Latest / In the Loop AI / Claude Fable 5: The Most Impressive AI Model Ever

Transcript

Claude Fable 5: The Most Impressive AI Model Ever · In the Loop AI

0:00
Ejaaz: Four days ago, Anthropic asked the world's leading AI labs to slow down their
0:04
Ejaaz: AI research out of fear that the AI models were getting so good that they would escape human control.
0:09
Ejaaz: Now, just yesterday, the same company released the most powerful model the world
0:13
Ejaaz: has ever seen, but it comes with a twist. It's one model, but there's two versions of it.
0:18
Ejaaz: Claude Fable 5 is the model that everyone gets to use.
0:21
Ejaaz: It's a class that sits above Opus, but it's heavily restricted.
0:24
Ejaaz: It comes with a lot of safeguards around it because version 2,
0:28
Ejaaz: called Mythos 5 is the unrestricted version which poses itself as a higher cybersecurity
0:34
Ejaaz: risk than its predecessor, Mythos Preview. It also
0:39
Ejaaz: excels at creating biological compounds, which could potentially be used as
0:43
Ejaaz: bioweapons. So it's only accessible by vetted partners that Anthropic approves.
0:48
Ejaaz: So this brings us to a fork in the road where the smartest, most intelligent
0:53
Ejaaz: AI model isn't accessible to everyone.
0:56
Ejaaz: Typically, for the entire history of software, you'd be able to pay for access
0:59
Ejaaz: to the same level of access that an institutional would.
1:02
Ejaaz: But this time, the new metric isn't how smart is your AI model,
1:06
Ejaaz: it's the gap between how smart is your AM model that you get access to and the
1:09
Ejaaz: ones others also get access to.
1:11
Josh: Well, how amazing is it that we have Mythos now available in our apps today?
1:16
Josh: I think that's the really exciting takeaway is not only is it better than the
1:20
Josh: Mythos preview, but it's available inside of your app.
1:23
Josh: So if you are listening to this, you can go and try CloudFable 5 right now with
1:27
Josh: those two restrictions.
1:28
Josh: It won't go anywhere near biology. It won't go anywhere near cybersecurity,
1:31
Josh: but this is the best model in the world and it is now available to everyone
1:35
Josh: to go try out and it's like every once in a while there there's this new technology that
1:40
Josh: kind of forces you to to reframe how you engage with the technology and i think
1:43
Josh: that's been my experience so far using fable
1:46
Josh: is that it's so different than any other model we've used that it kind of forces
1:50
Josh: you to reframe how you engage with them and to showcase that i want us to go through some demos of
1:56
Josh: attempts that people have done to kind of showcase the powers and capabilities
2:00
Josh: of fable 5 this new frontier model starting with a really bizarre demo in which
2:05
Josh: uh dan shipper who is a prominent poster on x he recreated the library of babble
2:11
Josh: and help me understand what's going on here you just because i'm seeing a lot
2:14
Josh: of visuals um and i understand that it's not a image or video generating model
2:19
Josh: so how is it able to create a real world's,
2:22
Josh: emulation so accurate so quickly?
2:25
Ejaaz: So the greatest part about this is it took one prompt and he basically asked
2:31
Ejaaz: it to read a book, the book of Babel or whatever the title of the book is.
2:36
Ejaaz: And he said, I want you to read this book.
2:38
Ejaaz: And then I want you to recreate one of the concepts that is described throughout
2:41
Ejaaz: the book, which is this library of Babel or Babel.
2:45
Ejaaz: And it did so in just under an hour, I believe.
2:48
Ejaaz: And if you look visually on the screen, it is this high fidelity 3D representation
2:54
Ejaaz: of what this library looks like. And it's infinite. So you'll notice in this
2:57
Ejaaz: video that he looks down, he looks up, and it's just infinite books he can access himself.
3:03
Ejaaz: And he even asked it to include some of the essays that he's written himself.
3:06
Ejaaz: This guy authors a bunch of analysis on AI, and he pulls open a piece that he
3:11
Ejaaz: wrote that is in the Library of Babel. The idea of the library is that it contains
3:14
Ejaaz: every piece of text that has been ever written. And so he gets access to it.
3:19
Ejaaz: It's a pretty cool example.
3:20
Josh: Yeah, that's super cool. The other ones that I've really noticed that it accelerated
3:24
Josh: in is kind of 3D world building, which is funny because you don't think of Anthropic
3:29
Josh: as a world building model. It's not a world model.
3:32
Josh: In fact, it doesn't have image generation capabilities. In fact,
3:35
Josh: it doesn't have video generation capabilities.
3:36
Josh: So how is it creating all this realistic visual design assets?
3:40
Josh: And it's just really good at math. And this begs the question is how important
3:45
Josh: is it to focus on image gen on video gen if kind of at its core it could do
3:49
Josh: all of this with math and what we're seeing on screen here is how.
3:52
Josh: A virtual one-to-one recreation of Yosemite National Park that was done with
3:58
Josh: a simple prompt asking it to create a recreation of it.
4:01
Josh: And then the model was smart enough to go off and understand the context required
4:05
Josh: in order to build an accurate representation and pulled that off.
4:08
Josh: It did things like it scanned the satellite imagery to figure out what the elevations
4:13
Josh: were like. It pulled topographic maps that it found to figure out specifically what the heights were.
4:17
Josh: It found imagery that any imagery that it could find about the park so that
4:21
Josh: it can reference it and show you i mean look at this resuming in on a waterfall
4:24
Josh: it feels like it's a one-to-one replica,
4:27
Josh: lower fidelity but something that can run inside of your browser it's really
4:31
Josh: impressive how far the model can go on one prompt and i think that's one of
4:34
Josh: the places in which fable stands out in particular is its ability to to reason through
4:40
Josh: your requests in a way that hasn't been done before we filmed an episode yesterday
4:44
Josh: that i would highly recommend listening to about how we've kind of progressed,
4:49
Josh: with our interaction of the model kind of moving up the extraption layer,
4:52
Josh: where first we engage with models, then agents, then harnesses.
4:55
Josh: Now we're just creating these loops where the agent and the.
4:58
Josh: Underlying intelligence is smart
5:00
Josh: enough to actually understand what's required to get you to your goal.
5:03
Josh: And I think that's what this is such a great example of this Yosemite one in
5:06
Josh: particular is, hey, I want you to recreate Yosemite for me so that way I can
5:09
Josh: fly around and I can enjoy it in a one-to-one replica.
5:12
Josh: And it does all the rest for you. And I think that level of critical thinking
5:15
Josh: is something that's novel with Fable 5 that we've never seen in any other model before.
5:19
Ejaaz: I mean, the breakthrough that we're talking about is visual and spatial reasoning.
5:23
Ejaaz: And I think it's important to explain the difference between this and another
5:27
Ejaaz: favorite version of a model that we speak about a lot on this show, which is world models.
5:32
Ejaaz: Typically with world models, it recreates the physical world around us.
5:35
Ejaaz: But most importantly, it understands that physical reality, understands how
5:39
Ejaaz: gravity works, it understands how different forces of nature works,
5:43
Ejaaz: and it applies it when an object has an action. So let's say you kind of like
5:47
Ejaaz: punch a puddle of water, it splashes, the droplets come up.
5:50
Ejaaz: This isn't exactly the same thing. It's still based on theory.
5:54
Ejaaz: This is still an LLM that ingests a lot of text and understands kind of like
5:59
Ejaaz: how the physics works in theory and then recreates what its version of it might
6:03
Ejaaz: be. And this is what we're looking at on screen. It's kind of like,
6:05
Ejaaz: it's known as spatial reasoning or visual intelligence.
6:08
Ejaaz: It's close to the thing, but it's not quite the same thing. Now,
6:12
Ejaaz: another example that I really enjoyed was from Ethan Moloch.
6:16
Ejaaz: Ethan Moloch is one of my favorite AI researchers that analyzes a lot of these
6:20
Ejaaz: new models, but he builds or tests it in really interesting ways.
6:24
Ejaaz: One of these ways was he rebuilt Snake.
6:27
Ejaaz: Now, I am kind of ashamed to admit how much time I spent playing this particular
6:31
Ejaaz: game only because it was like the best version of Snake that we see.
6:34
Ejaaz: And like I'm showing you on the screen right now, it's like incredibly high
6:38
Ejaaz: fidelity. It looks way better than the game I used to play on the Nokia that I had as a kid.
6:43
Ejaaz: But the point is, it's pretty cool. It introduces new power-ups and obviously
6:46
Ejaaz: like you know you can die in usual things
6:48
Ejaaz: Also, forget about creating the game. Claude Fable 5 is really good at playing
6:54
Ejaaz: the game itself. What you're seeing is an accelerated version of it playing
6:58
Ejaaz: Pokemon Fire Red, and it completes the game in, I believe, 50 minutes.
7:02
Ejaaz: And the way that it works is it takes screenshots of the game at any single
7:07
Ejaaz: point, and it basically makes a decision as to which button it wants to click,
7:11
Ejaaz: which step it wants to take, and it is just kind of like spread through a software
7:15
Ejaaz: run where it's able to do that.
7:16
Ejaaz: Now, as a kid growing up and watching, you know, a lot of Pokemon,
7:20
Ejaaz: playing a lot of Pokemon, trading the cards, this is kind of nostalgic,
7:23
Ejaaz: but also kind of scary. We were joking before we started recording,
7:26
Ejaaz: Josh plays Cod quite a bit or plays a lot of computer games.
7:29
Ejaaz: And I wonder about the time in the near future where you're going to be one
7:33
Ejaaz: of the buying a AI agent and it might actually be better than you.
7:37
Josh: Yeah, that's going to be a little traumatic for me on a personal note.
7:40
Josh: Just hurting my ego that I'm losing my game to an AI.
7:44
Josh: I saw another great example on the YouTube channel, actually,
7:47
Josh: where they were Mythos or Fable 5 was playing Factorio.
7:50
Josh: And Factorio is a game that I really enjoy, that I've been playing for a long time.
7:54
Josh: And it was doing it very, very well. And I'm like, oh, dude,
7:56
Josh: you're getting a little too close to home with this. I don't love it.
7:59
Josh: But it's incredibly capable. And we have another example here that shows a use case that is not a game.
8:04
Josh: Instead, it is a it's so cool.
8:08
Josh: So there's this guy named Todd Saunders. He's on X and he he posts this tweet
8:12
Josh: saying fable slash mythos is unbelievable was on a customer call today and had
8:16
Josh: Claude transcribing in the background and on screen we're showing a visual of what that looks like.
8:21
Josh: As they were telling me about the features they wish their current software
8:23
Josh: had, Claude was building the features in real time.
8:27
Josh: By the end of the call, I was able to show a fully working product with the
8:30
Josh: exact workflow they mentioned 15 minutes earlier.
8:34
Josh: Autonomous looped building triggers from a customer call.
8:37
Josh: And this is one of the most amazing things about the model is that it's able
8:41
Josh: to go off and do a lot of the hard work yourself, where it feels like recently
8:45
Josh: you've had to kind of be in the loop. You had to continue to prompt the agent to give it more context.
8:50
Josh: And with this model, it's very easy to give it a goal and give it a verifiable
8:55
Josh: outcome that it can match against the goal.
8:57
Josh: And then it will just go off and do those things. So as is on the customer call,
9:00
Josh: like how cool is that for a salesperson where you're listening to customers,
9:03
Josh: you're listening to complaints and in real time, you're fixing their problems.
9:06
Josh: You're building new software on top of it.
9:08
Josh: It's unbelievably capable. This was one of the demos that I found most interesting too.
9:11
Josh: And then this final example, it's just fun for networking nerds or just computer
9:14
Josh: science nerds in general. We're seeing a highway on screen with cars and buses and vans.
9:19
Josh: And those cars are not random. They are actually associated with specific packets
9:22
Josh: that are being pushed across the network. So it's a really fun and interactive
9:26
Josh: way to visualize network traffic.
9:28
Josh: And I could imagine this being great for a lot of educational purposes.
9:32
Josh: And one of the things that I did notice also on Snake Ejaz is that there was audio, there was sound.
9:36
Josh: It actually generated sounds and audio too. So there's a lot of modalities in
9:40
Josh: which it's actually performing pretty well.
9:43
Ejaaz: Can I add a bonus Easter egg, which we didn't see on the demo,
9:46
Ejaaz: but I know because I played it too much yesterday.
9:48
Ejaaz: There are power-ups that pop up, but the power-ups are software patches.
9:54
Ejaaz: So if you don't hit the software patch power-up, you end up with a random error
9:59
Ejaaz: in the game that you must avoid, like a random wormhole that could like suck
10:04
Ejaaz: you out of the game and like cause you to lose.
10:05
Ejaaz: So like it's coding in real time but you could fix it in real time and if you
10:09
Ejaaz: hit the power up it it like implements the code fix like immediately it was
10:13
Ejaaz: just that's very nerdy but like very cool and very creative not something that
10:16
Ejaaz: we've seen before now if you want to like move away from the
10:19
Ejaaz: The toy aspect or the retail adoption of how you can use this particular model
10:24
Ejaaz: there's of course a lot of enterprise use cases and the number one version of
10:28
Ejaaz: enterprises using Claude is through code specifically and there were a few examples
10:32
Ejaaz: that were included in the official blog my favorite one was Stripe,
10:36
Ejaaz: who did a code migration of 50 million lines of Ruby, which typically would
10:40
Ejaaz: take around two months and several software engineering teams.
10:45
Ejaaz: It took them less than a day using Fable. I believe they used two to three instances
10:50
Ejaaz: of Fable, but still, that's like two to three software engineers that kind of
10:53
Ejaaz: worked night and day continuously to be able to achieve this.
10:57
Ejaaz: I saw another example of a company which had a software engineering team of
11:02
Ejaaz: Opus 4.8 working on projects that would typically take him two weeks.
11:06
Ejaaz: He can now do it in less than a day as well. So the point is there's a massive
11:09
Ejaaz: leap in intelligence for coding specifically with Fable, which is the publicly
11:14
Ejaaz: accessible model. Now, before we continue,
11:17
Ejaaz: We also wanted to like not just look at the demos that other people have recorded.
11:21
Ejaaz: We want to try our own. So we have a few prepared for you today.
11:25
Josh: Okay, so I just, I actually have no idea what you've been working on with these
11:27
Josh: demos. I know you are building a demo. Please share with the class what the
11:31
Josh: prompt was, what you're building and what the outputs of this thing are.
11:35
Ejaaz: For sure. Okay, so one of my favorite breakthroughs with this model is the visual
11:39
Ejaaz: and spatial intelligence that we referenced earlier on.
11:41
Ejaaz: So what I did was I found this hand sketch, this hand drawn version of a floor
11:46
Ejaaz: plan of a blueprint, which I'm showing you on the screen here.
11:49
Ejaaz: It gives you the layout of someone's home. It has a garage.
11:53
Ejaaz: It has bedrooms. It's kind of clunkily drawn. It's not really high fidelity.
11:57
Ejaaz: If an architect looked at this, they'd be like, this is probably physically inaccurate.
12:01
Ejaaz: And then I fed it to Fable and I said, listen, here's a photo of a hand-drawn floor plan.
12:06
Ejaaz: I want you to rebuild it as a single, clean, self-contained SVG that is architecturally
12:10
Ejaaz: accurate. I want you to improve it where you can, improve it in a way that can
12:15
Ejaaz: like, you know, reinforce the wall structures, et cetera, like really high detail things.
12:20
Ejaaz: And it produced this, what we're seeing on screen right here,
12:24
Ejaaz: which is this really high fidelity floor plan. You've got the garage.
12:27
Josh: Architectural grade.
12:28
Ejaaz: Yeah. You've got the surface area measured in meter squared.
12:31
Ejaaz: You've got the swing angle of each door.
12:35
Ejaaz: You've got the entire like kind of like layout of this entire thing,
12:37
Ejaaz: including potential mock furniture, which gave me a second idea,
12:42
Ejaaz: which was like, okay, I asked it, I want to purchase this sofa.
12:47
Ejaaz: These are the dimensions of this sofa.
12:50
Ejaaz: And I want to place it in the dining room or in the lounge. Can you tell me
12:53
Ejaaz: if I can feasibly do this in this full plan or will it get stuck?
12:56
Ejaaz: Like, how do I, how do I do this? Do we have enough doors? Do we have enough
12:59
Ejaaz: space to like maneuver it? Like how would this work?
13:02
Ejaaz: And it said verdict, yes, but not flat. You're going to need to pivot the sofa
13:07
Ejaaz: on its side and kind of like pull it in like vertically.
13:10
Ejaaz: And it gave me this really cool mock-up of how I would do it.
13:14
Ejaaz: It would gave me route A where I take it in through the front door.
13:17
Ejaaz: And this is kind of like the sofa that you see here, but I need to turn it on
13:20
Ejaaz: its side and kind of like shift it through this gap that I'm highlighting on
13:23
Ejaaz: the screen here in green.
13:24
Ejaaz: Or that there's route B where I can take it in from the outside and I have a
13:29
Ejaaz: two meter, very spacious door opening, which I can bring it straight into the
13:32
Ejaaz: lounge and place it right there. So it's really physically accurate going off
13:36
Ejaaz: the comments that we made earlier.
13:37
Ejaaz: It understands the kind of like reasoning behind physics really, really well.
13:42
Josh: It's such a great companion. And this is this gets back to what we talked about
13:45
Josh: earlier, which is like the most complicated and difficult part about this model
13:49
Josh: is figuring out how to engage with it, what to ask it, because it's so capable of doing these things.
13:54
Josh: We love making artifacts for the show as a way to share kind of the ideas that we're talking about.
13:58
Josh: And this might be a good time to get into the benchmarks of how good this model
14:02
Josh: actually is relative to other models. And I was looking at this post that I
14:08
Josh: saw on X showcasing in particular Claude Fable 5 versus GPT 5.5.
14:13
Josh: And my first reaction is, holy shit, that's a huge leap.
14:18
Josh: So Fable 5 low mode is scoring over 10%, whereas GPT 5.5 extra high is getting about 5.7%.
14:28
Josh: Now, this is on Frontier Code Benchmark. This is a specific particular coding
14:31
Josh: benchmark. This is not across the board, but it gives you a sense of how much
14:36
Josh: more powerful this model really is versus,
14:39
Josh: all the others second coding benchmark that i'd say this is probably one of
14:43
Josh: the gold standards this is what a lot of models will use to benchmark themselves against each other
14:47
Josh: um this is the swb bench pro and fable 5 scores 22 points higher than gpt 5.5 which was already,
14:57
Josh: what is that 11 points below opus 4.8 so it's currently looking like gemini
15:03
Josh: 3.1 pro gpt 5.5 opus 4.8 and then fable is running away with it and this seems
15:09
Josh: to be the case with almost all of these other.
15:11
Ejaaz: Benchmarks i have a i have a better one for you right so a lot of people listening
15:15
Ejaaz: to this might be thinking okay well i don't code i'm not a software engineer
15:18
Ejaaz: why does this apply to me well there's another benchmark called gdp
15:22
Ejaaz: vow which tests it against real world tasks that take human experts like knowledge
15:27
Ejaaz: workers that you know do back-end admin excel sheets all that kind of stuff
15:31
Ejaaz: hours to do and they compare it to the model
15:33
Ejaaz: take a look at this so fable mythos 5
15:38
Ejaaz: basically achieves the highest benchmark score it's actually almost completed
15:42
Ejaaz: the entire benchmark so they're probably going to have to recreate an entire
15:44
Ejaaz: new benchmark for this but basically what this means is probabilistically if you were to
15:50
Ejaaz: blind test or blind pick the output work of a expert human that is really good
15:55
Ejaaz: at a particular knowledge work task versus this particular model over 50% of
16:00
Ejaaz: the time, you're going to be picking this model, which is just an insane stat to see.
16:05
Josh: Yeah, it's pretty unbelievable. And you have to like, I'm looking at these charts,
16:07
Josh: and you have to ask yourself the question, as Anthropica saturating benchmark,
16:11
Josh: are they running away with it? Like, where is OpenAI in this conversation now?
16:17
Josh: I have to imagine that GBT 5.6 is coming soon.
16:20
Josh: Is it okay maybe it's better than opus 4.8 but can it can it eclipse fable no,
16:24
Josh: and i mean we know that anthropic released mythos months ago so you have to
16:29
Josh: assume that like they've been continuing progress and iterative development on
16:33
Josh: new frontier models that are even more powerful than this and and is this is
16:36
Josh: this beginning to become a runway or are they actually still competitive with
16:40
Josh: each other or maybe we just don't have enough information to tell we kind of
16:43
Josh: have to see what the response from open ai is.
16:45
Ejaaz: Well you look at the cadence between model releases right like what was the
16:48
Ejaaz: time since 4.8 was released. It was like less than, I think,
16:51
Ejaaz: 30 days ago. So the cadence is getting...
16:53
Josh: Yeah, we filmed an episode on this not too long ago.
16:54
Ejaaz: Yeah, like I remember that episode, right? And we spoke about it and went through
16:58
Ejaaz: its benchmarks back then.
16:59
Ejaaz: So the point is, these model releases are happening faster, but the capability
17:04
Ejaaz: gaps are even greater, which tells me one thing, which is we're getting closer
17:09
Ejaaz: and closer to the AI models just building itself.
17:12
Ejaaz: They haven't been private about this either. Anthropic has publicly claimed
17:15
Ejaaz: that they have been using Mythos Preview to build this new version that we're
17:19
Ejaaz: talking about today, Fable, right?
17:21
Ejaaz: So we think we've reached a point where you could maybe call it a breakaway
17:24
Ejaaz: from Anthropic, where they basically have recursive self-improvement almost
17:28
Ejaaz: achieved, where the model can do all the research, figure out its own issues,
17:32
Ejaaz: and build a better version of itself.
17:34
Ejaaz: Now, I do want to ground us at this point in this episode, Josh, which is,
17:40
Ejaaz: Fable 5 is an amazing model, but it's one version of the amazing model.
17:44
Ejaaz: There is another version of this model, which is technically better than Fable
17:49
Ejaaz: 5, but it is not publicly accessible.
17:51
Ejaaz: It is restricted because it poses itself as a cybersecurity risk,
17:56
Ejaaz: not just a cybersecurity risk, but also a bioweapons risk.
18:00
Ejaaz: It is so good at biology and chemistry that it could feasibly create compounds
18:06
Ejaaz: and a biological weapon that could pose a risk to any sort of nation state.
18:11
Ejaaz: And so for that reason, it is under heavy restrictions and safeguards in the
18:17
Ejaaz: version of Fable where you can't get access to any of this.
18:19
Ejaaz: And only vetted partners and cleared government security initiatives are able
18:24
Ejaaz: to get access to this Mythos 5 thing.
18:26
Ejaaz: Now i put this to the test josh um and i i did a very simple example which was
18:33
Ejaaz: um can you explain how the mitochondria works do you want to bet what its answer was
18:39
Josh: Oh i'm gonna be best it's not touching that it's not touching biology yeah i
18:42
Josh: want to take a second to actually explain the nuances between the models because
18:46
Josh: when you say the word better i'm not sure it's better it's just more complete
18:49
Josh: one model is a complete model one model is a heavily restricted model and in the case of Mythos,
18:55
Josh: it's available for cybersecurity.
18:57
Josh: It's available for biology. And that's what we've seen with Project Glasswing,
19:00
Josh: where they're working privately with companies to kind of fix security vulnerabilities and figure out bio.
19:05
Josh: And in the case of the system card, I saw that it's accelerating some bio experiments
19:09
Josh: at a full order of magnitude, 10 times better. So it's really capable there.
19:14
Josh: The compromise that we had to make in order to receive it was it can't touch
19:19
Josh: bio, it can't touch cyber.
19:21
Josh: So it's just as capable everywhere else. It will not do that.
19:23
Josh: What happens is if you ask in the case like you did, how does mitochondria work?
19:27
Josh: It will route through Opus 4.8 for that answer and then come back and give you a response.
19:33
Josh: So it is as capable everywhere. It's just don't ask about bio,
19:36
Josh: don't ask about cyber because from my experience so far trying it and EGS, it seems like yours.
19:41
Josh: Anytime you get remotely close to those topics, it is just completely shut down,
19:45
Josh: routed through Opus 4.8 instead.
19:47
Ejaaz: Yeah, I think it's it's too aggressive, personally, like, as a former science
19:52
Ejaaz: nerd, I still spend a lot of time trying to digest like some of the latest scientific advancements.
19:57
Ejaaz: And like, listen, I'm not reading research papers. So I work with my best pal
20:00
Ejaaz: Claude to try and figure out, you know, what the latest takeaways are.
20:03
Ejaaz: Now, typically, I could slam that into Opus 4.8. And it would give me an amazing
20:07
Ejaaz: summary. And I could like ask it questions.
20:09
Ejaaz: Now, if I want to use Fable 5, it just simply won't read the paper.
20:12
Ejaaz: As soon as it sees anything related to chemistry or biology,
20:15
Ejaaz: it switches off and reroutes to 4.8.
20:17
Ejaaz: So I am not able to get access to the frontier LLM intelligence or brain that
20:25
Ejaaz: Fable 5 has for me, mythos, even though, you know, my intention isn't to build
20:28
Ejaaz: a bio-weapon by any means, I can't get that analysis.
20:32
Ejaaz: And so that's one version of it, right? Where like it is too heavily restricted.
20:36
Ejaaz: The other version of this is with the more intelligent models get,
20:40
Ejaaz: it's not just going to be super intelligent in one particular vertical.
20:43
Ejaaz: Like for us, it's like, you know, research and creating artifacts and the best content.
20:47
Ejaaz: It should also apply to any other profession, right? Whether you're a scientist,
20:51
Ejaaz: whether you're a mathematician, and whether you are building different kinds
20:55
Ejaaz: of structures or whatever it might be.
20:56
Ejaaz: The fact that it can get triggered so easily or the fact that Anthropic has
21:00
Ejaaz: very heavily restricted that capable intelligence in a way that like even people
21:06
Ejaaz: that have well intentions can't get access to it.
21:08
Ejaaz: In my opinion, is a bit of an issue. And listen, it's V1. I'm sure they're going
21:12
Ejaaz: to like release a bunch of versions of the safeguards where it like makes it a lot easier to use.
21:18
Ejaaz: But for V1, it's kind of like, I think it's overdone. It's important for people,
21:22
Ejaaz: I think, to understand how these safety classifiers work as well.
21:25
Ejaaz: Think of Claude Mythos 5 having an AI model or system that is watching it.
21:32
Ejaaz: And as soon as one of the red flags that it's been trained on is triggered,
21:35
Ejaaz: for example, anything to do with biology or chemistry, it gets switched off
21:39
Ejaaz: immediately and rerouted to 4.8.
21:42
Ejaaz: Now, there's four particular categories that Fable can't get access to.
21:46
Ejaaz: It is cybersecurity, for biology, for chemistry, and for distillation as well.
21:52
Ejaaz: And this is a key one which caused a lot of contention in the public ecosystem
21:56
Ejaaz: when they launched yesterday, which is if you were to ask about model training
22:01
Ejaaz: techniques or even just simple general questions around, hey,
22:04
Ejaaz: I have this AI agent, it's pretty clunky,
22:07
Ejaaz: how can I improve its harness to kind of make it go quicker or use less tokens?
22:12
Ejaaz: Automatically degrades performance. And this is the key change in this fourth
22:17
Ejaaz: category with distillation.
22:18
Ejaaz: Anything that Anthropic considers to be trying to derive its model to build
22:23
Ejaaz: another model, it gives you intentional poor performance. And it doesn't even tell you.
22:29
Ejaaz: Now, it says that this happens for 0.3% of cases, but my guess is it's probably
22:34
Ejaaz: happening for higher reasons. And listen, it's completely within Anthropic's
22:37
Ejaaz: right to do this. I get it. I understand it. You want to remain competitive.
22:41
Ejaaz: But it's just interesting to see when like you have this intelligence model
22:44
Ejaaz: where, you know, it's meant to kind of like blossom and create and help other
22:48
Ejaaz: people build different things.
22:49
Ejaaz: But they're being competitive when it comes to other models, I guess.
22:53
Josh: Well, we're getting to this unique intersection where like, they have mythos,
22:56
Josh: they've had mythos for a little while, and they decided to keep it private.
22:59
Josh: And that was okay, and somewhat understood because it was really discovering
23:04
Josh: a lot of zero day vulnerabilities.
23:06
Josh: And it seems like they worked pretty hard to figure out a way to not only improve
23:10
Josh: the quality of the model but actually make it public and i guess like the question
23:13
Josh: we're gonna have to start asking as these ai labs continue to create these like unbelievably
23:18
Josh: uh forward-looking frontier models is like to what capacity are we just happy
23:22
Josh: to have them like how much should we expect out of the labs when it comes to
23:26
Josh: delivering these models like in my case i'm pretty stoked to be able to use fable 5,
23:30
Josh: And I'm not interested in biology. I'm not interested in distilling the model.
23:35
Josh: I'm just like pretty stoked to do my day-to-day work with this capable model.
23:38
Josh: And in that sense, it's really fun and exciting and interesting.
23:41
Josh: And I think it's the start of a longer conversation.
23:43
Josh: We saw some legislation come in a few weeks ago, last week maybe,
23:48
Josh: about requiring AI Frontier Labs to kind of showcase the models privately with
23:53
Josh: the government to share what's coming down the line.
23:54
Josh: And I guess in this essence, I'm more excited to have the model versus not have
24:00
Josh: the model and have it have these constraints in the hope that it will slowly become
24:05
Josh: unwounded as they kind of improve and iterate on the quality and the kind of
24:09
Josh: like security set of this model.
24:12
Ejaaz: Yeah, listen, I think Anthropic is ultimately doing the right thing.
24:16
Ejaaz: I think that they can't just kind of diffuse this model to anyone and everyone
24:19
Ejaaz: because malicious actors, however few they might be, will actually end up doing
24:23
Ejaaz: something dangerous with this.
24:25
Ejaaz: That being said, I think the subjectivity and who gets to govern that subjectivity is important.
24:32
Ejaaz: Like, I can imagine a future version of an Anthropic model that isn't just necessarily
24:37
Ejaaz: really good at biology or cybersecurity.
24:40
Ejaaz: It might be really good at something such as trading, right,
24:44
Ejaaz: for example. And then the question becomes, who gets access to this trading
24:48
Ejaaz: model that is so good that it could break the stock market?
24:52
Ejaaz: And maybe if you are Citadel, who are closely aligned with Anthropica,
24:56
Ejaaz: I'm theorizing here, then they get access to it, but Jane Street won't get access to it.
25:00
Ejaaz: And so it becomes this heavily-based nuance that is only dictated by maybe the
25:05
Ejaaz: government and maybe Anthropica itself.
25:07
Ejaaz: There was talks around like Trump taking a stake in some of these AI labs to
25:11
Ejaaz: nationalize it for this exact reason, because it could pose a threat and they
25:14
Ejaaz: want to have governance decisions. It just gets a little murky and messy.
25:17
Ejaaz: And I think we're at the fork in the road.
25:19
Ejaaz: There's no going back at this point. We are now entering a phase where
25:24
Ejaaz: The model that you have access to may not be the most intelligent model for
25:27
Ejaaz: the specific thing. And listen, it may not be the thing that you necessarily
25:31
Ejaaz: do on a day-to-day, but it's a lot of things that other people do day-to-day,
25:34
Ejaaz: and they want to get access to this model. How that is governed, I don't know.
25:37
Ejaaz: The other restriction, which I found really interesting that I noticed in the
25:41
Ejaaz: footnotes of their system card or announcement blog post is on June 22nd,
25:48
Ejaaz: we ceased to get access to Fable 5.
25:51
Ejaaz: Now i think this was taken massively out of context because i think the reasoning behind this is
25:56
Ejaaz: it's because it depends on availability of compute so if by june 22nd anthropic
26:01
Ejaaz: has more available compute to distribute to users
26:03
Ejaaz: then it wouldn't be the case but on the case that it is it would shift to a
26:07
Ejaaz: pay-per-usage model which means that you buy credits and if your credits are
26:11
Ejaaz: consumed you then need to buy more credits kind of like the api model is that right
26:16
Josh: Yeah, according to the blog post, it says from today through June 22nd,
26:19
Josh: Fable 5 is included on Pro, Max, Team, and Seat-based enterprise plans at no extra cost.
26:23
Josh: On June 23rd, we'll remove Fable 5 from those plans. Using it after that will require usage credit.
26:28
Josh: If capacity allows, we'll extend the included window. After this point,
26:32
Josh: when sufficient capacity allows us to do so, we aim to restore Fable 5 as a
26:36
Josh: standard part of subscription plans.
26:38
Josh: We intend to do this as quickly as we can.
26:41
Josh: And yeah, it sounds like Fable 5 consumes a lot of compute. When you load it up inside of the,
26:46
Josh: app it says fable is the most capable model and draws down usage twice as fast
26:50
Josh: as opus so you have to imagine
26:52
Josh: that it consumes a lot of gpus they're clearly using those gpus for a lot of
26:56
Josh: things i think the idea is to give a preview and then extend that for as long
27:00
Josh: as possible or just continue to extend it perpetually based on compute
27:03
Josh: i think to your earlier point we're very much at a fork in the road
27:07
Josh: when it comes to these models being capable enough to
27:11
Josh: really make a meaningful impact in the world and we've spoken so much about
27:14
Josh: alignment and ai safety and it's kind of been this like open-ended fuzzy thing
27:19
Josh: where it hasn't really practically applied to anything that's happened before
27:23
Josh: and we're finally at a moment in time in which the models are becoming capable
27:26
Josh: enough to have that conversation about
27:28
Josh: ai alignment ai safety you're starting to see why a lot of the teams are taking
27:32
Josh: it so seriously because it is the singular question is like answering what you
27:36
Josh: just said who gets access to this model how is it going to be restricted who
27:39
Josh: gets to decide that and that's why the alignment and safety conversation.
27:43
Josh: Is so important. And while you start to see a lot of the company cultures within
27:47
Josh: these companies align around these different priority sets that separate them from each other.
27:52
Josh: So this is, it's, it's a new day. It's a new era today.
27:56
Josh: We are moving into the, the next frontier of models.
28:00
Josh: It was pushed forward a considerable amount and in a way that I don't think
28:03
Josh: we've experienced in quite a long time. And it's really exciting to see.
28:06
Josh: I'm very excited to play with Fable, spend some time kind of generating outputs,
28:10
Josh: figuring out what it's most capable in that could help us in the day-to-day.
28:14
Josh: Like for me, if they never told me it wasn't gonna do bio or cyber,
28:18
Josh: I'm not sure I'd ever come across it because that's not really within the realm of uses that I have.
28:23
Josh: So I'm excited to just kind of play with it and figure out best use cases for
28:26
Josh: this. In terms of pricing, what I found really interesting is CloudFable 5 is
28:30
Josh: only twice the price of GPT 5.5.
28:34
Josh: I believe it's $10 per million input tokens, $50 per million out, and,
28:40
Josh: gpt 5.5 is five dollars and thirty dollars out so pretty close and much more
28:46
Josh: capable so if the case that it does get removed from subscriptions it is still
28:50
Josh: available from api it is not
28:52
Josh: as expensive as i think a lot of people thought this is significantly cheaper
28:55
Josh: than what i believe mythos preview was early on.
28:58
Ejaaz: Yeah it's my new favorite model i've been using it relentlessly for the last
29:02
Ejaaz: 10 hours um one thing that
29:05
Ejaaz: Uh is a really strong capability that it has is long horizon tasks like this
29:09
Ejaaz: model is engineered from the ground up to be able to work like a dog for six
29:15
Ejaaz: to 12 hours at a time on whatever project that you have.
29:18
Ejaaz: And it has this loop function, which basically says, if you come across a problem,
29:21
Ejaaz: don't ask me, try and figure out yourself and do the thing, build the thing.
29:25
Ejaaz: That's why we had people build world engines from scratch that we demoed earlier
29:29
Ejaaz: and these games from scratch, all from a single prompt, the library of Babel.
29:33
Ejaaz: So if you have an idea or if you've been pondering on a project that you've
29:37
Ejaaz: been putting on for a while, because you're like, I know I could probably vibe
29:40
Ejaaz: code this, but I don't want to spend like an hour doing this.
29:42
Ejaaz: Now you just need to write one detailed prompt and you should be able to do
29:46
Ejaaz: this. So my prompt for the listeners of the show as we wrap up this episode
29:50
Ejaaz: is get access to this model.
29:53
Ejaaz: Try it out. And I'm curious what your thoughts are on it. Do you think the restrictions
29:57
Ejaaz: affect you specifically?
29:59
Ejaaz: Or do you think it's a really good general purpose model and you're happy with
30:03
Ejaaz: how it's presented itself. You don't care about Mythos 5 in effect.
30:07
Ejaaz: And also, what other projects are you going to be doing with this?
30:09
Ejaaz: Are there any kind of demos or use cases that we haven't covered that might
30:13
Ejaaz: be specific to you in your leisure or your particular work that you might want to apply this to?
30:17
Ejaaz: Let us know the feedback in the comments to this video or DM us on X.
30:22
Ejaaz: Our profiles are linked below.
30:23
Ejaaz: We want to hear back from you. But I think that brings us to the end of this
30:28
Ejaaz: episode. We have now like a new world-leading model?
30:32
Ejaaz: OpenAI is going to have to answer to this, but it seems like Anthropic is running
30:35
Ejaaz: away with it. Josh, any final thoughts?
30:38
Josh: That's it. This is a new, it's a new era today. Like, I feel like we should celebrate.
30:42
Josh: This is a new frontier that has been pushed forward very far in an industry
30:45
Josh: we care very deeply about. So it's exciting to see. I'm stoked to use it.
30:48
Josh: I'm curious to hear what the best types of prompts or use cases are that anyone
30:53
Josh: who's listening has found.
30:54
Josh: If you enjoyed this episode, don't forget to share it with a friend who might
30:57
Josh: also want to try Cloud Fable 5 and experiment and get their feedback on how
31:01
Josh: it's being used to improve their life, improve productivity,
31:03
Josh: whatever use cases it may be. But as always, thank you all so much for watching
31:07
Josh: and we will see you guys in the next one.