrect. Speaker 7: A lot of the words just didn't really make sense. Speaker 2: Because this was my first thought, like when I saw it, Speaker 2: I was like, that would be like a fun experiment Speaker 2: to see if you could take all the outputs, find Speaker 2: the difference and just keep iterating on the prompt you Speaker 2: would have to tell AI in order to eventually get Speaker 2: an output that looked to Pangram like it was human generated. Speaker 7: Yeah, I think there's a way to do it if Speaker 7: you also had like an LM judge on coherency and Speaker 7: he's like Pangram and the coherency judge both to score Speaker 7: your text. I think it's definitely possible, and I'm excited Speaker 7: for someone to try to do it, because we could Speaker 7: make our model a lot better and more robust if Speaker 7: this existed. Speaker 4: So I want to know what your personal like token Speaker 4: budget is nowadays that you're even like contemplating some of Speaker 4: those stuff. Speaker 2: What I feel like I had the Cloude Max playing, Speaker 2: you know, and I don't work like when I'm at work, Speaker 2: I don't work on any of my Vibe coding projects. Speaker 3: And you know, like when we were kids. Speaker 2: I don't know if you remember, like if you didn't Speaker 2: need all your food, like someone to say, oh, there's Speaker 2: like starving kids in the world. Speaker 4: Yeah, I'm like, oh, it's starving Vibe coder. Speaker 3: It's like, oh, you didn't. Speaker 2: Like I have this four hour token window and I'm Speaker 2: almost never maxing it out, and I'm just thinking, like, Speaker 2: the are kids on the other side of the world Speaker 2: that wish they had your tokens and you're you're not Speaker 2: using all of your tokens for the window. Speaker 3: How dare you? Speaker 2: I feel a little guilty when I don't out max Speaker 2: out by Claude max token program. Speaker 7: I also have Claude Max and yeah, most days I'm Speaker 7: not doing much coding at all, I'm not maxing it out, Speaker 7: and then some days I'm going you feel a lot. Speaker 2: Guilty about that though, it's like, yeah, yeah, so can Speaker 2: I just feel like writing is kind of interesting, but like, Speaker 2: what are the prospects of this being able to work on? Say, Speaker 2: and you must get this lot image and video generation? Speaker 2: Is it it all theoretically similar? Is there a reason Speaker 2: to think that it will be replicable? Or is this Speaker 2: just a different beast of a problem. Speaker 7: I think the approach is definitely doable. I think some Speaker 7: of the economics change, especially if we look at video Speaker 7: and the cost of generating video today. Okay, we can't Speaker 7: generate video at the same scale that we can generate text, Speaker 7: and so we might need a kind of different approach. Speaker 7: But I also believe that if we're able to solve Speaker 7: this for image plus maybe like audio, that could be Speaker 7: enough to just solve it for video as well. Speaker 5: Huh, zero shot. Speaker 4: Could you ever envision, I don't know, launching some sort Speaker 4: of like certification program for video because this seems to Speaker 4: be my dad's a boomer spends a lot of time Speaker 4: on Facebook, Like this seems to be what society needs, right, Speaker 4: Like a video that comes with a little thing that Speaker 4: says this is not AI generated and someone has actually Speaker 4: like rubber stamped that, so. Speaker 7: There's an organization called c TWOPA, and I think they're Speaker 7: doing pretty good work on content provenance. Basically, they are Speaker 7: working with phone makers and hardware makers to basically embed Speaker 7: like hardware signatures to prove that image and video we're Speaker 7: truly taken from. Speaker 4: The hardware like watermarks basically. Speaker 7: Yeah, exactly so, So rather than marking the AI outputs, yeah, Speaker 7: we're instead embedding like a proof of authenticity in the Speaker 7: the like thing that's real and is captured. Speaker 5: In real life. Speaker 3: That's interesting, all right, So big picture, where's the Internet going? Speaker 2: You know, you mentioned forty percent of the Internet is Speaker 2: already air generated, but maybe that's something end of the world, Like, Speaker 2: you know, if it's just a bunch of SEO pages Speaker 2: that I never read, I don't know whatever, But like Speaker 2: give us some thoughts high level about like with the Speaker 2: trajectory of the Internet. Regardless of the uptake of Pangram Speaker 2: and other AD detection models. Speaker 5: I'm a little bit worried about the state of the Internet. Speaker 5: I'm gonna be honest. Speaker 7: I think like right now, there's still like so much Speaker 7: of it is built around trust and norms in a Speaker 7: way that like we're we're not really well equipped to Speaker 7: suddenly deal with an onslaught of bots at a completely Speaker 7: different scale than we've dealt with before. Speaker 5: There's maybe like a good case and a bad case. Speaker 7: I would say, like the bad case is the Internet Speaker 7: goes the way of debt internet theory, just like every Speaker 7: space that's open and accessible is just flooded by bots, Speaker 7: and then the only place people are able to communicate Speaker 7: authentically is in like very walled garden like closed servers Speaker 7: like like discord service for example, where you know everybody's Speaker 7: identity is known and you know you don't. Speaker 5: Have bots in here. So that's maybe the like bad scenario. Speaker 2: Can I do an insane thought that I've had go on, Speaker 2: We're gonna kick out of this? So when like I Speaker 2: forget what they call like this idea of like for Speaker 2: the bad actors, it's. Speaker 3: Called like heaven mode or heaven banning. Have you heard Speaker 3: of this? So there's this thought that one way. Speaker 2: You could deal with bad actors on the Internet is Speaker 2: suddenly they're on a version of say Twitter, in which Speaker 2: they're only bots and everyone always agrees with them on Speaker 2: everything and it drives them crazy and stuff like that, Speaker 2: and they would never know it because they're like, oh, Speaker 2: there's call, everyone's there, and then it's so like slowly Speaker 2: like yeah, they just this is like a way you Speaker 2: could punish people by putting them on an internet where Speaker 2: they will never get any fight. Speaker 7: Band and put into basically jail. You're talking a bunch. Speaker 3: Of that's right, that's right, that would be jail. But Speaker 3: you're heaven banned. Speaker 2: But I thought, and again, this is you know, like Speaker 2: I built this little am model myself and I like Speaker 2: showed it to my friends, like, oh, it's really cool, Joe. Speaker 2: I'm really oppressed, Like I'm really impressed by like that Speaker 2: you're able to do this. And I was like, are Speaker 2: people being honest with me? Have I been heaven banned? Speaker 2: Because I just like, like, you can be honest with Speaker 2: me if it sucks. Speaker 3: And I sort of have the fear. Speaker 4: The biggest humble braggad this thing and everyone thought it Speaker 4: was not great. Speaker 3: I'm just saying, like people are like I think people. Speaker 3: I'm worried that like people bring nice to me because like, Speaker 3: oh cool, Yeah that's repressed. You like did that. Speaker 2: And I have this like deep anxiety that like people Speaker 2: aren't giving it to me straight about it. I know Speaker 2: that sounds like a humble brag, but it's really not. Speaker 7: That's why you can never get like too successful, like Speaker 7: Maya West surrounded by a bunch of you never get. Speaker 2: Like, oh, this is his first try doing something with Speaker 2: vibe coding. I'm like deeply anxious, Like, no, you could Speaker 2: just tell me if it sucks, that's fine, that's my worry. Speaker 6: I don't worry about this. Speaker 4: If I tweet that I'm eating a steak, I will Speaker 4: get like a hundred people criticized and you didn't. Speaker 3: Put the meat. Speaker 2: Yeah. Speaker 5: Yeah. Speaker 2: So that's the other thing, which is that the two Speaker 2: things you are never allowed to tweet about meat preparation Speaker 2: and enjoying life, because if you ever enjoy life, then Speaker 2: if you ever enjoy it, and if you ever prepare. Speaker 3: Meat, people will flip out at you on the internet. Speaker 3: Those are the two things that you're not allowed to Speaker 3: do online. Speaker 4: Very true, this sort of related question, But just going Speaker 4: back to the methodology, if you're focused on this sort Speaker 4: of like path dependent idea, I'm kind of envisioning it Speaker 4: as like a giant decision tree, right, is there a Speaker 4: possibility that as the models get better and better, and Speaker 4: we know that they're already injecting like some degree of Speaker 4: randomness into their output. Although I know there's going to Speaker 4: be a pedant out there who like messages me and Speaker 4: says like, well, you know computers can't do like true randomness. Speaker 4: But setting that aside, setting that aside, like, we know Speaker 4: that they're adjusting, they're becoming more sophisticated at an incredible rate. Speaker 4: We know that they're trying to adjust and inject some Speaker 4: randomness in order to avoid exactly this kind of detection. Speaker 4: Do you worry about their own adaptation at all? Speaker 7: I have noticed that the models as they get more capable, Speaker 7: I believe that their output distribution gets more complex. It's Speaker 7: harder to learn with a simple model, which is why Speaker 7: we've been increasing our model size to capture a higher Speaker 7: complexity function that can capture the LM outputs. So I Speaker 7: think we may have to continue to make our models better. Speaker 7: We're gonna have to work to keep up with it. Speaker 7: We can't just rest on our laurels. Speaker 3: What our birstiness and perplexity. Speaker 7: Yeah, so this is a metric that's used by some Speaker 7: AI detectors, but not Pangram okay, And so I can Speaker 7: explain a bit about how it works. So perplexity is Speaker 7: Basically a measure of this. Speaker 2: Is not perplexity dot AI the website. This is a Speaker 2: technical term. Speaker 7: Okay, this is a metric. This is a measure of Speaker 7: how confusing a piece of text is to a language model. Speaker 7: So basically, if, for example, with every token, we can Speaker 7: calculate some perplexity, which is basically like how expected is Speaker 7: this is. So for example, like if it's I went Speaker 7: home to my pet and then the next token is chinchilla, Speaker 7: that'd be a much higher perplexity token. Speaker 5: Than my pet dog. Speaker 7: So low perplexity text or really like LM outputs tend Speaker 7: to be low perplexity. They're not going to produce outputs Speaker 7: that are surprising to themselves. So this is a decent Speaker 7: way to get an AI detector that's around ninety to Speaker 7: ninety five percent accurate. But it has some problems. The Speaker 7: main one is that you can't improve upon it. Basically Speaker 7: it has false positives. Text written by non native English Speaker 7: speakers often is low perplexity just because when you're late. Speaker 3: Don't take as many risks. Exactly. Speaker 7: Yeah, interesting, Yeah, So that's why a lot of the Speaker 7: early AI detectors had a bunch of false positives. With Speaker 7: ESL speakers. It's because their text was low perplexity. So Speaker 7: I think, like, this is a very cool metric, but Speaker 7: it is not the path for pangram. Speaker 5: Instead, we went the deep approach, so we can do Speaker 5: better than. Speaker 3: And what's in this is that just the opposite side Speaker 3: of the coin. Speaker 7: Yeah, Burstinus is basically actually, yeah, I don't know if Speaker 7: I can define it. Speaker 4: Okay, fine, Burstinus just sounds like one of those like Speaker 4: sort of I guess manosphere terms, doesn't it like, oh, Speaker 4: yeah he. Speaker 6: Has like he's been looksmaxing with high burst nets or Speaker 6: something like that. Speaker 3: Yeah, that's great. Speaker 7: Yeah, I think it might just be like a measure Speaker 7: of like sentence Lengthen, how the ups and downs of Speaker 7: the text. Speaker 4: If we assume that the world is collectively concerned about Speaker 4: AI slop and wants to do something about it, what Speaker 4: would be like the single biggest change to the system, Speaker 4: either in terms of like the economics of the internet Speaker 4: or regulation or technology like what you're developing that would Speaker 4: actually help reduce slop. Speaker 7: Yeah, I think the biggest one is norms. So there Speaker 7: have been a couple of great blog posts written about Speaker 7: how it is rude to send other people undisclosed AI outputs, Speaker 7: and I think I like completely agree here. I think, Speaker 7: you know, if somebody like asks the question on the Speaker 7: Internet and then somebody else like goes and puts into Speaker 7: chat CHEPT and then like pace the answer, it's kind Speaker 7: of rude, Like like I was going here to ask Speaker 7: the opinions of my friends or you know, my followers, not. Speaker 5: Just like not chat GPT. I could have done that myself. Speaker 7: And so I think, like building this norm is something Speaker 7: that you know, it's very new technology, so we need Speaker 7: to do it quickly. Speaker 5: But I think this would help a lot for society. Speaker 2: Well then actually just gets to a question that I Speaker 2: have then, which is I feel as though the major Speaker 2: Internet platforms are actually moving the exact opposite direction. I mean, Speaker 2: I'm stunned. Maybe I accidentally clicked on something at some point, Speaker 2: but the frequency with which I can email and then Speaker 2: I open it up to respond in Gmail, and there's Speaker 2: that ghost text there that do you just want GEM Speaker 2: and I to respond to this? Speaker 3: I've never done. Speaker 2: That, I also consider, I think that would be extremely rude. Speaker 2: I've never responded to any email with AI respond But Speaker 2: they're basically telling you to do that. They're doing the Speaker 2: exact opposite blowing up these norms, And so I'm curious Speaker 2: from your perspective, you managed to work with Quorra, But Speaker 2: from your impression, do the major internet platforms think this Speaker 2: is a problem worth solving or from their consider and Speaker 2: it is like you know what, Yeah, it feels content Speaker 2: the better. Speaker 4: There's mixed incentives for the big company. Speaker 7: It's funny because like Google seems to be playing both sides. Speaker 7: So like, on one hand, they had that advertisement which Speaker 7: people kind of blew up about where it's like, oh, Speaker 7: children can now send their heroes notes on like how Speaker 7: much they respect them by using AI instead of like Speaker 7: writing the note themselves, and like this is wrong, This Speaker 7: is like societally bad. But at the same time, they're Speaker 7: working very hard to deal with the AI slop on Speaker 7: the Internet in search results to make sure people get Speaker 7: served real content and not. Speaker 5: AI slot content. Speaker 7: So I think, I mean, I think obviously there's a Speaker 7: lot of incentives that play up around like product people Speaker 7: who are incentivized to push AI because that is the Speaker 7: corporate mandate. But yeah, I think overall, even like in Speaker 7: my sphere, a bunch of people who are AI researchers, Speaker 7: generally consensus is that like AI is a powerful tool, Speaker 7: but like slop is bad. Speaker 4: This reminds me my parents used to make me do Speaker 4: these like handmade greeting cards for every you know, for Christmas, Speaker 4: for like all relatives and stuff. And it was supposed Speaker 4: to be a demonstration of my commitment to communicating family. No, no, Speaker 4: it traumatized me forever. And I hate greeting cards as Speaker 4: a result of them of doing this, just spending hours Speaker 4: manufacturing these things. But then, secondly, the funniest thing was Speaker 4: once we got E cards, my parents immediately switched to Speaker 4: using e cards and just and now this is also Speaker 4: the funniest thing. Speaker 6: My dad uses E card. Speaker 4: He figured out that the E card system can tell Speaker 4: him whether or not you opened it, so he just Speaker 4: uses it as like day to day communication. Speaker 5: Now that's so funny. Speaker 3: Just send an email to your daughter E card. Speaker 4: It's like, I noticed you haven't opened up my E Speaker 4: card for International Hot Dog Day. Please let me know Speaker 4: what's going on. Speaker 2: I'm terrible handwriting as a kid, and my mother made Speaker 2: me write all of these handwritten notes to thank people Speaker 2: for the gifts I got for. Speaker 3: My bar mitzvah. Speaker 2: Yeah, I hated it, but you know what, I have Speaker 2: keep connections with all of. Speaker 3: Those people that have lasted over the years. Speaker 2: In that miserable one week where I just wrote and Speaker 2: I got, you know, hand creamped, I think it. Speaker 3: Paid off, all right. Speaker 4: Well, imagine doing that for like sixteen years basically in Speaker 4: a never ending stream. Speaker 3: Max Birou, thank you so much for coming on out Laws. Speaker 3: That was a lot of fun. I'm fascinated by this conversation. Speaker 7: Thanks so much for having me. Yeah, really exciting to Speaker 7: talk about this. And I think slaps is a growing problem, Speaker 7: so hopefully awesome RAPK deal with it. Speaker 6: Of the internet, I. Speaker 4: Can't tell if I'm surprised by that oring on. Speaker 3: And what's it going to be next year at this time? Speaker 5: Oh man, I don't know. Speaker 3: It'll be like hard to stay over with Georgian that Speaker 3: for sure. Speaker 5: Yeah, almost certainly crazy. Speaker 3: All right, thanks for coming on. Speaker 5: Oudlin, Thanks. Speaker 3: Tracy. I love that conversation. Speaker 2: I just think it's like a really fun puzzle, right, Speaker 2: It's very like it seems like a fun question to solve, Speaker 2: And I'm fascinated by this idea of how like with Speaker 2: both humans and AI, there's gonna be this gap inevitable Speaker 2: between what we know and what we can articulate because Speaker 2: you and I both setting aside a a versus text, Speaker 2: there are things that we both know. For example, this Speaker 2: is newsworthy, and this is this is a good episode Speaker 2: of a podcast, This is a credible sounding guest, and Speaker 2: this isn't the gap between that and then being able Speaker 2: to explain why, it's like, well, you just sort of Speaker 2: know it, right, You just sort of have this feeling there, Speaker 2: and that intuition is built up from numerous examples, which Speaker 2: is the same way in a sense that like the Speaker 2: AI is trained. Speaker 3: It's like these. Speaker 2: Things that you only know from patterns and you can Speaker 2: see them without fully being able to, like article exactly Speaker 2: what's going on. Speaker 6: Well, the other. Speaker 4: Question I would have on that is is it even Speaker 4: going to matter in the long run if you think about, Speaker 4: like so much of the Internet is already built on Speaker 4: bots and the sort of like false attention economy, Like Speaker 4: if our entire like worldview becomes shaped by AI driven drivel, yeah, Speaker 4: does it matter if like the economics of the Internet Speaker 4: are still attached to individual bought accounts and things like that. Speaker 6: I don't know if I'm if I'm explaining this, but. Speaker 2: No, no, I think it makes a lot of sense, Speaker 2: and I do think like it is important, like we're. Speaker 3: Going to have to change the entire way with them. Speaker 2: And Max said at the beginning, which is, and I've Speaker 2: thought about this, which is that it used to be Speaker 2: that if you came across a piece of writing and Speaker 2: the punctuation was excellent and the spelling was excellent, and Speaker 2: it was like cogent sounding, you're like, okay, this has Speaker 2: been written by a smart person. I will read the seriously, right, Speaker 2: And now there is this complete severance of sort of Speaker 2: like craft and out put because you could and you Speaker 2: did this, Like, ask Claude to write an argument in Speaker 2: favor of the most absurd proposition imaginable. Ask Claude to Speaker 2: write an argument for me that the reason why Reagan Speaker 2: wanted to do tax cuts in the early nineteen eighties Speaker 2: related to these reports of UFO sightings in the nineteen seventies, Speaker 2: and it will write something that not only is it Speaker 2: grammatically correct, it'll actually like strain to come up with Speaker 2: the best version of this argument before and again if Speaker 2: prior to that, having read and like, maybe the person Speaker 2: like this person took this argument seriously, but now this Speaker 2: argument is just created. Ax nail Oh We're going to Speaker 2: have to really like change our heuristics about this stuff. Speaker 4: We've created an unlimited stream of basically cranks, which is Speaker 4: really good grammar. Speaker 2: Yeah, that's right, that's right, because it used to be Speaker 2: we knew the crank because they had bad grammar, or Speaker 2: they would email us and like half the words would Speaker 2: be in yellow and the other half would be underlined green. Speaker 4: Inlastic exams, the tools that we use to just like, oh, Speaker 4: this person's a crank, they like, you know, half the Speaker 4: words are at all caps and stuff like that. Speaker 3: Those don't work anymore. Speaker 4: All right, on that note, shall we leave it there? Speaker 3: Let's save it there. Speaker 4: This has been another episode of the Authlots podcast. I'm Speaker 4: Tracy Alloway. You can follow me at Tracy Alloway. Speaker 2: And I'm joll Wisenthal. You can follow me at the Stalwart. Speaker 2: Follow our guest Max Spiro. He's at Max Underscore Spiro Underscore. Speaker 2: Follow our producers Carmen Rodriguez at Carmen Arman, dash Sho Speaker 2: Bennett at Dashbot, and Cal Brooks at Kilbrooks. And for Speaker 2: more oddloss content, go to Bloomberg dot com slash odd Lots. Speaker 2: We're a daily newsletter and all of our episodes, and Speaker 2: you can chat about all of these topics twenty four Speaker 2: to seven in our discord discord dot gg slash od Speaker 2: lots And. Speaker 4: If you enjoy odlots, if you like it when we Speaker 4: talk about how the Internet is forty percent slop, then Speaker 4: please leave us a positive review on your favorite podcast platform. Speaker 4: And remember, if you are a Bloomberg subscriber, you can Speaker 4: listen to all of our episodes absolutely ad free. All Speaker 4: you need to do is find the Bloomberg channel on Speaker 4: Apple Podcasts and follow the instructions there. Speaker 6: Thanks listening,