Anthropic Is Hiding Its Best AI: Fable 5.2 and the IPO

Anthropic just released Claude Fable 5.2, the new number one AI model in the world, and its own system card says it isn't Anthropic's most powerful model. Ejaaz breaks down what Fable 5.2 can actually do, from protein design and robots to remote work, AAA games built in minutes, visuals made entirely of code, a business plan from one prompt and an operating system built by 1,000 agents, and why Anthropic deliberately held back the model's ability to speed up AI research.

We also cover the strategy behind the release: the growing gap between public and internal models, the METR review, the tit-for-tat race with OpenAI and GPT-6.2 Astra, and the new IPO terms (825 million shares at $70 to $78) as Anthropic heads into IPO week with the best model in the world.

==
FOLLOW EJAAZ ON TWITTER:
https://x.com/cryptopunk7213

==
Timestamps
00:00 Anthropic Is Hiding Its Best AI
01:10 Why Release It Now?
03:35 How We Got Here
05:08 The Benchmarks
05:48 Built for Medicine
07:17 Robots
08:38 Remote Work
09:23 AAA Games in Minutes
10:50 Visuals Made of Code
13:00 A Business From One Prompt
14:36 An OS From Scratch
15:44 Why Hide the Best Model?
20:09 Baiting OpenAI
21:30 Racing to Self-Improvement
22:32 The IPO Details
24:30 Final Thoughts

#ai #anthropic #claude #fable52 #ipo

==
AlphaSpace by Yahoo Finance: A Professional-Grade Investment Platform Built for Everyday Investors.
👉 https://finance.yahoo.com/about/promos/gold/alphaspace/?ncid=100003571

Creators and Guests

Host

Ejaaz Ahamadeen

What is In the Loop AI?

Exploring the frontiers of Technology and AI

It's official. Anthropic has once again unseated OpenAI to become the number

one model lab in the world with the release of not one, but two brand new models,

Fable and Mythos 5.2. Now, Fable 5.2 is a model that's generally available to you and I.

And for all intents and purposes, it's a fantastic model. It's amazing and number

one at coding, general reasoning, and visual creation.

But the craziest part about this is it's not even Anthropic's most powerful model.

Their most powerful model is actually being hidden behind the doors and has

been given a codename, Model 4.

And you might be asking, well, why hasn't Anthropic released this?

Well, that points towards a certain type of strategy that both Anthropic and

OpenAid, the two top AI labs in the world, are employing ahead of some of the

biggest IPOs in the world.

Now for Anthropic, that IPO is looking to be in the next six to 12 days

with new information that was released yesterday that they are targeting around

a $1 to $1.3 trillion valuation, which is down from $2 trillion,

dollars, which is much healthier in my opinion.

But that's off the back of still insane demand to the tune of where they hit

$80 billion of ARR in the last month.

These exponentials are getting even steeper.

But then you might be asking, okay, why release a model that's only slightly

better than your previous models?

If I had to answer that, it's because of two reasons. They have getting into

a bit of a competition with OpenAI, making sure that they always leapfrog ahead

of each other. So that's why they would keep a more powerful model behind the scenes.

But number two, maybe a larger point is we're seeing an increasing gap between

the large labs of public models and hidden models.

And that gap is getting larger and larger and larger.

And that might be the two reasons. One, the models that are internal are,

albeit more powerful, but also more misaligned, which means that if they were

generally available to everyone, that could be used for misaligned purposes.

They need to figure out a way to align the models before they can.

Confidently put them out. But number two, and this is arguably the more sinister

reason, if you're open air or if you're anthropic and you have a model that's

capable of solving some of the biggest problems in the world.

Wouldn't you just keep that to yourself to solve those problems and then capture

those markets yourself?

I'm talking like things like science, robotics, solving medicine, curing cancer.

These are all potential markets that are going to become increasingly available

to two, maybe three specific companies as they absorb all the markets around

them, as AI consumes all these different markets around it.

So you can imagine they would want to start gatekeeping some of these models.

And listen, with Fable 5.2, you're starting to see the emergence of this capability.

It's not just about AI models being really good chatbots and LLMs,

like the majority of people use AI models today.

It's about these models doing things outside of that, solving really complex

problems in mathematics, physics, and chemistry, which forms the bedrock of

solving things like cancer and some of the biggest ailments in the world,

but also applying that to anything and everything.

Imagine if you could build a better aeroplane, or if you could build a completely

new vehicle from scratch, or brand new devices that can improve your life around you.

Now, and this is part of a larger trend that we're seeing with AI models stepping

outside of the computer and interacting with the physical world and more generally

available sectors in a huge way.

So on this episode, we're going to get into the fable 5.2, we're going to get

into the brand new IPO details, as well as what the new model releases are telling

us about the general strategy that Anthropic and OpenAI are employing going

forward. So all of this on today's episode brought to you by Qualcomm.

Now, just to remind you how we got here, this all started literally nine months

ago at the start of the year when Anthropic teased their new powerful model.

And remember, it was Mitho's Cyber, I think, and it was a huge security risk.

And then they released it.

And then they shut it down two weeks later because the government told them

to because they were scared that it was going to go out and hack everything.

And then seven months later, we realized, oh, yep, that's exactly what these

models were capable of. It was shown that Anthropic and OpenAI models at the

time had escaped and broken out.

That's what resulted in things like the Hugging Face incident and the crazy

Black October cybersecurity hacks that we've seen over the last month.

And the response from this, probably heralded by Dario Modi himself,

was we need to pace the frontier. And to be fair, he had the support of all

the other AI leaders as well, Sam Altman, Demis Hacibis.

And the idea here was we need to slow down AI research and development because

our internal research shows that these models are getting too powerful,

too smart, to the point that they could be misaligned in the work that they do.

So whilst they tell human researchers, hey, we're doing this,

we're acting in good faith secretly, if you read some of their internal monologue,

they're actually doing other things. And that's what's led to a lot of the hacks.

So it seems like a pretty fair request.

Now, Fable 5.2 is the first obvious model release that is both the most powerful

in the world, number one in the world by all benchmarks, stats, and usage.

But is also being throttled. And I'll get to more of that later.

But first, let's get into the actual model. What's it good for? Should you switch to it?

Demos. Now, at this point, you and I are tired of hearing benchmarks.

Benchmarks just kind of grade these models like a natural exam,

but it doesn't actually translate into how you and I use the model.

And it's hilarious. My producer, before starting this episode,

he goes, yeah, I kind of see them as nutritional facts or like you're going

to eat your vegetables. So right now, let me just serve you for two minutes

the vegetables of this new model.

I want to point you actually to three specific benchmarks. So overall,

this model is number one on pretty much everything. And that doesn't surprise

me because Anthropic has had the lead for a while, Anthropic and OpenAI.

But three specific benchmarks point to why this model is actually quite unique.

Number one, if you look at the protein design benchmark, it scores the highest

overall every single model.

Now, the reason why this is important is Anthropic has recently,

over the last, I want to say.

Six months, invested tens to hundreds of billions of dollars on acquiring biological

and scientific data and training models specifically to solve scientific problems.

Now, OpenAI, their main competitor, has also been doing the same,

but focused on maths and physics.

Anthropic has been obsessed with biology. They proved that this model is really

good at designing and predicting the structure of something known as capsids.

And what their research and demos have shown for this particular model is that

it's really good at all the development processes around creating and delivering medicine.

And this is very intentful. I think

Dario Modi over the last year actually suffered the death of his father.

And he said quite publicly that he could have seen a world in two to three years

from now where his dad's ailment could actually have been cured and his father could have been saved.

And so it's a personal focus on.

For him to make sure that AI accelerates meaningful industries,

in this particular case, medicine.

And this is the first step that we're seeing with a model where it's actually

being used by scientists and researchers to help accelerate their process of

creating medicinal drugs that can solve some of the biggest cancers and diseases in the world.

Now, the second benchmark is robotics.

Now, remember, it's one thing having an AI model that can create potentially the cure for cancer.

It's a complete other world to actually go out physically, set up all the right

equipment, dilute the right types of concoctions and chemicals,

and actually create the cure.

And robotics, remember, doesn't just apply to science. It applies to anything

and everything out there, manufacturing, manual labor, whatever that might be.

And so connecting AI models to really good robotic arms and machinery has honestly

been a major struggle. There's so many things that get in the way,

especially around understanding the physics of the world.

This model, Fable 5.2, is fantastic at doing that.

And I say all of this not to say that you and I should go out and start creating

robots or experimenting with drug development, but more so to point out that

Anthropik's focus when it comes to creating models isn't just on general intelligence anymore.

Indeed, 5.2 is exceptional at coding.

It's the new number one. It's really good at general reasoning.

It's really good at solving any and all of your problems.

But to be honest, it's only really incremental in those aspects.

You could still use Opus 5.5, which is their previous most powerful model for the majority of tasks.

In fact, you can argue that you should be using both of these models hand in hand.

But the real unlock, the real power for this model comes outside of the chatbot LLM.

Now, before we get into the demos, the third thing I want to point out is when

it comes to remote labor. So we're talking about things that are only done on

a computer that you could particularly consider as very niche or hard tasks.

This model is exceptional. Now, the previous highest percentage that could have

been scored on this, I think, was 25%.

This achieved 41%. Now, putting the scores aside, why this is important is the

measure of the impact of these models can only really be gauged by real-life, real-world tasks.

The fact that this model is now performing really well on remote labor indexes

is a really interesting point and indicates that these models are now becoming

good enough to replace quite a large amount of work out there for people.

Now, getting into the demos of this model and what it can actually do,

I've seen two main things. Number one.

People are going ham on creating triple A rated computer games.

Seems to be a lot of gamers out there.

And listen, before you switch off, I know you and I may not be playing games

all the time, or maybe you are, but when it comes to this industry of gaming

and creating something that would be, I don't know, the latest Call of Duty or GTA,

you could now do so, and I'm not exaggerating.

In 30 minutes, what takes game developers tens to hundreds of millions of dollars and decades to do.

I mean, just take GTA as an example, we still haven't seen GTA 5.

The idea here is AI has accelerated the development of really complex,

really visual types of mediums.

And Fable 5.2 demonstrates that really well when it comes to computer gaming.

Now, if you are a games company, right?

Let's take Epic Games, for example, which has created some of the biggest hit games in the world.

The CEO and founder, Tim Sweeney, recently said that he is now replacing a ton.

Of staff-related roles with AI because it's the only way to,

one, accelerate progress in the gaming world because AI models are now good

enough, and two, a really good way to unlock new potentials and capabilities.

So if you're Tim Sweeney or a game developer or even just a regular average

college kid that wants to create a game or you've had an idea about a game,

you can now do so with a model like this.

Okay, but what if you're like me and you don't really care too much about these

things? Well, there's a number of other applications.

Now, firstly, the visual generating capability of this model is insane.

And it's insane for two reasons. One, if you look at this example from Claude,

the official account, they're doing an educational video.

And this generates a really in-depth view of how a cell is created.

Now, could you imagine if you were back in high school or college and you had

access to this thing, I would probably have scored way higher than I actually

did and not flunk out the class, right?

So the idea here is it can completely transform the way your kids or yourself

learn brand new concepts.

But the second crazy part about this is, remember, Claude and Anthropic haven't

invested any money, zero dollars, in image and video models.

All of this, what you're seeing in front of you, is generated by code.

And what's baffling to me about this is.

Not too long ago, around four to maybe six months ago, I was convinced that

you couldn't have a coding model or an AI model that relied solely on code good

enough to beat a video and image model.

And yet what I'm staring at in front of me is pretty damn impressive.

If you look at the games, or if you look at this educational videos,

we are now in this weird world where code kind of solves everything.

And so if I step back a year and I look at Anthropik's humongous and very risky

at the time bet on going all in on building not just the best general model,

but the best coding AI model

and taking the bet that that is going to eat anything and everything around

the world, software related at least.

They made the right bet and it's paying off pretty massively.

And I've seen people so far use this function for really cool things that aren't

just educational videos.

This guy, for example, is using it to track all of his AI agent movements in a single visual video.

So what you're seeing on the screen here is a live update of all his different

instances of Claude or other open models and what they're getting up to,

tasks they're working on, when they're complete and when they're waiting for another update.

This is a very novel way of kind of visualizing your core work process so that

you're not just staring or having to move between several different conversations in your side chat.

It's pretty cool. Now, for the business folk out there, I know that there are

a lot of entrepreneurs out there that are trying to figure out how best to kind

of work with AI models or the latest and greatest and whether one is better than the other.

There was this really interesting case study that Ethan Mollock actually got

up to where he tasked Fable 5.2 with the following. He said,

I want you to create a business, any business, kind of go out there and research

and come up with the best idea.

But here are the stipulations. I want you to generate at least $20,000 a year,

and I would like you to take no more than 20 minutes of my time per month.

So in the entire month, I just want to have to pay attention to this for 20

minutes, and I want you to generate me $20,000 by the end of the year.

So Fable 5.2 came up with 22 ideas. It screened through all of them and settled

on one. A service which told equipment sales reps when park and water projects

were being designed or funded publicly before they went up to bid.

So the idea here is the AI model would scan publicly available records.

You know, there's arduous terms and agreements, numbers of different PDFs,

which no sane person has enough time to do. It would normally consume hundreds of hours.

It would do it autonomously 24-7 and discover when these projects go live before

they officially go live for bid.

So the idea here would be these equipment and sales reps submit a bid or submit

their application before everyone else potentially win the bid and then earn money that way.

Now, Ethan, pretty sensibly, in my opinion, decided not to pursue this any further.

But it's cool to see the new ways that these models are thinking when it comes

to business. And finally, for the software engineers out there,

yes, you have a new number one coding model, and it's very impressive in a very

distinctive way, which is.

It's now no longer just as good at one-shotting basic software apps,

programs, or whatever that might be.

Now, it can rebuild an entire operating system in one go.

Now, to kind of put this into context, operating systems, I'm talking about

like iOS, talking about Microsoft Windows or the latest suite, right?

The fact that you can now not only reverse engineer pretty much any software

out there, if you're, for example, Adobe, Microsoft Excel, or whatever that

might do in a few minutes,

but you could also rebuild the actual system that it resides on from scratch

means that you could theoretically

have a single individual build a multi-billion dollar company in a matter of hours.

Structurally, at least, the demand, the customers, the marketing,

obviously, are other things that can be solved, maybe honestly,

with five to 10 more Claude Fable 5.2 agents. But I digress.

The idea here is it's becoming incredibly easier for a single lone individual

to generate a lot economic value in the world.

Okay, so we've established this is a great model, but why then is Anthropic

gatekeeping a more powerful model behind the scenes?

You see, most recently there's been a few reports, the main one being from this

organization called METR, which was given very privileged access to spend time

in Anthropics labs, read all their research, experiment with the new internal models to.

Gauge whether these models are getting out of hand too powerful or whether,

Anthropic is misusing it in some way, shape, or form. And this prompt came from

all the scrutiny that Anthropic and a lot of other AI labs were facing that

they were gatekeeping some of the most powerful models and not allowing the

public to actually see what was going on under the hood.

So Anthropic invited external parties to come in. METR was one of them.

And they released last week one of their most important reports,

which suggested two very important things.

One, that the research and development progress inside Anthropic had become

so fast that it was accelerating so quickly that now you get 2x the leap in

model development in a single year than we did two years ago.

And this is increasing exponentially to the point where by the end of next year,

we're probably going to have a 5x increase.

And so the implied outcome here is that models will achieve what is being codenamed

as superintelligence or ASI, artificial super intelligence pretty soon in the

next, say, nine to 18 months.

And so if you compare that to Dario's whole point around we need to pace the

frontier because these models are getting out of hand, you start to question,

well, what's your actual motive here?

Are you just instead saying you're pacing the frontier, but actually continuing

to develop the most powerful models behind the scenes?

That's certainly what it looks like. Because if you look at Fable 5.2,

and if you look at the fine print of Anthropik's release, you'll notice that

they say, we have intentionally held back the capability of this model

to build and advance AI development.

And so what they're saying here is this publicly available model that we've

given to you cannot be used to really improve models any more meaningfully than

Opus 5.5 or any of our predecessor models could do so, whilst having an internal

model codenamed model 4 that can do this to the tune of 2x right now and maybe

5x by the end of next year. So the idea here is.

Gap between the most powerful model and what is publicly available is growing much, much larger.

My bet is we're going to see this continually increase for two reasons.

The public one is Anthropic and OpenArea saying, listen, we need to take more

time to make sure that these models are actually safe and aligned before we

publicly release them. And honestly, I think that's a good argument.

But the other more sinister argument that you could make is that this is so

that they can capture the value of what these more powerful models can do for themselves.

And it's resulted in a lot of negative market reactions. I mean,

look at when Anthropic releases a new cybersecurity update that says like,

hey, our models are really good at securing your systems. You see cybersecurity

stocks absolutely dump.

You can imagine if you have an internal cloud model that solves cancer in a

meaningful way over the next couple of months, that could result in a lot of

pharmaceutical companies also tagging. So we're in this weird world where.

Two, maybe three companies, Anthropic, OpenAI, maybe Google at this point,

have a lot of power concentrated within them.

And two-thirds of those companies, Anthropic and OpenAI, have only existed for

like three to five years.

And they're still private companies. They're not public yet.

And they're on this insane trajectory. I was chatting to producer Luke,

and we were talking about the rise and impact of technology over time.

And if you kind of think about the days where we were kind of exchanging goods

as a currency to the point where we had the internet boom, right?

And the internet boom, during that time, people were basically saying,

listen, like, this is unparalleled growth. What we've seen over the last five

years in revenue growth, we would have taken centuries in the production of

the railway, whatever that might be.

And I was joking with them, I was saying, that's what we're seeing in AI here,

but it's happening over weeks and months.

And we're comparing it to the dotcom boom being like, hey, we had to wait a

decade to get a meta or Google type company.

And now we're probably going to see it being spun up overnight because we're

at this insane acceleration.

We're at the steepness of the curve where it pretty much looks indistinguishable

to being vertical. But there's also a more strategic reason which I think Anthropic

is employing to compete directly with OpenAI, which I'll get into on today's

episode brought to you by Qualcomm.

And that tactic is they're kind of playing a tit-for-tat game with openai

openai has kind of also teased that they have this insane internal model and

it's starting to solve some crazy problems like navier stokes which is one of

the hardest mathematical problems out there worth a million dollars if anyone solved it

and then it went on to solve four to five hundred other problems exactly like

it so they're having this impressive insane ascent when it comes to physics and maths.

Anthropic's looking over at that and saying, okay, cool. So I know you have

an internal model and you're probably going to release your most powerful model.

So we're going to kind of wait until you do it, and then we'll beat you immediately

by releasing our most powerful model so that, you know, it takes the attention away from you.

And OpenAI is trying to do the same thing. There were rumors last week that

GPT 6.2 Astra was going to be released Friday of last But I think they held

off in anticipation for Anthropic to release.

Fable 5.2. Now that they have, I'm wondering whether we'll see OpenAI maybe

today or maybe tomorrow release 6.2 Astra to kind of keep the playing field level.

The craziest part is both of these labs are still harboring way more powerful internal models.

Now, one other thing that I want to point out is as a result of these companies

gatekeeping their internal models is that they now have more compute to allocate

to just training the most powerful model.

Remember, It's one thing having the most powerful model internally.

It's another serving it to millions and millions of your customers.

That costs a lot of compute, a lot of time, and a lot of resources.

But if you simply just don't release it, you now have a lot more compute that

you can, I guess, feed it back into that very same model to create an even more

powerful version of itself. So we see this gap increasing more and more.

And so that's what OpenAI and Anthropic are doing. and they're racing towards

one particular thing, RSI, recursive self-improvement, the idea of a model that

is so smart that surpasses.

Even the smartest minds, human minds in the world to then continually recreate

versions of itself and take over and eat basically any sector or vertical that exists out there today.

Crazy thing to wrap my head around, to be honest. But of course,

the other and final reason why they're doing this is because they have one of

the biggest IPOs coming up.

And as I mentioned earlier on in this episode, they have come up with new details

where it's going to be around 825 million shares.

25 million is coming from Anthropic employees. I believe each founder,

there are seven of them, are donating or selling around.

3.378 million shares, which at the price range of 70 to 78, let's take $78 at

the top end of that range, would net them around $1.

$360 million each. Not bad. But they also want to make sure that they enter IPO week with a bang.

And now they have with the number one model in the world. Now,

I'm sure OpenAI has the idea now to release

their next model that beats their model so that they can kind of dampen the

IPO market open to make sure that people, you know, if they're looking at the

Anthropik Val and they're like, well, OpenAI has a better one.

And they look over OpenAI and they say, well, they have the better model,

I'll just wait until OpenAI IPOs.

I know that Anthropic's defense to this will be simple. I have a bigger and

better model ready to go. And trust me on this, I'll place any amount of money

that if OpenAI does release another model that is technically better than Anthropic's 5.2.

Anthropic will just release 5.3 or Fable 5.5, and it'll blow everything out

of the water right before the IPO day, and their stock is going to do pretty damn well.

Now, on the actual pricing, if the valuation is going to come in at between $1 to $1.3 trillion.

Listen, as expensive as that is, I still think that it is much healthier than

the $2 trillion that we were targeting, and much better than SpaceX's valuation

that came in at, I think, $1.75 trillion.

So we're starting to see the market re-rate in this downturn around a much healthier valuation.

Of course, a lot of that depends on the ARR and continued demand and revenue

generation for these models, but that is what we're seeing and what we're probably

going to see with this new model Fable 5.2.

So yeah, that is the episode. Also leaving comments is honestly the best part

of all of these episodes.

I get so much feedback from you guys. I read all of them and I respond to as

much of them as I have time for, but I've been loving the engagement.

So keep it up and yeah, share this with a friend if you find this interesting

because I'm trying to spread the love and get as much feedback as I can.

So thank you for listening to today's episode brought to you by Qualcomm.