So, Fable just got banned. What are your
thoughts on that?
>> Yeah, so it might be the start of the
permanent underclass. And I didn't
really use this term at all until the
last 72 hours. Since Fable came out, I
started more thinking about are we in
the permanent underclass because it's
twice the cost of Opus and it's heavily
restricted compared to Mythos. You know,
the guardrails were pretty crazy. It was
rejecting even normal questions. But now
after they banned it, the US government
told Enthropic they cannot serve it to
anybody outside of the US. And since
they didn't have any system or
infrastructure for this, they just
pulled it completely. So nobody can use
Fable right now, no matter whether
you're US citizen or not, this really
might be the permanent underclass
because only people working in the
government or people directly at
Enthropic have access to Mythos and
Fable and us the normies, we don't. And
even when they add it back, it's not
going to be available to people in
Europe, Asia, Africa, Australia, most of
the world. It's only going to be for US
citizens.
And yeah, I mean this is highly highly
concerning and uh most people probably
won't think about the second and third
order consequences of this. But if you
go into the future where AI is much much
better and that's like the number one
thing that separates a true AI believer
versus people who are in the AI field
just for money, just for hype, just
because it's trendy, is that do you
believe in the technology? Do you
believe that AI models will become
better and better with each passing
year? So far that has been true. So if
we go into two years from now, 3 years
from now, access to AI will be even more
important than access to electricity, to
access to a car, to a computer, all of
us can imagine how much crippled we
would be if someone took away your
smartphone, your internet, your
electricity, your car. Like you couldn't
function in this world and you
definitely couldn't compete with others,
let alone in business. AI is going to be
that times 100. If someone takes away
your AI access three years from now when
the AI is literally hyper intelligent
genius with access to all your softwares
they can do anything on the internet
they're going to completely [ __ ] you
and people need to realize that this is
a wakeup call the US government banning
Fable is a taste of what the future will
look like today I feel severely less
productive than in the last 3 days and
it's true like Opus and GPD 545 they're
not as good as fable and this is just a
small taste of what's coming.
>> So what was the impact on your workflow
from this ban specifically?
>> Yeah. So all the coding is taking longer
because the models are not as competent.
They're also making a lot more sloppy
mistakes. It's kind of funny because
even a week ago if you asked me I would
say like Opus is amazing. Opus is you
know the best entropic model. It's very
powerful. But today I was working with
it heavily because Fable wasn't
available and it just made made a lot of
stupid mistakes you know you know it
changed some links but it removed like
the tracking link that was the main link
for the video and a lot of like dumb
stuff that like even interns wouldn't
make and it's weird because it's the
jagged intelligence that's the concept
of AI where they're hyper intelligent in
one domain but very stupid in another
domain and they just make a lot of dumb
mistakes and hallucinations and GBD 5.5
right now is the best coding model Opus
is probably generally better in terms of
like daily use, but both of them are
significantly worse than Fable. Fable
just understood all the context. It
pushed through to completion. You didn't
have to tell it what to do. It just test
like anytime you made a coding change,
it tested the changes. Then it deploy it
tried to deploy them, right? So if it
was missing API key, it tried to log in.
It it went and, you know, clicked
through browser use. It did a lot of
more things without you having to guide
it. So it was a lot more intelligent
entity, a lot more complete of an agent
compared to Opus and GBD 5.5. So yeah, I
mean when Fable came out, literally in
the first couple hours, I pushed updates
to all of our internal softwares and I
got so many things done. Now today I
just feel like I'm working with like
tools, you know, really GPT 5.5 and
Opus, they feel like tools. You need to
guide them heavily. You need to tell
them what to do. Fable felt like
something more like like almost like its
own entity like it knows it understands
on a deeper level what are you what
you're doing why you're doing it how
good work looks like it's it's
completely different
>> so you don't think giving a fable to the
masses is dangerous
>> I mean you you can say like giving fire
to the masses is dangerous you know
giving cars to the mass is in dangerous
anybody who owns a car can just like
drive it into a mass of people in theory
you know obviously us Europe Europeans,
we don't do that, but some other
cultures might do it. Anything giving to
the masses in is dangerous, you know.
So, I I don't think that's a strong
enough of an argument to prevent it.
Same thing with the internet. Like when
the internet was coming out, there was
all this fear-mongering. Same thing with
Bitcoin. Oh, Bitcoin is used by
criminals. So, any technology can be
used for good and bad. But I think we
can all agree that it shouldn't be
centralized in the control of a few. Who
are those few? Is that the US
government? I mean over the past, you
know, 12 months, a lot of shady stuff
came around with the top people in US
government. Do we want those to be in
charge of the world's most powerful AI?
Is it the people leading the AI labs?
Okay, maybe that's a bit better. At
least they understand technology more,
but still, do we really trust a handful
of people on Enthropic or a handful of
people at OpenAI? I mean, Sam Ultman
doesn't have the best track record. You
know, everybody understands that he's
pretty good at, you know, getting his
way. So I I would say these are not the
best people to be in charge of super
intelligence no matter what. The
question is who is you know I don't
think a full democratic system where
average Joe on the street who's zero
knowledge of LLMs or how tokens even are
generated should have the same opinion
as Dario Amod definitely not. But like
the people at the US government are are
way more clueless than the average
YouTube viewer. I mean most people
watching this video probably know more
about AI than some like US Secretary of
State or some top people in the US
government. Yeah, these are the ones
making the decisions for billions of
people across the planet. We are not in
the US and we're affected by this
massively. AI is people say like okay at
least open AI is in the in the lead. You
know when CHB came out it was basically
all open AI. Now there's Enfropic,
there's Gemini, there's Grog but all of
these are US companies. Then you have
the second wave which is Chinese
companies which I have to give massive
credit for because they're open sourcing
all of their models. So even if the
Chinese government was like let's ban
it, it doesn't matter. The model weights
are downloadable online. You know
millions of people already have them.
All of you can you know me and you guys
you we can download these models on our
laptops on our hard drives create
backups and just run them. We can
fine-tune them. We can post train them.
All of that is available and you cannot
take it back. So
really AI is not distributed. There is
not that much competition. It's really
US versus China. Unfortunately, here in
Europe, we are falling behind heavily.
The best European company right now is
Mistral and it's very behind the top
Chinese models and the top American
models. So AI is not really, you know,
it's not it's much more centralized than
people think. I'm just going to put it
that way. It's really US and China and
US is uh it's completely different
structure. US is more decentralized
government and more centralized
companies. China is a centralized
government and the companies are
decentralized. All of the Chinese models
are open source and uh it's fascinating
how different their approaches are and
right now it seems like all it takes is
somebody and the US government and they
can just shut down all AI models. It's
not impossible to imagine like they're
going to even remove GBD 5.5 or when GBD
5.6 is ready and OpenAI wants to release
it. Maybe the US government shuts that
down as well. Like who knows?
>> But speaking of Chinese models, why do
you think that a lot of these Chinese
models are able to be so cheap?
Yes, that's a great question. Um, they
China is really great at publishing
their research and at innovating novel
techniques. So, they have like some
breakthroughs in terms of inference and
also they have a lot of restrictions on
chips. So, that forces them to be more
efficient to, you know, get more done
with less resources. But there's also
this theory I saw on Twitter which I
think more likely is true. With each
passing day, I I'm starting to believe
it more and more. And that is that
instead of subsidizing which a lot of
people think like anthropic and openai
are heavily subsidizing their
subscriptions you know Dylan Patel
posted a research that like the $200
plan at cloth gives you like $8,000 API
worth right so if you had to pay API
pricing you you would have to pay
roughly $8,000 to get the same amount of
usage as the cloth max plan gives you
but for CHBD was even crazier the chbd
pro plan which also cost 200 was like
$14,000 worth of API credits, right? So,
either
they're subsidizing it super heavily, or
the second theory, which I think is more
likely, is that they have actually
disgustingly high profit margins, like
over 90% on the API tokens, which allows
them to put these subscriptions out at,
you know, a discount and maybe they
break even on the subscriptions, maybe
they lose a bit, maybe they profit a
bit, doesn't matter. But the API tokens,
that's where Enthropic and OpenAI could
be insanely profitable. Like we're
talking 80 90 plus percent profit
margins, which would make sense why
these Chinese models are a lot cheaper
because maybe they're closer in terms of
how expensive they're to run to
inference. Maybe just the closed
companies are pricing them massively
heavily because they're in the lead and
they can afford it and people will pay a
premium to get the most highest quality
tokens, which I definitely am one of
those people. I'm happy to play pay a
premium to get the highest level of
intelligence.
This might be a plausible theory where
the cost to run these models is actually
similar, but the closed companies just
charge a lot a lot more.
>> So getting a bit back to Fable, do you
think we're going to get it get access
to it again? And if we do, what will you
do?
>> Yeah, so this is highly speculative.
Obviously, it's still a developing
story. It's like less than 24 hours
since this happened. Uh I think the
first step will be US government or like
US um citizens will get access. So
Anthropic will figure out some
verification system whether it's a
social security number. Upload your
passport. I mean most Americans have a
passport so maybe that's a bad idea you
know upload your ID whatever something
like that to verify is this user a US
citizen? Is their IP address in America?
Stuff like that right? So obviously I'm
not a US citizen so we'll have to figure
out a way. Um but whether everybody in
the world would access maybe like I
don't know if someone anthropic you know
does a good job explaining it to the US
government then maybe they lift this
completely and everybody else in the
world can can uh access it. Maybe they
implement a new set of oversight rules
where somebody at the US government can
actually look at the at the messages
because that was the biggest issue with
Fable is that like they
had a data retention policy that even
people using through API had to send the
data to enforc and they allow they were
allowed to keep it for 30 days which
prevented a lot of companies from using
it. We can touch on that a bit later,
what that implies for companies and why
local models and self-hosting is so
important. But the main issue is is like
even if they allow it,
it's it's it's like I I'm I'm having
like some hunch that this is not as
simple as it seems. Part of me thinks
like Anthropic kind of wants this. If
you think about it, this is like the
biggest no notoriety for them. This is
the biggest like, oh my god, this
company is dangerous technology. the
biggest promotion from the US
government. It like not all of this
feels organic to me. You know,
anthropic, you know, Dario Amod and the
entire leadership, they were the biggest
fear-mongerers for years. You know, they
were comparing it to nuclear weapons.
And you know, again, I saw a tweet where
someone was making fun of it, like, oh
yeah, compare AI to nuclear weapons 800
times and and you're acting surprised
that you get regulated. It's like they
they were kind of asking for it. So, I
don't know. Like anthropic, it's very
conflicting. They they first say like
mythos is way too dangerous to release.
You know that also at the start of April
it's like this we're not going to
release this. This is cyber security
blah blah blah. Fast forward two months
and they release it even more powerful
version of it albeit with guard rails
and and you know safety layers. But then
like few days later they get regulated
by US government which is what they were
kind of asking for all all the time and
now they're ask like saying it's a
misunderstanding that they're acting
surprised. like all of this feels very
weird, you know, it something isn't
adding up and I think the official story
is not what actually happened. So, I'm
going to put it that way. To answer your
question, very first thing, if I got
access to Fable Now for 1 hour, I'm
asking it
like all of the questions. First of all,
I'm creating long-term plans for
everything. Okay? So like long-term road
maps and long-term plans for all areas
of my life, how to scale my business,
all aspects of business, all of the
startup ideas I have, all like every
task I have like to completion, every
single like possible scenario, all
geopolitical conflicts where they could
lead and the implications of that and
you know different paths to technology
and the future basically trying to
create as many long-term plans as
possible so that whichever one is true I
can then if fable gets taken away I can
use less powerful models and agents on
those steps which were orchestrated and
created by Fable. When it comes to like
personal use, I would just ask it as
many look first of all like I would just
ask it as many questions as possible. I
would just run it in some loop through
API to generate data sets basically
distillation without you know
distillation obviously it's against
entropthropic terms of service so I
wouldn't do that obviously I'm just hypo
you know hypothetically speaking but I
would ask it like questions about life
you know I think one of the best prompts
to LMS is like based on all the
knowledge you have what is the one habit
that humans could do to succeed in life
these types of prompts you know based on
all the knowledge you have and all the
research
What is like one health tip that would
make your life better? Or like what's
the number one mistake that humans make
that ruin their relationships? Stuff
like that to just like get the limit the
limit testing questions. And I would do
that for hundreds if not thousands of
different areas of life and business and
everything and you know turn that into a
data set of sorts to just have all these
Fable tokens so that if Fable gets taken
away again I can look at them and I can
read them and it's like remnants of a
intelligent entity that no longer
exists.
>> So do you think Fable was AGI?
>> I mean it's the best claim for AGI that
there has ever been. Obviously the
definition of AGI has shifted over the
years. I remember when CHGBD came out it
was uh like I think if you show that to
when ShvD came out Fable with all the
tools and you know CLI and all the
entire harness people would say like
yeah this is AGI for sure because you
know ask yourself would you prefer
talking to Fable or the average human on
the street and even say like the average
knowledge worker right like someone
working in IT or whatever or above
average like legit I would prefer
talking to Fable
over 90% of humans like it's just more
intelligent it's more capable it's it
has more understanding so do we call
that AGI or do we not call that AGI
because some tools are missing because
it you know struggles to navigate the
web because it's heavily restricted some
websites don't allow agents and bots to
access it doesn't have a phone number
like those are all like limitations that
could be solved even if the models
stopped improving we would get far more
powerful AI tools purely by the
connectors, the integrations, all of the
tooling around it, that alone would
still have at least five years of
massive growth. Even if you take models
from last year and you give them proper
tools, you know, proper harness, they
can do a lot more. So, nobody really
knows how fable how powerful Fable is
because ultimately these models were
limited by a the humans, you know, what
the humans give them and b all of the
tools, the harnesses. If the model had
the same access as a human in terms of
it can do banking, it can make payments,
it can use the email, it can use a phone
number, I think it would economically
outperform most humans. Obviously, it
cannot do stuff in the physical world.
So, ironically, people like doing
construction, they're probably safer
than people like us, you know, in terms
of their job not being replaced. But on
the web, anything on the software,
especially like there's like sites where
you can like pay humans to do some
things. I don't know. like I I'm
struggling to find things that it
couldn't do. So yes, I think Fable has a
really good claim at being AGI.
>> So potentially if everybody in the
future had access to a model with the
power of Fable, but it was open source,
no restrictions. Do you think that would
be dangerous?
>> Again, like the the issue is the
dangerous argument is really um really
flawed. Like anything is dangerous, you
know, like giving a kid a book can be
dangerous. He can use that kid to smack
another kid on the head. You know, like
giving a rock to a homeless man can be
dangerous. He can throw that rock at
somebody. Like something being
dangerous, it doesn't mean it shouldn't
be done. Obviously, I think that would
be the greatest blessing for humanity if
model like Fable got open source and
that would also prevent for any
government from seizing power over it.
Like anthropic, the biggest risk for
anthropic is they should be worried of
being nationalized, right? Like America
is like, "Oh yeah, we're capitalist.
We're anti-communist." China bad. China
bad. If it's kind of funny if you watch
the Bloomberg interview with Dario Ammo,
he's like, "Oh yeah, we don't want the
bad guys to win." First of all, who are
the bad guys? But most importantly, who
decides who the bad guys are, right?
People have all these criticism of
China. So, oh, China this, China bad,
China this. China is releasing all their
models open source, right? If you forget
everything else and look at their
actions, they're the biggest good guys
in all of AI, in all of technology.
China is contributing the most to the
world by far. They're open sourcing
their re research. They're open sourcing
their models. They're not charging crazy
prices and API. They have the most
costefficient AI plans out there. Like
China is doing the rest of the world a
huge favor. Imagine being like a
developing country. Imagine being like
a, you know, random European country
where you don't have your own AI
startup, your own AI research lab. You
can get so far ahead with Chinese models
compared to US models. US models are
closed. US models are heavily guard
rails. you cannot fine US models. I mean
I don't know like what what is this
notion that the China are the bad guys?
If you just use your eyes and look at
the actions in the AI space, China are
the good guys in the AI space. Okay, we
we can you know it's other time we can
talk about politics, history, communism,
capitalism, whatever. But in terms of AI
space, China is doing much more good for
the world than America.
>> So given all the recent events, what are
your plans moving forward? Are you going
to change the way you operate with AI,
the way you operate in business, your
personal life given the current
situation with Fable?
>> Yeah, I'm definitely moving in the
direction of self-hosting, of
fine-tuning, of local AI models, a lot
more. And you know, again, there's so
many more levels before you can
pre-train a large AI model. That
requires a lot of skill, a lot of
capital to get the compute, a lot of
time, talent, all that stuff, right? We
don't have that. Like we're falling
behind here in Europe. It's kind of
crazy. 100 years ago Europe was the
pinnacle uh continent and now you know
because of regulation, two world wars,
all this [ __ ] We're falling behind.
But you don't have to start by
pre-training a new AI model. You can
start with fine-tuning. You can start
with post- training. You can start by
having enough hardware to run these
medium-sized and large AI models
locally. You can even start by
downloading the weights. You know, all
of you guys should go to hugging face
and download the weights of some AI
models. Just put it on a hard drive, you
know, whether it's Miniax, whether it's
Kimmy, whether it's GLM, whether it's
Quen, just download these models just in
case. Who knows what's going to happen?
And again, let's let's not like say it's
only the US government. The EU is even
crazier than the US government, right?
Right now the the Americans are giving
ideas to the European bureaucrats in the
EU in Brussels and it's not impossible
that like hugging face will get seized
right or like open source models will
get outlawed. I mean the Biden executive
order wanted to do crazy [ __ ] like
specific computer I think it was like 10
to the^ of 26 flops or something like
that where any model that was trained on
more than that will get seized. But this
is just math. This is mathematics. This
is matrix multiplication. You know,
again, people on Twitter were memeing
Trump with like a paper just like doing
matrix multiplication by hand. Like,
come and take it. That's what AI is at
the end of the day. You know, if you
scale, if you that's the problem. Like
people don't really understand what AI
is. They think it's like, oh yeah,
Claude has feelings, blah blah blah, all
this [ __ ] It's next token prediction.
Okay? And this at at the inner core is
math, matrix multiplication. If you do
that enough, turns out you can get some
impressive intelligent answers from it.
But anybody with enough compute and with
the right skills can run that math to
create AI models. And again, I would
encourage everybody watching this to
learn more about local AI models to
invest a bit into some hardware. You
know, buy more RAM, VRAMm, GPUs, upgrade
to a better MacBook, Mac Studio,
whatever it is. Whether you prefer
Nvidia, whether you prefer AMD, whether
you prefer Apple Silicon, download some
models locally, have them, learn how
they work, get into fine-tuning. We
really need to distribute this. It needs
to be decentralized just like Bitcoin is
decentralized. Central banks are very,
very evil. And I'm I'm not going to get
into the discussion. That's another
two-hour podcast. The same thing is if
central AI, like AI is going to be way
more important to our lives than
banking. Way more. It's not even close.
Like again, I'm I'm telling you guys, in
two years, in three years, in five
years, if someone takes away access to
your AI, you're [ __ ] You're going to
be beyond crippled. You're going to be
useless.
Just these agents will be way more
intelligent, way more competent with all
the tools. They'll be able to do any
tasks and you can spin up hundred of
them. So, you'll have an army of agents
doing all of your work. You're going to
be orchestrating them, and then if
someone turns the switch and disables
them, it's over. So, you better get that
urgency right now, not in the future
when it's too late. And I I feel like
I'm crazy because I'm one of the only
people who understands this. I have a
lot of friends in AI and when Fable came
out, I was like in disbelief because so
many of them either didn't try the model
or tried the model and like ah it's too
expensive, you know, whatever or like
used it and were like, oh, this is
impressive and didn't really change
their trajectory, change their business
plans or anything any of their plans.
And now when fable is taken away, they
don't have the urgency. A lot of them
are also Americans, so they're like, "Oh
yeah, we're going to get it back. We're
Americans, you know, please, please, US
citizens." They don't have that urgency.
We are Europeans. We are not on this
fable access list. So this puts me in a
very unique position where I have enough
experience with local models,
fine-tuning, all that stuff. I'm
definitely not an expert yet, but I made
enough videos on it, so I understand.
And also I understand on a philosophical
level the importance of decentralized of
libertarian values of not having a
central entity controlling access to the
most important technology of all time.
So that puts me in a unique position
where I actually understand the urgency
of the moment and I I I can see the
second order and third order
consequences of this. If nothing
changes, we're screwed. We're completely
[ __ ] So we cannot let a few people
whether it's a government, private
entity, whatever, control who has access
to super intelligent AI.
So obviously for people who are locked
in on AI, it makes sense to download
their own models. But what should they
tell to the people in their lives who
aren't really focused on AI? What should
they tell them to do?
>> I mean, rather than telling them what to
do, well, they should first tell them
the message I just spoke, right? how
important it is to have the greatest
technology decentralized, open source,
as unrestricted as possible.
Second of all, they should like do
something rather than telling them,
right? they should spend a couple
thousand dollars to spin up a home
cluster AI cluster at home their you
know mini supercomputer whatever you
want to call it that you can run
inference you can find cute models on
and you can serve that as a API for your
family for your friends for your
neighborhood for your village like you
know up to you how much compute you have
and you know get into building models
like if someone is intelligent competent
in AI create AI model for your community
I mean I was shocked a couple hours ago
that the IT department, municipal IT
department at Rio de Janeiro, which is a
completely random country in like Rio
Janeiro is a city in Brazil. This is a
completely random country in the AR
race. Nobody even thought Brazil is in
the AR race, but turns out they now have
some impressive open source models
because people at the municipal IT
department of Rio de Janeiro city are
cooking, fine-tuning. I think it was
some quen model or whatever. But that
that's crazy, you know, like huge shout
outs to that municipal IT department.
They are cooking and that's what people
need to be doing. People who understand
this the gravity of this moment they
need to get into fine tuning. They need
to take, you know, some of these Chinese
models are impressive, post train them.
They need to, you know, watch some Andre
Karpathy videos and learn about how to
build a small model from scratch and
start creating these custom models for
the communities instead of relying on
openic.
So my last question is do you think
these model or these AI companies are
responsible for actually open sourcing
these models due to the fact that they
basically took all of humanity's
knowledge in order to train these
models?
>> Yeah, I mean definitely you know
enthropic and open they're like oh no
please don't distill our models it's
against term of service. They distilled
all of human knowledge. They took away
all the novels, all the art. And again,
I'm not for like the doomers, you know,
I'm definitely not on the side of like
all these leftists, but the best
argument that these people who are like
protesting AI have is that it is stolen
data. All of this data was on the
internet, usually voluntarily
contributed for free by, you know,
novelists, artists, programmers, people
contributing highquality tokens to
humanity. And these companies, whether
it's Google, Enthropic, OpenAI, XAI,
Meta, they just scraped it. They scraped
the web. They they stole that, you know,
they took they took the data and built
the models on top of them and benefited
massively. Like these companies are
trillion dollar companies, right?
Whether it's Enthropic reaching trillion
dollar valuation, OpenAI, Google adding
another trillion, I mean Nvidia added
multiple trillions, SpaceX is valued at
two trillion. All of these companies
benefited massively from scraping the
web and training AI models, proprietary
closed source AI models on top of them.
So yes, I think they should be if the
government should be doing anything,
they should be regulating them so they
open source models, not a [ __ ] ban.
That's terrible. It shouldn't be a ban.
It should be like, yo, Enthropic, you
must open source models. That's it. That
would be insane. And that would unlock
an era of abundance
never before seen. Like imagine if Fable
got open source, right? And you could
fine-tune it. You could run it. I mean,
it would be expensive to have the
hardware, but you could run it locally,
you know, at your office, at your home,
and you could post train it. And like,
there would be such an explosion of
research papers, so many different
versions of fables, so many different
API. Some people would figure out how to
run it inference quickly. Some people
would figure out how to quantize it and
run it on a MacBook. It would be a
renaissance of progress of humanity.
Everybody, again, most people don't want
to create bioweapons or or anything like
that. people who figure out how to solve
diseases, how to solve math problems,
how to advance technology. That would be
insane. That would be incredible. And
that's if someone is watching this for
the government, that's the type of
regulation you should be pushing for.
They took all of humanity's knowledge.
They distilled it. They scraped it.
The model should be open source.