Episode Transcript
[00:00:05] Welcome to Prompting Curiosity, a podcast for the AI curious. No coding background required. I'm your host, Dr. Shantae Cofield, also known as the Maestro, and I created this show to explore what these AI tools actually are. Really, though, are the files in the computer, how to use them, and what they might mean for how we think, work, create, and move through life. Whether you're skeptical, intrigued, or already experimenting, you're in the right place. All that I ask is that you stay curious. All right, let's get into it.
[00:00:38] Hello, hello, hello, my curious people, and welcome to episode 59. And also welcome to September, uh, of Prompting Curiosity. I'm your grateful host, the Maestro, and today we are talking about frontier models. Now, frontier model is a term that I've absolutely used before. And I don't want to assume that you actually know what it means, because in full transparency, prior to doing the background research for this episode, I only had, like, oh, there's Hear the noise. That's Rupert. I only had a, you know, a loose understanding of what it meant. And so I was like, let me, um, dig and see if it means something else besides what I have come to understand it to mean. Uh, and so today we're gonna keep with the what the series and briefly cover what the A Frontier model is. I think this is going to be a little bit of a shorter episode, but, uh, Curiosity has no size requirements. So jumping into the main topic here, most people think. I, at least I think that I am most people, that when you hear the term frontier model, you it refers to, uh, like the most established AI models that are out there, right? Like ChatGPT, Claude, Gemini, and technically it does. Right? But that's not the actual technical definition. So let's start there. So the term frontier model was first coined in July of 2023, uh, and that was when anthropic, Google, Microsoft and OpenAI, they formed, wait for it, the Frontier Model Forum. We're going to refer to that as the fmf. It's just easier. So the fmf, it still exists today.
[00:02:17] It was started as, uh, what we'll call an industry body that was focused on ensuring safe and responsible development of AI models. Right. So they're like, we are about safety. The FMF is about AI safety.
[00:02:30] Per the fmf, a frontier model was defined as a model that beats everything else that's been widely deployed for the past 12 months. Right. So like, the most capable model, the EU AI act, which I spoke about maybe last episode, I don't remember which episode it was, but either way, that ACT now bases that demarcation of a frontier model off of what they call the training compute amount, which is basically how much math is done to train the model. So it's not like the true size of the model, it's, um, how much math is done in training it, which is notable because, because of that, that like that very objective measure. Now, any new companies that want to build a model that quote, unquote, you know, complex, and if they want to operate in the eu, they have to build the safety evaluations, the reporting, all the compliance infrastructure that's required by the EU based on that size. Right. Because that's what they have, uh, you know, made the cutoff as. So to recap here, I know I'm talking a lot, saying a lot of words. To recap here, the term frontier came about in 2023, and that was when all the biggest players in the AI game at the time, they got together and they said, hey, we're going to form an AI safety committee that includes only us and we're going to call it the Frontier Model Forum. Um, thus the forum was self governed and self named and it set the standard for what a frontier model was. Right.
[00:04:01] Worth noting, I think, uh, the FMF does still exist and it's expanded to include two more members. So it went from four to six. Um, and it brought in Amazon and Meta, which I, I feel like you would have guessed. Right. And that happened in 2024.
[00:04:15] So what does the FMX FMF actually do?
[00:04:18] They publish safety research and technical white papers on things like biosafety, like biosafety, threshold Cyber Defense. And we're going to talk about this later, but it's called Adversarial Distillation, where basically you train another model, you use a model to train another model illegally.
[00:04:33] Um, they, the FMF runs the AI Safety Fund, which, and I could be wrong about this, feels like it's more for show than anything, given the fund launched at, uh, $10 million in October 2023 and has yet to grow, despite the fact that these AI companies have absolutely insane valuations. Right. They're doing deals that are, you know, hundreds of billions of dollars then. And billions of dollars. And this is insane. And it's like maybe not hundreds of billions, I mean, maybe that's exaggerating, but either way, hundreds of millions of dollars. Billion and into the billions.
[00:05:06] And, uh, the fund is $10 million, that's like a penny. So I'm like, this feels like it's all for show.
[00:05:13] Um, and also as members of the fmf, um, all of the members, all six of them. Now they signed an agreement to share information about vulnerabilities, threats, um, and like capabilities of a concern that were specific to, to AI. Right. So if you're anything like me, AKA you're hella suspicious and skeptical of these fucking companies, hopefully, or you're likely asking why they started this thing in the front first place. Right? Why did they start the Frontier model form in the first place? Ak, what did they have to gain? Because there's no reason, there's no way that they're just doing this out of the goodness of their hearts. Yeah, right. So they did have things to gain and you are right to ask that question because it did in fact benefit them. So three things at least the three things that I want to cover in this episode. So one that they gained from forming the FMF is what's called regulatory capture. So at the time governments were actively writing AI regulation and so informing or by forming the fmf, it allowed those four companies at the time to define the category that they'd be regulated under. Right. Because like the governments really didn't know, they couldn't say. So they're like, hey, we'll step up and we'll help you out and help ourselves out.
[00:06:25] Second, brand positioning. So also at the same time, by 2023 and even definitely now, but in 2023 AI was already being associated with deep fakes and bots and just like a negative perception. So by calling themselves Frontier, they, they position themselves as, you know, the serious safety consciousness, you know, companies, the serious safety conscious models. Uh, and then lastly, which I kind of spoke to a little bit before, is the legal coordination. So antitrust law prevents competitors from sharing sensitive technical information. Right?
[00:06:59] By forming the FMF, OpenAI anthropic Google, they were now legally able to share this information, right? Because they're part of this forum. So they could share this information about threats, quote, unquote threats like Chinese adversarial distillation. So I've talked about this before, but uh, distillation, like I said earlier, is when you, you query a model and you use it to train a comp, a uh, competing model. Adversarial distillation would be like, hey, a Chinese company is going to train its own model on Claude, right? By asking it a bunch of questions, getting the answers, seeing, you know, the outputs and then just train, using it to train its own model. That's called distillation. So by forming the FMF, it allowed them to partner together. OpenAI anthropic Google to partner together to talk about this and share and be like, hey, is this happening? Is this happening? Whether or not it's fucking happening, I don't know. But there's a legal coordination piece. So, yes, forming the FMF and categorizing themselves as Frontier was absolutely net positive.
[00:08:07] But speaking of these Chinese models, right, by the FMF's own definition of Frontier, which is that a model that beats everything else that's been widely deployed for the past 12 months, these new open weight models like Kimi's, you know, K3, check, um, out episode 55. I talked about open weight models. Um, but according to that definition, Kimmy's K3 would qualify, you, uh, know, and, and be able to use that moniker of Frontier.
[00:08:35] But what I'm seeing in practice, because you all know I just, in these AI silos. But what I'm seeing in practice are labels like Frontier grade or near frontier, which to me supports the idea that the term frontier is really more so kind of tied to like a specific club.
[00:08:53] Uh, right. This is tied to a club and the government relationships that are built around that club. Not so much about. We're not solely about benchmark scores, right. It feels kind of like the old boys club of AI. So to close the loop here, right. By definition, a Frontier model is either one of the most capable AI models that already exists or one that is big enough, as measured by how much computing power went into training it, how much math went into it, that regulators think it could cause serious harm if something goes wrong. So then it would earn that moniker, that title of frontier.
[00:09:27] Right? So by that you read the term, you hear the term Frontier model, your brain's automatically going to think Claude, ChatGPT, Google Meta. Because they do fall within that standard. Right? However, given the politics by all of this, I don't think you would actually be wrong if you colloquially interpreted Frontier model to mean one of the most established models all. Uh, right. But, but, but, but as with all things AI, we'll see how and if this changes in the next five minutes. All right, I told you shorter episode. Real quick before I officially wrap it up, how I used AI this week. So if you're new here, welcome. Uh, each episode I share a quick example of how I used AI that week. This week I used Claw. That's my guy. I French assistant to build a workout tracker for my mom. The whole thing took me like two hours, maybe that long. Um, I planned it with Claude and then I built it with Claude code. And, uh, the New thing that I did this time for those who like to nerd out, is that I used Google Sheets as the database, right? So if you are going to build some sort of personal web app and you want it to remember things across sessions, you have to have a database associated with it. So usually I use Supabase, but I have maxed out their, you know, it's a generous free tier. Generous, generous enough. Um, and I've been wanting to play around with just using Google Sheets as the database. I'm like, is this a single person? And I don't need all the functionality of Supabase. Right. One of the main things of Supabase is uh, what's called row, uh, level protection. So this just means like everyone can have their own uh, like login. That's essentially what happens with that, is that you can protect each of these rows where the information would go in. And so this way it's specific to a specific user or like a sign in state. Right?
[00:11:15] I don't need that. It's just for my mom, I need to be able to see as well. I'm programming for her and so I just need to be able to see it. She needs to be able to see it. Nothing. There's no like, you know, sign in that's needed. And that's also annoying because then you get signed out. You don't need any of that shit. Uh, so I was like, I've been wanting to do this. I have two other apps, personal apps that I'm like, I actually could switch the database over to uh, Google Sheet, but I want to try it out first because changing things, I don't want to say that changing things is annoying because I'm not the one fucking doing it. Like Claude, you know, cloud code is doing it, but it's just like I would rather build it from the, from the get go. And so that's, I was like, let me try it with this one. Um, and it was great. It worked out great. It is really cool to see how far I've come in my capabilities and my understanding of things. Um, the first thing that I built out was actually a website for my brother. And then I built out that um, like the pill tracker for my sister for like medication tracker. And there was just like so many errors and issues and like things weren't working and like I didn't know what to do. I didn't know all you do. Even now something's not working. I like just, you know, screenshot it and put it back into Claude and I'M like, this is what I say. But, like, you know, yes, part of this speaks to how much better Claude is than it was a year ago. But, like, understanding the process and having these reps, having gotten so many reps, it helps so much because now I understand, like, I plan the thing out. I know what to ask Claude. I know what. What framework I want to use. I know the tech that I want to use. I know how I want it to be built. I understand how to test things.
[00:12:40] I, uh, understand how to plan things. And so I. And, you know, I use Claude, and I actually plan things out, and then I have it make a markdown file or have it, you know, just copy this text that I can copy, and I, uh, just paste that into Claude and then into cloud code. And it goes through one phase at a time. Because I asked Claude for that. I'm like, hey, so I want to build. These are the things that I want. These are the specs on it. And, you know, ask me any questions that would be helpful. And I want to be able to just paste this directly into cloud code, But I need it to go, you know, stepwise in individual phases. I don't just want a one shot of it, because that's when things get messed up.
[00:13:12] And there was no errors. There's no issues like, is working, knock on wood. But it is really cool to see the progress with this and how. Just how functional this is. That's awesome. You know, I really do believe that these. These personal web apps are definitely one of the best use cases for vibe coding. And it was really cool to just be able to create a solution for my mom, right? And I customized it, made it purple to. For every color. Personalized it, like, says, hey, mom, here's your workout for today.
[00:13:38] Um, you know, I built it while she slept, and then it was waiting for her in her text messages in the morning, right? And it looks like an actual app. It's a web app. It's a website, right? But it looks like a web app, and she's able to use it. It's super intuitive. It can be exactly what I want it to be. And then, you know, she tries it out, and I'm like, all right, what don't you like? What do you like? And then I can change that, and that's fucking awesome to me, right? And it's literally like, no sweat off my back. I'm like, all right, cool. I'll make this thing for you. So I think it's dope. I don't know if you've been playing around with with building your own personal web apps. If you have, would love to hear from you.
[00:14:11] But that, my friends, is all for today. Hopefully you found this episode, uh, helpful. I got a little ask. If you found it helpful or you find any helpful episodes helpful, consider sharing them with someone who you know is curious about AI. Right? AI Curious. Let's grow this podcast. Let's get some more listeners here. Not that you folks aren't great, but just, you know, get some new people. Don't forget I also have a companion newsletter and blog, the Curious Companion, that drops every Thursday. And it's basically by basically, I mean exactly the podcast episode in text format. So if you prefer to read or you just want a written record of the things you know, easy to follow along, join the newsletter fam or check out the blog. You can head to prompting curiosity.com forward/newsletter or forward/ blog. Or keep it simple and just check out the links in the show notes.
[00:14:57] As always, endlessly, endlessly, endlessly appreciative for every single one of you. Until we chat again next Thursday, stay curious.
