Thanks for watching!

welcome again to another episode of cloud

unplugged we have two big topics uh well

one big topic one smaller one um we

have anthropic uh accusing alibaba of

basically a distillation attack against

their model so reverse engineering um what

they can work out how the model was

built um what type of features and

capabilities it's providing

to accelerate their own AI growth.

And then we also have the fact that

Mapbox just got twenty percent more

expensive because of AI.

So if you're out to buy a MacBook,

you better go and get a bank loan.

So how are things going with the Anthropic

story, Salman?

What was going on there?

Well, John, here's the question.

Why would you want to spend hundreds of

millions of dollars to train a model when

you could just learn from one that already

exists?

And that is the question Alibaba asked.

And what they did,

what Anthropic realized that between April

and June,

so I think the timelines are between the

twenty second of April and the fifth of

June.

Anthropic identified that there's a

coordinated campaign by Alibaba's Qwenn AI

models.

You know,

the Qwenn AI models that perhaps you've

used or perhaps you like them or not.

They used almost twenty five thousand

accounts and they conducted around twenty

million interactions with Claude and we'll

come to what an interaction is in a

second.

The goal was to systematically extract

Claude's capabilities.

how it clothes reasons,

how it uses the tools, how, you know,

the coding abilities to train their own

smaller coin model.

So this is what Anthropic found out.

And what they did was as soon as

they identified,

and they called it corporate espionage

because it is corporate espionage,

what they found out is then they went

to the U.S.

senators, a couple of U.S.

senators, and told them about it, that,

you know, this is what's happened.

And they sent a letter to the Senate

Banking Committee.

And then two days later,

once they identified it and sent the

letter to the Banking Committee,

you remember when Mythos and Fable got

banned from non US use

That ban got put in place.

But the initial thing is that Alibaba was

using these models, the Claude models,

to train their own coin models.

That's the headline story, John.

And of course,

there's a lot to it than just what's

happened.

But this is what they found out.

Do you think,

because I don't know if you saw,

but Anthropic...

have obviously beaten OpenAI on the

market.

I think OpenAI dropped to like twenty

seven percent of market share for

enterprises.

But it sounds like maybe because how many

subscriptions were they getting from

China?

So we're saying that basically that twenty

percent,

that additional twenty percent that made

them forty seven percent market share was

really coming from China.

And maybe there is no,

maybe they're on a par.

I mean, were they making money from China?

We don't know.

I mean, that's right, right, John?

Look, you know, we said that Airbnb,

not Airbnb, Uber came around and said,

well,

we're not really seeing a return on

investment.

Well, Alibaba did.

They're training their models.

They've got really good ROI.

They've got the best ROI.

Yeah, the best ROI out of that.

So, John, it's like,

if you look at an interaction,

it's anywhere between five hundred to a

thousand tokens, right?

Roughly.

Just talking about roughly five hundred to

a thousand tokens.

Let's just say a thousand tokens.

And there's about twenty nine million

interactions.

So it's about, you can say,

between ten to twenty billion tokens that

were used, right?

Roughly, roughly.

It's just rough, rough estimates.

is a reasonable amount of money,

like maybe in hundreds of thousands of

dollars.

So perhaps not a lot in the grand

scheme of things.

And we're like, turn it back on.

We've got an IPO coming up.

Come on, guys.

Come on, China.

Get some more subscriptions going.

We're about to IPO.

So the thing here is that, of course,

this is not the first time that's

happened.

They also found out in the past that

Deep Seek carried out about one hundred

and fifty thousand of these interactions.

Moonshot AI, again,

another model from China,

about three point four million

extractions, right exchanges.

And there was another one called Minimax,

about thirteen million.

I mean, there's still still a lot,

but not at this level of almost thirty

million interactions.

Right.

So this is.

this has happened in the past of course

anthropic every time they find it they

improve their anomaly detection they've

been trying to figure out how they can

make it more secure because this is one

of the things that people have been

talking about this model distillation

where people can use your model to learn

the secrets that you're using but then you

know do you know how they do it

though as in like i mean i i

let's put you on the spot because you

may or may not know but i'm not

sure how

how do they use that information to then

improve their own model and accelerate the

learnings through another model?

Yeah, I think,

I don't know the exact details, John,

but it's feeding whatever output of those

specific questions that they have and bake

it into the model patterns, right?

So they ask specific questions,

specific patterns,

and then just feed that into the model

itself, perhaps.

Right, they're basically prompting.

So they've got a load of,

specific tailored specific that are then

used as a mechanism to then obviously

extract data in a format that then helps

them with the model that then they can

then accelerate their own training using

the app of another model correct and yeah

that's that that's correct this feels like

a sense of um

karma isn't it isn't that called karma

where you train a model based on every

other people's work but you're not happy

that somebody else trains their model and

your model that was trained on other

people's work just so i understand the

slight hypocrisy here so they're not happy

that somebody else is getting ahead based

on the data that they train their model

on which obviously wasn't theirs um but

they're not happy that somebody else does

it to them is that is that what

i'm hearing here

That's exactly what you're hearing.

That's exactly what some people are

saying.

They are saying that, well,

isn't what anthropic getting the taste of

the bit of the taste of their own

medicine because they trained it on data

that already exists, right?

Of course, it's open.

But this comes under John, of course,

with all of these commercial things,

fair usage policy, right?

So this is protecting their IP to some

extent.

Because if you use fake accounts,

bots or whatever,

about twenty five hundred twenty five

thousand of those i think legally it does

it you know it does count as corporate

espionage but this is why this is why

anthropic went to the senate banking

committee and the us the the commerce

department to restrict maybe to put in

some of these uh some

some of these guard rails around to

protect themselves.

I think it's a bit more geopolitical than

technical, but that's what's happened.

What do you think about all of this?

I just find it ironic.

I mean, I don't know if it's ironic,

but just hypocritical.

Just the hypocrisy levels of when it

happens to others, we're okay with that.

When it happens to me,

I'm really not okay with that.

Like the double standards as a business.

You know,

it does feel a little bit ridiculous that

they have an issue with their own business

potentially getting risked because they

don't work for the state.

The state hasn't got investment in

Anthropic.

It's not a government model, right?

So it's a pure commercial risk.

Let's be honest.

Because there is no,

you don't need to go to the US

Senate for

something that is a real commercial threat

potentially it doesn't become a national

threat because somebody can produce a

model like yours because say the argument

was that their quen model was on a

on a par uh with fable or mythos

then what would they have done like you

wouldn't need to block because they've put

the investment in they've trained the

model to the same standard so it's just

equal measure so there's nothing to do

right to anthropic anthropic couldn't do

anything

in that situation where China's built a

comparable model to them, right?

Yeah.

But I think, John,

based on the back of this,

the ban on this export control of Mythos

and Fable shows that this is being treated

the same way as other technology.

That's an export ban that doesn't get sold

to every country.

So this is where this is going.

Yes, the government is not involved,

but the government sees the potential on

what these models can be used.

Don't forget that Anthropic are the ones

who denied US government usage of their

models with the removal of guardrails so

they can use it for whatever they wanted

to use it for.

So it's the same company.

And then they went back on it.

But here's the thing, though, John,

if you compare these models,

I was just looking at the cost.

these models right i don't know i've never

used any of the alibaba's quen models but

the entropic if you just take opus opus

is five dollars per million input tokens

and twenty five dollars per million output

tokens right now the mythos and fable of

course they were more expensive yeah do

you want to guess how much quen is

oh i mean is it done in dollars

you've done it in dollars you haven't like

i'm just doing a dollar now i'm not

changing it to yen or anything like that

yeah okay so that's good um

I don't know.

I guess what are we talking?

Probably I'd be annoying if I'm right as

well, which is kind of the risk,

isn't it?

That I fluke myself.

I fluke myself into the number.

I'm going to say ten.

Ten dollars.

Ten dollars.

Okay.

It's five cents per million input tokens

and twenty cents per million output

tokens.

Why are we not using Quen then?

Why are we doing nonthropic then Salman?

So what I'm telling you is that there

are some models out there which are way

more because you could train based on

other models.

I don't know because

The market,

the way they present themselves is that

this is more affordable model to use.

Of course,

I've never checked the capabilities.

There's a bunch of benchmarks out there

that you can see always put Claude Opus

to the top.

But okay, fair enough.

Claude is the best Opus.

The Opus model are the best at the

moment.

But do you need that?

Maybe you can get away with Quinn.

It's five cents.

John,

this is like hundred times cheaper than

the models that we use every day.

Outside of the cost,

which is obviously ridiculous,

that there's such a disparity between the

costs.

But you're in a global economic situation,

right?

We've got globalization.

Companies now have a global footprint.

You've got people in Europe using

Anthropic, et cetera, et cetera.

Obviously,

there's a whole sovereignty thing going

on, which we've mentioned several times.

So I think everybody's looking at everyone

on the risk of not owning something,

especially a frontier model,

where if for some reason the adoption –

kind of goes through the roof and the

ROI does start to get recognized.

And then there is a disruption to

employment.

Let's just assume that path gets taken.

You don't own anything.

All of that value has gone to a

specific company.

And that company has nothing to do with

your country, right?

It's not European tax.

It's not going back into Europe or any

European country.

It's going into the States and it's

obviously grown the US economy.

And so knowing that you're in a globalised

position and that you've got a competitor,

Quen,

that's then sneakily trying to accelerate

its AI adoption.

Is it a little bit more underhand?

Is it like someone maybe taking steroids

in the Olympics?

Maybe.

Maybe it is a little bit like that.

But they're all competing.

They're all competing at the end of the

day,

and you can't suddenly start be okay on

one hand.

When we make the rules, it's okay.

And we're controlling the globe.

That's okay.

And so long as you follow our pattern,

that's fine.

But if you try and do it as

well, that's kind of not okay, guys.

It's not okay that you improve your mind.

It's not okay that you can monetize yours.

Do you know what I mean?

It just feels a little bit like that.

And you're charging five cents.

What are you doing?

Yeah,

charging five cents is a bit outrageous.

They can't make any money out of that,

surely.

But, John, it's the scale, right?

So it depends how they train,

how they run these models.

And, of course, here's the thing, though.

Sorry, just a quick one, though, on that.

Yeah.

Does Alibaba Cloud have an open source

model as well?

Have they bothered with that or they

don't?

They're just all commercial?

I have no idea.

Yeah,

I know you've got obviously some open

source, but I don't think they do,

do they?

At the moment,

the only country that's doing

doing a lot of open source model is

China, right?

So China's like deep seeks coming out as

open source.

There's so many models that come out.

That was it.

That's where the thread was.

Yeah,

deep seek is coming out as open source.

You've obviously got Gemma as well,

which is Google, isn't it?

And I have just fact checked, John,

because we can do that with the use

of internet.

Quinn is most widely adopted open source

AI families.

That was it.

I thought it was open source.

You're right.

So I guess the risk is then beyond

commercial because it basically means,

well, if this local model,

open source model gets more widely

adopted, you know,

then how do you make money?

If it's on a par and you can

run it yourself,

then we're going to go out of business.

So we need to stop it.

So, yeah.

you know another thing is right so they

caught this attack but they caught it

after the fact right yeah so you mean

has it already got better so what you're

saying is john we need to get quen

it's now we didn't get quen john i'm

telling you it's basically you know went

up right now let's go and get it

Yeah.

You know, in Scooby-Doo,

at the end of it,

when they catch the perpetrator and they

take the mask off, they take, like,

the Quinn's mask off.

Actually, underneath it's just Opus.

So I think that's what it is.

Oh, yeah, underneath it.

You know, it's one of them.

So it's just Opus.

Oh, no, it'll be Mythos.

That's the illusion,

because we don't know Mythos,

because we've never heard of it.

The mask illusion, that is Mythos.

And you actually see, like, Kaiser Soze.

Yeah, exactly.

But you know, the thing here is, John,

yes, over time,

they've improved their anomaly detection,

trying to figure out because it's kind of

hard to figure out these attacks and the

patterns that they exist.

So they can see how they how they

catch these people, right?

People were doing this.

But imagine all the other people who are

getting away with distillation.

All the other sports.

I'm not saying it's happening.

Perhaps some of the other competitors.

That sounds so guilty right now.

That was such a,

the way you phrased that was like a

real,

it was like you really wanted to confess,

but had to hold yourself back on the

fact that you've got something going down.

Is that what you've been up to?

Listen,

I'm not going to say any of the

SAR models that are going to come out.

I'm not talking about any of that.

You'll find out later on.

I'm just telling you, John.

SAR AI.

I'm just telling you that this kind of

attack is... But again,

the model is there for the usage.

If you stick within fair usage policy,

you're learning from the teacher, John.

Is that legal?

Yeah, they learned from the best.

They copied their model and was like, oh,

you can steal and train.

Yeah, let's follow suit.

I do think, though,

somebody was using Quen.

That's why in the back of my mind,

because I remember having a conversation

around somebody was using the Quen open

source model for quite a lot of their

coding and then there was some internal

discussions not to go off on too much

of a tangent but there's internal

conversations around actually you could

conserve your money by running something

like quen to kind of get going to

a certain extent say we're going to write

some code or whatever you're going to do

you could use your local quen model get

it so far you know to iterate to

a certain point and then use the polish

and the rest of it maybe like the

getting it reviewed,

making those changes with something then

like Anthropic so that actually you can

kind of save your costs by trying to

get to a certain point.

So you don't obviously hemorrhage tokens.

And actually some people have already

started to do that as a bit of

a way of working,

which is kind of interesting.

you know what john there's probably a very

good topic to discuss at a future uh

cloud and plug right okay i get it

i can i take the hint i take

the hits on them and that's all right

how do we how do we i'll drop

that no sorry sorry um i'll drop that

just for another time

How do we efficiently use our tokens,

right?

Because you're absolutely right.

It's all about the tokens, John.

And if you can use a cheaper model,

it's all about them tokens.

I need to mark the token max.

If we can use some of these cheaper

models to do most of the groundwork, then,

of course, why we can't polish it.

I kind of do the same with...

With Claude, I use Haiku,

which is much cheaper than Opus to get

going with a bunch of stuff.

And then I switch over to Opus when

Haiku starts to lag a little bit and

it's not really doing what it's supposed

to do.

Because the Claude models,

they're not five cents per million input

tokens, right, John?

So I have to do something about that.

Yeah.

But what about, I guess,

moving on to MacBooks?

And the fact that they're going to cost

twenty percent more.

I mean, they weren't cheap.

MacBooks are more on the expensive end of

the price spectrum,

but they're going to go up by roughly

fifteen to twenty percent.

Fifteen to twenty five percent this week,

they said.

Um, all because of the AI demand,

obviously the, the,

the memory and the chips that are in

your laptop also go into the AI servers.

Now you've got the competition obviously

happening on a shortage, um,

of like memory, et cetera.

And that's obviously put the demand up and

obviously supply and demand it's then made

the cost higher.

And therefore this is a trickle down into,

um, into obviously the, the,

the laptop market or in this case,

MacBook, but.

What do you think about that as in

now consumers are starting to feel the

pinch on this AI hype?

Yeah, John, supply and demand, right?

Because the hyperscalers are buying up all

this memory that goes in not just

MacBooks, by the way, it's also iPads,

iPad Pros and all of them.

Everything has gone up by like,

twenty percent of the cost.

So the RAM

hyperscalers are buying everything right

so they're buying the gpu chips they're

buying the rams so supply and demand

there's not enough memory available in the

market for to be put into these these

machines it's costing a bit more money for

apple to build it and the problem here

is it's not just

just Apple,

pretty soon we'll see this trickle down

into other devices, too,

and probably phones as well at some point.

So unless the bubble bursts and people

start realizing that bubble is expanding

way more than it should,

I think the price is just going to

keep going up.

So this is basically what we're doing,

John, you and I.

If we buy an iPad or a Mac,

we're subsidising the arms race.

That's what we're doing.

You and I, John.

We are responsible for this.

The booze.

Do you remember the last time you were

saying that they spoke to students and

obviously with the job market and job

market threat and I think it was Eric

Schmidt, wasn't it?

Or something was doing his talk and

everyone was booing him every time he

mentioned AI.

And now...

just the general public is going to start

booing because they'll be like, right,

I'm going to get the new iPhone or

I'm going to go and get the new

MacBook or I'm going to go and whatever

device.

And then they're going to realize it's

like three or four hundred pounds more

expensive than it was last year.

And then they're going to be booing and

be like, yeah,

that's

That's because of AI, mate.

They were like, well,

you can use AI to create all the

funny images that you wanted, but sorry,

unfortunately,

you're going to have to pay a bit

more.

Exactly.

Now you're feeling the pinch.

Yeah,

it's a bit crazy when it starts to

hit consumers like that,

as in the trickle down effect that's kind

of happening now because of the demand.

And that's a little bit.

Crazy, really.

Correct.

I mean, we also see the models themselves.

Slightly over time,

they'll get more expensive.

We've seen co-pilot go into usage-based

consumption.

Well,

that's okay because you are in the market

for using the model, so you understand,

okay,

I have to pay a little bit more.

But people who are buying iPads and

laptops,

they're not in the market for models.

Not everybody, John,

as much as you and I like to

think, not everybody's using AI.

Exactly.

Do you know, though,

how you can make your money back on

all of this if you want to make

some money?

You're a content guy,

so you can produce your content.

You can now monetize it.

You can now monetize your content.

And you can make your money back.

So all the AI bots that are scraping

your content and stealing it,

you can gate it now and get some

payment.

Get that money's coming in for all that

content to pay for your new MacBook that

the AI was stealing to begin with.

Or maybe you're using AI to make the

content that's then making your MacBook

expensive to obviously create it.

So then you've obviously got to get the

money back to pay for it.

You caught me.

You caught me, John.

You caught me.

It's an interesting time.

Yeah.

Interesting times and the money game.

But yeah.

Sorry about that.

Amazon, yeah,

have also released something because

they've been in New York talking about

that.

So I just thought it was a bit

of a slight interlink between costs and

money and things like that.

I thought I'd weave that one in.

um so i guess really we need to

get on quen um asap i think is

the the summary save your anthropic money

go and give it to alibaba because they've

pretty much stolen everything from

anthropic anyway so don't even need to you

don't need to be spending all that cash

now on uh anthropic you can now invest

that into a five cents price point with

uh with using quen and um

save that cash so you can buy a

MacBook.

Exactly.

That money has to come from somewhere,

John.

The extra two,

three hundred pounds has to come from

somewhere.

So might as well stop using the Opus

models and use the Quen models.

I have used the deep seek models before.

Absolutely fine.

But I haven't used them in anger as

much.

Perhaps it's time for me to have a

look into the Quen models.

Yeah, to be fair, I haven't used.

I do know people that have been,

but I haven't used it either.

cool so the next episode i think we

said though we were going to break down

a story and we've cheated and we've

basically done news again i'm pretty sure

in the last episode we said next episode

we're gonna you know go deep on a

topic and then we've like next episode

arrives and we haven't so apologies for

that that's just us uh being a bit

short for time but i think next time

we are going to get into a very

specific topic i think we're talking about

um

the trust of an AI agent and what

goes into it and how can you,

from a security perspective,

people are using it today,

What does it really mean when you're

giving it access to things that you have

and what's the imperative risk there to

you?

And then what does it mean to a

business as well when you have not just

you as an individual,

but your entire workforce all in mass

using AI and connecting things up?

And is it quite a risky thing to

just roll out or are there ways of

protecting it so you can get the value

out of it?

So that will be probably what we're going

to discuss next time, all being well.

No promises, though.

But that's where we want to be for

the next episode.

Cool.

Cool.

All right.

Well, a bit short and sweet today,

but we should be back next week with

a bit of a deep dive around Agentic.

Cool.

Thank you.

Bye-bye.