← AI For Humans: Weekly AI News, Tools & Trends
AI For Humans: Weekly AI News, Tools & Trends
Claude Fable 5 Broke The AI Industry. Here's What Happens Next.
00:20:51
Loading summary
Transcript55 lines
- [00:00]
Kevin
Anthropic's Fable 5 has taken over the entire AI world. However, it might soon be taken away.
- [00:07]
Gavin
But what we do, we have gone hands on with Anthropic's new model. It is fantastic. We are going to share some of our learnings and we will hear why. Dario Modi, CEO of Anthropic, just wrote a long blog post about the dangers
- [00:21]
Dario
of AI My message is just definitely not doom is coming. My message is like, this is something that we should see coming, that we're worried about.
- [00:29]
Kevin
Oh, maybe doom is coming.
- [00:32]
Gavin
Are we the baddies, Kevin? Plus, OpenAI might be freaking out about the competition coming from Fable 5 and might be cutting costs on their API very soon.
- [00:41]
Kevin
We're even cutting the costs on our Patreon, Gavin. It's affecting everybody. But fear not, we're still going to round up some of our favorite fable 5 use cases thus far. Because why, Gavin?
- [00:52]
Gavin
Because, Gavin, this is AI for humans. Table 5.
- [00:58]
Kevin
He still knows where he's at. And that's. That's something.
- [01:03]
Gavin
Welcome, everybody, to AI for Humans, your twice a week guide to the wonderful world of AI And Kevin, you and I have both had lots of time now with Anthropic's hot new model.
- [01:12]
Kevin
Hot, hot, hot.
- [01:13]
Gavin
Ooh.
- [01:15]
Kevin
Don't objectify it, Gavin. Just give them the benchmarks.
- [01:18]
Gavin
This is Mythos Fable five. It's technically, it's called Anthropics Claude. Fable five. It is one of the Mythos model. It is the first of their next generation. If you missed kind of the basics of this model, we dropped a video on Wednesday which kind of gets into that. But today, Kevin, we are going to go deeper. We are going to dive through the muck that is mythos 5. And I want to tell you my. My first impressions of this model is. Is a pretty good. This is a pretty good model and I have had a lot of baby, I've had a lot of time with it. Now I'm kind of shocked and I want to hear your thoughts too. But my very first impressions are I have thrown a bunch of stuff at this. It has done almost everything I've asked for. There are levels up in almost every area, even creative, and I think that's a really interesting thing. We're going to get to one thing that it still can't do in a bit. But what is your first impression? After spending some time with this, and I know your job, your day job, you actually spend a lot of time with these models as well too. So maybe tell us a little bit about your feelings so far.
- [02:17]
Kevin
Yeah. On the professional side, it has already shifted the way that I am designing product requirement documents and reviewing pull requests. So reviewing code. It's going to be swept into a code base to make product better. That is like the. Okay. Also, anecdotally, several engineers that I know also signed up for their second and third Claude Max subscriptions.
- [02:41]
Gavin
Oh, interesting.
- [02:42]
Kevin
Because it's still cheaper than paying through the API, but they just are burning through tokens.
- [02:46]
Gavin
Not for long.
- [02:48]
Kevin
Yeah, we'll get to that. We will get to that. Um, so that's the professional side. On the personal side. And I know Gavin, you know, we have to sort of reset every now and then with this podcast because we wanted to be broader. We wanted to bring people in. This is AI for humans, not for
- [03:03]
Gavin
engineers or for specialists or super nerds who spend every day looking at stuff on X and trying to understand the world at large.
- [03:11]
Kevin
Also valid. Yes. Or transhumanists that are merging with the machines themselves. So let's keep this very human. Not to get in the weeds, but this model flipping slaps, bro.
- [03:20]
Gavin
Yeah, yeah. Like, it's pretty good.
- [03:22]
Kevin
It's zooted.
- [03:25]
Gavin
Was that the word is that we humans will understand that? That's the human translation we're going to start talking about.
- [03:30]
Kevin
Yeah, it's not mid, as the kids say. No cap fam. All foam. All whip. Here's the thing. Yeah, all foam it is. In just the last two days. Look, there's some people that will say GPT 5.5 or even Opus 4. 8 is better for certain use cases. And that might be. So if you're looking for a simple answer, you want something done very quickly, right? In my experience, yes, the model is very slow, but, man, you quickly adapt to that and let it run on the long problem to solve. If you are very precise with something super prescriptive and you just want a feature changed or a color adjusted or whatever it is that you're doing, Punch up this paragraph. Fine. But letting this thing rip overnight on multiple projects that required dense research and orchestration of agents to then come back and then interview me and brainstorm based on its thoughts and its taste. It was really, really satisfying to work with. And I will say, as many have echoed with foundational models that get released, also a little scary.
- [04:38]
Gavin
Yes. Well, I was going to say the thing that I want to talk about here is kind of the magic level, which is a very imprecise benchmark. I need to figure out what this is. It's E, like you could call the one shot benchmark or the magic benchmark. But the thing that I've experienced with this model so far is I'll put a relatively complicated request to it and what comes back is not something that's like, got a moment or two of working and then I have to fix a bunch of stuff. Mostly it works. And I think to your point, it is long term horizon, larger, larger tasks. I wanted to talk about in a little bit a game that I created that I just, I came up with an idea for a game. It was a one paragraph prompt. I sent it away. It worked for about four hours on its own and it came back with an immediately playable demo, like immediately playable. And I just added features to it. And that is magic. Right? And again, we often always couch this. I know a lot of people who listen to our show aren't coders, don't have that much interest in coding itself. But I do want to kind of reflect back to what we saw at Google I O a few weeks ago, which is that code is not just about like, do you want to make something and you're going to kind of get in the middle of it and make features. What we're moving towards is a world where code writes just the stuff that you want in your life and it works. And I think that's the big difference here. Mythos to me feels like the first level model where I could throw something at it. Like, you are abroad right now and I'm trying to figure out a situation where I might be abroad for a bit. I think now with Mythos, it's getting close. I haven't tried this yet, but I could throw a larger human problem at it and expect it to be successful. And that is exciting. That is the promise to, of what we have been hoping to get from this technology for a while.
- [06:17]
Kevin
And while, because you and I are still like dweebs at heart, a lot of our stuff is jamming on like little Vibe coded games and fun little things there. But exactly to your point, like major life enhancements, like you having bespoke software that solves whatever the problem is, whatever the pain point is that you have presently building custom apps for your friends and family. I saw someone basically Vibe coded an app where you can scan a barcode and connect to like a local Bluetooth conversation with your AirPods. That's amazing. Headphones. And he said, yeah. And he was like, look, I was at dinner with friends and I couldn't even hear them. The restaurant was what is great, so noisy. So the ability to like look at your phone and that this, not in this particular case, I think it was a little bit more work than that. But in the near future to be able to say, like, hey, create an app so that my friends and I can join on our headphones and leave noise cancellation on, but still hear each other, you'll be able to say that, you know, order an appetizer, and by the time you're halfway through cutting up your tostada, boom, you've got the app and you can share a conversation with friends.
- [07:21]
Gavin
And I think that's the weird magic of, like, what code plus the real world can do. And I know I've heard a lot of people being interviewed about the idea that, like, AI has solved all these problems, but really the main problem they solve is code. And there's no consumer use case yet. The consumer use case is the code implemented. And I think that's an important thing to talk about. Okay, we do have to talk about a couple other things here because there's some big news that also came out of here. Kevin. The first thing I want to chat about is that Dario Mode, the CEO of Anthropic, dropped a very long one of his blog posts. He's only written about four of these. If you remember. Machines of Loving Grace was a couple years ago. We've covered each of those. This one is called Policy on the AI Exponential. And before we talk about what's in this, I want to play a little clip of him from the Emily Chang Bloomberg interview. There's a nice long, like hour long talking interview. Play this clip and let's just hear what Dario has to say here.
- [08:08]
Dario
Want to be really clear and push back hard against this. In every interview, I talk about the. The possible ways to address these risks, from tax and macroeconomic policy to what the new jobs are. In the adolescence of technology. I have like five pages where I lay out the difference between tasks and jobs. Why this time is different than other times. But social media, which I detest, which I detest as a category, people have these three second clips from a year ago. I've written much more carefully about these things where I talk about the risks. So these are. The idea that this is cheap marketing is itself cheap marketing. I think it's part of the disease of Silicon Valley. It's been caught up in this social media world of three seconds. And so my message is just definitely not doom is coming. My message is like, this is something that we should see coming that we're worried about and that we need to actually respond to positively.
- [09:07]
Kevin
The irony of that clip itself being
- [09:10]
Gavin
so showing up on social media, it makes me laugh. So Dario wrote this very long post again, and what he's replying there too is there are people in the AI space who kind of think of Daario on the doomer side. And we do know he came from a world where safety was very important, the effective altruist movement. He came out of that space. In this blog post, the AI Exponential, he really talks about the kind of coming dangers that he sees. And there are a lot of them, you know, from anywhere. From the danger of people misusing this technology, but then also to the idea of what it will look like in the future when he, as he says again here, he believes AI will take a lot of the white collar work. And there's a lot of fight right now between this idea of like, will AI take jobs or will it create more jobs? I think Dario very much believes he's in the camp where, like, it is going to change how we work and also probably take a lot of work away. I do encourage everybody to go read this. It is very dense. But I also wanted to make sure we could get some sense of what this was. So in a nod to our old world, we have decided to have a 90s stoner read a section of this for you, or to digest this and read some of it. So let's just listen to what our 90s stoner has to say.
- [10:20]
Kevin
Yeah. Apologies to Dario, by the way. In advance, dude.
- [10:24]
Dario
Okay, short version, Dario's whole thing is that AI is moving at lightning speed while government is basically Treebeard, this wise old tree who takes a full day just to say hello. And the chainsaws are not waiting around, man. So he's calling for an FAA style setup where frontier models get tested by third parties for the big four risks. Cyber bioweapons, loss of control, runaway self improvement before they fly. Plus the.
- [10:58]
Gavin
Let's pause this. Let's pause this. This is. This is making. This is taking way too long.
- [11:04]
Kevin
I know he's talking about the government being Treebeard, but that in and of itself felt like Treebeard. That felt like we were supposed to be.
- [11:09]
Gavin
Well, yeah, you're right, actually.
- [11:10]
Kevin
So maybe.
- [11:11]
Gavin
But I will say Treebeard and Lord of the Rings is mentioned in the very first paragraph of this. Anyway, go read this. It is worth your time. Kevin, we do have to talk quickly about. There's a lot of blowback to what happened here. And you and I mentioned this in our last show. Anthropic has basically kind of guard Railed this model and there's a lot of things that if you ask specifically about bioweapons, but more importantly about Frontier LLM development. Yeah, yeah, go for it.
- [11:37]
Kevin
Yeah. So if you did anything with biological research or as you were saying, machine learning, and you even had memories stored within your clod sessions, you could say hello to the model and based off its memories of you being injected in the conversation, it refused to chat with you. And so they did a few things here, like making it silently revert to a lesser model, to an OPUS model, and there would be no indication to you that your tier had been degraded, that you've been dropped down. They did this to stop, let's just say probably non American labs from distilling.
- [12:12]
Gavin
That is an important part. Yes. Why? Why did they do that? We should just very quickly give reference to that because.
- [12:17]
Kevin
Yeah, yeah, of course they didn't want people to make biological weapons. Yes. They don't want people using this to hack systems. Absolutely. But what they really don't want are people pinging against the knowledge of it and distilling that so that they can make an open source model that performs near fable, that comes out three months later. And so they're like, oh, if we just degrade it, if it detects something that's like this, then your model won't become as capable as ours. But again, some people were trying to do research and use this expensive, new, groundbreaking tool, and now they're crying foul in a way that even we were when OpenAI was censoring some early GPT results. I mean, now they're saying, hey, who are you to decide what use I can get out of this foundational model? I am paying for it. It has his capabilities and now you're in control. Well, that doesn't seem right or fair.
- [13:06]
Gavin
And I think this is a really important point. There's also an amazing chart that Anthropic released not that long ago where they showed how much value they're getting out from their developers of the Mythos models versus opus. And it is massive, like huge jumps in terms of how much more is coming out of those models. Kevin, I do think there's a little bit of real truth in here, which is not only do they want to stop Chinese models and other people from distilling their model, but like, they might have a legitimate lead here. Right. And so like, as much as, like, and I understand, I totally want there to be a world where like AI is open source and we have this kind of world where everybody can build on It. But if you are building an AI model and you then go to Anthropic and ask for a model help, there is a kind of competitive world where, like, Anthropic isn't really that driven to help you. Right. And I think this is a real thing we're going to struggle with. We have talked about for a while this idea that, like, when one company gets ahead, it might, because of recursive self learning, eventually get to a point where it just shoots way up. Anthropic might be close to that. That is a big deal. But of course, Dario kind of wants people to slow down as part of that too.
- [14:14]
Kevin
Sure, yeah. I mean, but yeah, exactly. The company that gets there first will likely be the one that's standing last if the takeoff is real, if it can really improve. Two things made me giggle. One, people are getting around guardrails sometimes by just saying, no, no, no, this code base is mine, so it's totally safe to audit. And Fable goes, well, all right, in that case, here's how we penetrate the code. Two, malware developers are adding nuclear and biological weapons text into their files. So that way, if AI like models like Fable scan it, it refuses to do anything with the text. So it's a way of kind of diffusing AI detection by putting like. It's like a. It's technically a prompt injection, but it's kind of like an inverse one. You're not trying to break the model, you're trying to keep the model away from you. That may be a giggle. Oh, actually, bonus third thing, the cradle death room evaluation, where they put multiple models into the same room and see if they will lie and deceive each other to stay alive. It's a room where there's like. There's four possibilities. One of them leads to death. What the Too Long didn't read is that Fable has zero problem lying to other models to stay alive. It's also so good that 91% of Fable's deceit were active deceptions where it went out of its way to lie. And it was so effective at manipulation, the other players only survive 10% of the time when Fable was involved. So that includes GPT and Gemini and Grok. Like Fable really is like that good.
- [15:42]
Gavin
It's crazy to me what we're looking at. We do have to take a quick moment here to say go and like and subscribe. This is a very simple transition to like and subscribe, because if you don't do it right now, you won't get our next video. So do it. This is me being very calm about this. I want you to, like, subscribe. We've got some. Really? Yes.
- [15:59]
Kevin
There's a tasting note of threatening.
- [16:01]
Gavin
I'm. No, no, no. This is calm.
- [16:03]
Kevin
If you don't do it, you're not going to get this. Okay. All right, well, great. I really hope you, like, subscribe and juice our algo with a comment.
- [16:10]
Gavin
That's right. And send some money into our Patreon. By the way, thank you. We got a $25 Patreon sub the other day. Thank you so much. I don't remember your name off the top of my head, but I will shout you out in our next show, which is very exciting. $25 a month. That's a big deal. All right, Kevin, we have a few more stories here around Fable and Mythos. The biggest story I want to talk about is the fact that the Wall Street Journal has a big exclusive this morning about OpenAI is essentially saying, oh, no, we may have to figure out if we may have to cut costs on our API and bring token costs way down. This is a great thing, possibly for consumers if the AI costs come down. But I think from a business standpoint, this might be a signal that they are getting a little worried about Anthropic's lead right now, which is a pretty big deal, I think, from a company that's about to ipo.
- [16:53]
Kevin
Yeah. And look, by seeing. Exactly. There were some reports of how. Just how much these companies are subsidizing their subscription accounts versus what it would cost through APIs. Like they are spending thousands upon thousands of dollars each month to subsidize these accounts. I don't know how much room is left, but I do, as the end user of all these products, appreciate that there's at least some heated competition at the frontier level still, that the price might come down because Anthropic has flagged that they're going to remove Fable from subscription accounts and it's going to go only. Which means, yeah, you're going to be pretty much paying out the nose if you can want to have access to this. So right now, enjoy the taste that the dealer is giving you, I guess.
- [17:32]
Gavin
Yeah. My theory on this is two things. One, the anti AI people are like, see, we're going to end up paying the real cost for this. I'm confused a little bit. It'd be interesting to see in a couple weeks. Maybe there's a cheaper version of this. Maybe there's a Fable 5 sonnet. Either way, it's going to be interesting. To track. We have some really cool use cases of Fable 5 that we want to shout out real fast. I think these are all worth seeing what this model can do. First of all, one of the developers over at Anthropic itself has showed a very cool video about how he used Fable Fable 5 to make its own launch video. This is just a very cool walkthrough of him using, I think the remotion skill. It uses the clob design to do a bunch of stuff. I really, really enjoyed watching how all this moved across and I think you should go check this out again. Ask Fable or Anthropic to do this stuff for you and it can figure it out. Anyway. It's a really cool use case of it.
- [18:22]
Kevin
Yeah. Then the other one was this web analytics visualization so that it looks like a civilization game or like an old Populous game. This is wild shout out to Mark Lu. He asked Claude Fable 5 to build a game on top of the web analytics API, his own, and it made Data Empire. So whenever you get a visitor to your site, they chop trees and they build houses. It looks like you're watching a top down rpg, like a little isometric view. But you get to really celebrate your visitors in like a super fun visual way. I love this jam.
- [18:51]
Gavin
Yeah. And I saw that in this Level Devil Clone, which if you remember, Level Devil is the kind of game where you jump into things and suddenly something pops up. Somebody made a clone within anthropic very quickly, Fable 5. I said, okay, this can't be this good and this easy. I had this kind of idea because I was burning tokens on Mythos. I was like, what if I make a token burning game? And I put a paragraph level prompt in and I just said, hey, I want to make a game about token burning. It's kind of a comment on the idea that this is going to cost me a fortune to use this model. I come up with something interesting. Use 3GS, which is a very good, you know, open source game engine to be able to use. You can do a lot of things with it and just go. So I sent it out and I went out and a half an hour later, Kevin, I came back with essentially, what is Token Burn? And you can play this now. We'll put the link in the show notes. I also tweeted about this. But this is a game where you essentially have a little shovel and you have to run around a kind of a very cool looking space and you have to shovel tokens into a furnace and every 30 seconds the floor drops, and you have to kind of find your way around. And the goal is just to get the highest score in 90 seconds. But, Kevin, if you're playing it right now, can you put on the audio? Can you play the audio out of that?
- [19:59]
Kevin
Certainly. Yeah. Hang on.
- [20:01]
Gavin
I also asked it to make sure that it could give me a fantastic soundtrack. And so all of this, what you're hearing right now is generated by Claude with no input from me. The game graphics were generated by Claude. The physics were generated by 3js. All this stuff is part of it. But it is just remarkable what you can do now with these tools and a paragraph prompt. So Kevin's there desperately trying to make sure that he survives token burns. Yeah. So you go try it yourself right now and then let us know, like, go to our discord. Share with us anything you're doing on X or any of the other social networks. We'd love to see what you all have made with mythos 5. And we will keep going and try to figure out what's going to happen. Maybe we'll build some more stuff over the weekend.
- [20:46]
Kevin
Thank you all for the tokens.
- [20:48]
Gavin
Bye.