In the Loop AI

It is finally here. Anthropic announced Claude Duet, their 24-7,

always-on, persistent AI agent, which is meant to act like an assistant towards

all your personal stuff and your work stuff.

Now, to be honest, we've seen quite a few of these pop up recently.

We've got GrokBot from Elon Musk, Muse from Meta, and we've got Dots from OpenAI,

all across various different price ranges.

And the big question is, where's Anthropic been? And to be honest,

they've been pretty quiet when it comes to personal or consumer-facing products.

They've always been focused on the enterprise stuff.

But finally, we see what their personal agent looks like. And I got access to

it, early access to it, for the last 48 hours.

I've been playing around with it, and it's been amazing. And I'm keen to share

my thoughts and insights into this pros versus cons, pricing,

accessibility, and all of that.

But what really makes me excited about these agents is the fact that it's becoming

the new way to communicate with the internet. Typically, you and I would scroll

on our phones, navigate to different apps, open up a web browser,

type in an address, Google search, click links, scroll through buttons.

It's confusing and annoying. Speaking to an AI agent and helping it navigate

and do all the stuff for you is going to be where I think the future is headed.

So these AI agents are very important to test. And finally, we have a new frontier one.

Other things I'm going to be talking about in today's episode is going to be.

Earnings. We've got MetaEarnings, Bloom Energy, and a bunch of other companies

that give us an insight into the heavy capex spending that we're gonna expect

in the markets going forwards, as well as new raises from OpenAI and DeepSeek.

Let's get into it on today's episode brought to you by Qualcomm.

So Claude Duet, as I mentioned, is Anthropic's personal AI agent that they're giving towards you.

Now, the difference between this and, let's say, regular Claude that you end

up using every single day is that it's personally tailored for you and it's

meant to be proactive. So let me explain what that means.

It's meant to effectively kind of read your mind or understand what you might

need to get resolved or solved the next day.

So it works similarly to some of the other AI agents that we've reviewed on this show before.

It is an instance of Claude, the model Claude AI, that has access to its own personal computer.

Now, this personal computer looks like your laptop, but it lives in the cloud.

And what it can do with that laptop is pretty much anything and everything that you can do.

It can download apps. It can scroll through websites. It can see and watch videos.

It can do pretty much anything you can do. And that's powerful because it could

end up being a really helpful assistant that can take a load of work off of

you, both on your personal and professional life.

But also it could work 24-7, which means when you go to bed, this thing keeps going.

It kind of reviews all the context learnings that it's had about you,

and it improves, it self-improves over time.

I actually wrote a tweet about this. Anthropics launches their own personal

agent, which competes directly with Metamuse and OpenAI Dots.

The reason why I say this is because it's built in a very similar way.

But what Anthropics Claw Duet has as a distinct advantage is it's built off

of your own context. One of my favorite experiences actually with this agent

was onboarding, which barely existed at all.

I just tapped into the app and it just popped up and said.

Here's what I know about you. And the list that it gave me was very convincing

because of course, it has access to all my Claude account data.

So it's like being introduced to a really familiar old friend that you've known

for decades and decades.

And what I like about it is there's no real handholding that is required when

it comes to interacting with this agent.

And I find that pretty hilarious because my experience with Claude,

I use like Fable 5.1 or whatever the latest models are, it's quite verbose.

It keeps on talking and droning on at length. And listen, I appreciate the knowledge

that it brings to me, but every single response is like an essay to read.

And what I liked about using Meta Muse's product is that it was concise.

That was the first thing that kind of like popped into my head.

I was like, finally, just a succinct answer.

Claude Duet is that for Anthropix product. And I think it does a fantastic job.

But enough about that. Let me actually show you what I've been using it for

and how it's been helpful to me. So I have it on my phone. I also have it as

a desktop access as well.

And I had this demo where I said, hey, can you map me a five mile walk through

Manhattan streets in New York that I haven't done before and include maybe some

weird sites along the way that's not so typical. So I do a lot of walking in

the city, try and get my cardio in.

And I am tired of passing the same neighborhood spots. And I just wanted a new route, something new.

And also maybe some novel kind of like landmarks to see that a typical New Yorker

that has been here for God knows how long wouldn't even know about,

right? I like exploring.

And what it went on to produce was basically an interactive map.

I don't know what it's linked to. It's linked to Google, as I can see here.

And I'm just kind of like, you can see it's easy to move through.

And then it says, listen, And here's a uptown walk of roughly five miles through

the top of Manhattan, where most people never go. Now, I want to point out a

few things here that is different between this and regular Claude and all the

other AI agents that I've used so far.

It goes, I can't see your city strides map.

So I picked the far north tip, which is the part you're least likely to have

walked. Two things there.

Pulled context that didn't live on Claude about City Strides.

City Strides is a separate app that I use to track.

It's my goal to walk every street in Manhattan. Good luck with me for that.

But it realized that that app

might have useful context and data towards answering this question for me.

And I didn't tell her about City Strides. I haven't told Claude,

my regular account, about it at all.

But because it had access to my email account, it was able to have the necessary context about this.

Now, you'll notice it says, I can't see your City Strides map.

That's because I haven't logged it into an account. And there's a debate that

I'll get into in a second about whether you want to give it access to it.

But it's nice to know that it had that context so I could pick up maybe an idea

about, hmm, maybe I will give you access to this so that you can give me a better answer and give me

a path through streets that I haven't quite crossed, but would check off new

points on my City Strides app. I thought that was pretty cool.

The second thing here is I picked the far north tip, which is the part you're

least likely to have walked.

So it has context and awareness of where I've walked before.

And what I like about that is it's orienting itself around where I am and giving

me the best possible input to where I can have. It's intuitive and it's instinctive.

And I haven't really seen this from any other personal AI agent,

maybe Metamuse actually.

But again, what I'm realizing is these agents are really dependent on the context

and data. Now, I just got back from Japan and I am a Pokemon fan. I'll admit it.

I enjoyed trading the cards as a kid.

I've watched all the shows. I love it. And so what I've recently been getting

into is collecting some of the older cards. It's very nostalgic.

It's kind of like a fun hobby that I picked up.

My fiance thinks it's a very expensive one, but hey, I'm just playing around.

And I asked for what a Shadowless Charizard is going for right now.

For those of you who aren't Pokemon connoisseurs, a Shadowless Charizard is

a rare version of the original Shiny Charizard card. It's like one of the big

Pokemon back in the olden games.

And it pulled on recent sold prices. It gave me a rough range,

but I didn't want a rough range. I wanted actual live listings to prove that

I actually knew what data it was pulling on.

And so I said, great, can you pull up any eBay listings and display them here?

And it went on to do exactly that.

And what I found was when I checked on the market rates online,

that was exactly what it was meant to be.

The next feature that I'm about to talk about, which is arguably the most exciting

part about Claude Duet, makes it more valuable.

And that feature is, it's incredibly proactive.

And what I mean by that is it kind of reads your mind in advance in a way that

I haven't really seen with any other personal AI agent, maybe bar Muse if I

had to, you know, put a stake in the ground.

Now, Claude Duet kind of has an insight unlike any other AI agent.

And so what I mean by that is it gets access to like maybe your Gmail account

or if you give it access to your Claude history, which is automatically baked

into, which is actually quite valuable for it.

And so it starts to suggest tasks or things that you might want it to help you do.

And what I found is the big problem with AI in general, and for the majority

of people actually, is it doesn't matter if there's a flashy new model that

comes out if you don't know what to use it for.

Now, the big flashy thing that's gone viral over the last six to nine months is coding.

But 99% of the world don't code or know how it would be useful to them to code things.

So they need to be kind of told, hey, you can use this to do A,

B, and C now, which is like magic. And you never got access or were able to do that before.

This is what Claude's personal agent Duet can do for 99% of the rest of the

world and why I'm so excited about AI agents in general.

So going back to the proactiveness, there's a tasks tab where you can basically

tap into it and it shows you overnight as you've slept, when you wake up,

things that it's thought about.

It's learned about you, your habitual things. And it's like,

hey, maybe you might want to cook up a cool dinner tonight. And I actually looked

into a few recipes for you because you've talked about A, B,

and C in your emails to your fiance or whatever that might be.

And I think you might like this. and I actually decided to do the grocery shopping

for you. It's actually in your Instacart thing. If you wanna click buy, go for it.

If you don't, totally cool, I could just exclude this, but just being helpful for you.

The other cool thing is from a professional standpoint, it's gotten really good

at understanding the types of things that I want to learn about at different times of the day.

So without prompting it, it gave me just kind of like a brief or a roundup in

the morning about all the hottest things that's happened in AI.

So the idea here is the proactiveness leads to a very familiar magical feeling

where it's like, wow, how did you know that I wanted that? That's really cool.

Yes, I didn't know you could help me do that. Help me do that thing.

And so this proactiveness is to a level that I haven't seen compared to other

AI agents. Let's, I guess, get into a bit of a comparison here.

When I've been using Claude Duet and I compare it to GrokBot,

Claude Duet is a really cool hybrid of personal AI agent versus work AI agent.

So it kind of like balances both really well in a way that isn't too polarizing.

It's not like I'm talking to like, you know, two different people.

It's just a very uniform experience.

When I use GrokBot, it is a much more work-oriented thing. They spin up a million

bots to go work on a task for seven hours at a time. And that's all well and

good, but I wouldn't really use it for all my personal stuff as it stands today.

And then when I compare it to OpenAI's Dots, I think they're kind of gathering

their feet underneath them right now. It's a good product, but it's been a slow

rollout, limited release in terms of usage capacity. So I'll reserve my opinion on that one.

And then very quickly on Metamuse, this is the most alike. And I've been using

Metamuse for a lot of personal and professional stuff.

I think that Metamuse has the winning seat when it comes to personal stuff.

And it doesn't surprise me because Meta is known for building personalized social

media products. That's their bread and butter.

That's how they create some of their award-winning products.

It doesn't surprise me that they're better at it than that.

But when it comes to an all-round use agent, Claude Duet is pretty damn good.

Now, what does it look like when it comes to pricing? This is a very important

thing to discuss because some cases you don't want to be forking out hundreds

of dollars for a brand new product. Do you need to get a new subscription?

Blah, blah, blah, blah, blah. Now, Claude.

Is in limited release right now for pro and max plan users. So the cheapest pro plan is $20 a month.

The max plan comes in at around, I think, $200 a month, which is very extensive.

But if you want to get access to this thing, if you have the $20 a month subscription,

you should be able to sign up for a waitlist. And the idea here is Anthropic

is going to roll this out to many, many more people over the coming days and weeks.

Now, comparing this to the alternatives, Grogbot, I believe the cheapest accessibility

point is also $20 a month. So it's competitive there.

OpenAI's Dots is the most expensive. I think you need to be a max or their version

of a max plan user, which is 150 to 200 bucks a month.

I think OpenAI is struggling with compute right now, which is the reason why

they're slow to roll it out.

Hopefully, or I'm guessing the idea is that they roll this out to many more

individuals that are actually gonna find a lot more use to a personal agent,

which is probably the free slash lower $20 a month tier.

But then the most competitive here is still Metamuse, which comes in at a price of $0.

It's free to use. So if you're trying to debate whether you want to use a personal

AI agent at all, go check out Metamuse and play around with that. If you want to...

Familiarize yourself, you're kind of like, okay, cool, I really want to kind

of like lean into this and really understand the magic of this thing.

Go try out Claude Duet, sign up for that waitlist. It's a fantastic product.

Now, of course, probably one of the red flags that you're voicing out aloud,

and I can't quite hear you right now, but I'm guessing is how on earth can I

trust Anthropic and Claude with this thing?

Now, they haven't really given us insight into the architecture of how it's been built.

But my intuition, my guess is they've built it similarly to how Meta has built

Muse. Claude Duet runs its own virtual machine, which lives in the cloud.

So this is the computer that I mentioned earlier. It can kind of like run,

download, do whatever the hell it wants.

And that should live in a secure box or sandbox, which Anthropic presumably

doesn't really have access to it.

Now, when it comes to passwords, when it comes to Claude Duet paying for things,

when it comes to Claude Duet doing anything that would be considered major,

like accessing an important file or writing something or clicking send.

It won't do it without your explicit approval.

And that's baked into the safeguards, as it says, and it was made in the announcement as well.

Now, when it comes to sharing your passwords or seeing your passwords,

I believe it doesn't actually have the capability of doing this,

which is exactly the same experience that you have with MetaMuse as well.

Then the final thing I'm gonna say about this is more of a wider thesis or observation

that I'm having on personal AI agents in general.

I think a lot of people will hear the term personal AI agent and think,

ugh, I don't even know what that is. I don't even know why I would want to use

that thing. Don't I just use my LLM chatbot?

Won't I just use Google Gemini? Why the hell would I care about this thing?

Personal agents are going to be the new way that you interact with the internet.

Now that's a big statement because typically you would need to navigate through

a browser or you would need to specify the device that you are going to interact with it on.

And now you can just remove all that manual labor, the clicking,

the scrolling, the, is it this page? Do I submit this document?

Do I, is the information here or should I look there?

And you just outsource it to this super intelligent being that is working in your favor.

And that's the distinction between whether it is a regular clawed chat window

and whether it's your own personal agent. It's trained off of your data and

context and it self-improves over time.

If you download Claude Duet now or any other personal agent,

you start using it and you're thinking, oh, I don't really know how different

this is from the regular chatbot version, wait a few weeks and give me an answer then.

Because my experience using Metamuse versus using it three weeks later has been

significantly different.

So it's this new world where all the big frontier labs have realized that it's

not just about the smartest model.

It's not about plowing in billions of dollars in training the sexiest,

smartest model that can solve all the mathematical problems in the world or

code something from scratch in a few seconds.

It's about what's more intuitively a better product experience.

How can we get more people to engage with AI and actually get more use from them?

Personal agents is the window to that world. And I would argue that my future vision is.

We'll interact through these things. They'll go out and do all the buying,

shopping, research, browsing for all of us so that we can just be left to,

I don't know, use the internet for entertainment purposes, watching videos,

interacting with our friends.

It will automate all the other manual stuff for us, which I think is very cool.

But the other thing is, imagine having an always-on version of yourself on the internet.

That could sound a little bit dystopian, especially if it's kind of like a digital twin type thing.

I'm still figuring out Now, my thoughts about that, because my mind instantly

jumps to Black Mirror, if I'm being honest with you. And I'm like,

I don't want to end up in a world where there's a digital version of me just living on the internet.

But I think we're moving some steps towards that. And I don't necessarily think

that would be a bad thing if it ends up making each of us.

More productive or having a more enjoyable experience. And that was my whole

takeaway with Metamuse, by the way.

And I keep mentioning Metamuse, not because I love the product,

although I do enjoy it a lot, but more so to point out that that's where I've

had the most experience with.

And I'm kind of like parlaying that into my Claude Duet experience,

where I'm guessing I'm going to have in a few weeks time.

But the idea here is that it's become a sort of default app that I now use.

Which I haven't really seen in a long time. You know, we all use our email apps.

We all end up going on the browser tool or whatever that might be.

But what I found is these personal agent apps have taken up more of my time

whilst giving me back more time.

And what I mean by that is it takes up more of my time when I'm scrolling on

my phone. I just go onto the app and I talk to it or whatever that might be.

But it also reduces the time that I'm spending on other apps.

So it's kind of like optimizing or making things more efficient or for whatever

reason, I'm being more drawn to using that thing because it ends up being a better experience.

And honestly, this could be quite dangerous or an insight into the competition

that these businesses are going to face.

What happens when you don't need to navigate to a website? What happens to advertisers

when your eyeballs aren't going to be on Google search anymore on their sponsored

link results? It's just going to be on your AI agent.

These advertisers need to start then pitching these AI agents.

I don't know. This deserves its own episode. And honestly, at this point,

it's become its own episode. I know this was meant to be around it,

but I don't know. I just, I enjoy talking about this stuff. Um,

but yeah, that is called duet.

Go and check it out. Okay. Because I've been talking for too long,

I'm going to give myself room to speak about one more, one, maybe two more topics

very quickly, because I think it's important to raise and I don't want a week

to go by before I mention it.

So let's continue on today's episode brought to you by Qualcomm.

So I saw this tweet that OpenAI is raising money.

And that struck me as weird because I was like, aren't you going to be IPO-ing

soon where you're going to be

raising a lot of money? Why are you raising money in the private sector?

Turns out that's apparently what's happening and they're raising $30 billion.

The rumor says that Kotu is leading it with some investors from the UAE and

money is coming in or the money will be used to purchase extra compute,

which is kind of the whole point about why I want to talk about this.

The CapEx spending is getting absolutely insane. And we're seeing this not just

from OpenAI, but also DeepSeek.

The Chinese adversaries or the big AI lab DeepSeek has been raising around $15 billion.

I think they recently closed it at a, I think it was a $70 to $90 billion valuation.

And I think the broader reason for this comes down to AI CapEx spending.

And I want to address this for a moment because when I look at the markets,

we've been bleeding for a little bit now.

It's not been too long, but it's been about like a week and a half,

which given the absolute bull run that we've had over the last year and a half,

people are getting concerned.

And they're questioning whether we are in a bubble and whether this thing is

going to pop anytime soon.

My question is, are they taking on any leverage or financial risk to do all

of these things. This is the largest amount of capital raise that we've seen

concentrated to such few companies.

By the way, Elon Musk is still doing the same thing as well.

I think even though SpaceX has gone public, he's also rumored to be raising

another $40 billion to do the exact same thing, buy as much compute and GPUs as they can.

The idea here is it's extremely concentrated on one specific sector for one specific thing.

And that concerns a lot of people, I think rightly so, because we don't know

or have enough proof yet that the compute will result in many great and major things.

Now, if you ask the optimist guy in me, I'm like.

This is going to result in an optimistic future, because I think AI is going

to transform so many different things. But for the bears, for the critics,

I completely see your point of view.

And my take is, if these companies, specifically Anthropic, OpenAI,

DeepSeek, are doing this on a levered basis where the credit financing is bad.

Then this is something we need to keep an eye on.

It is a big bet, but they don't have as much money to spend.

They don't have billions of dollars in the cash account to spend versus some

of the more mature companies.

And I say that because Meta had their earnings as well.

And without going into the entire earnings that Meta had, they're doing a stock

split for something known as class C shares, which is non-voting shares.

But the idea here is they should be able to raise a heck ton more money to,

you guessed it, spend on AI CapEx.

And Meta from their quarterly earnings has finally gone into negative cash flow on their spending.

Now, Meta has been the most aggressive spender, or Elon Musk probably beats

suck, but the second most aggressive spender when it comes to acquiring GPUs

and compute, for the same reason that I described.

They believe that having more compute allows them to build a better AI model.

And I would disagree with that if it weren't for Meta mounting the most impressive

comeback over the last 12 months, where they went from being the worst Frontier Lab provider,

when they should have been one of the top to one of the near top Frontier Lab providers.

And it's not because they have the best model yet. They're training multiple

big, big models, but they have one of the best AI products out there.

One of my favorites, Metamuse, right?

So the idea here is there's just this race to coagulate and capture as much

compute as they can without necessarily knowing what the end product is going

to be. And I'm just going to call it out as what it is. It's a big...

It's a big risk. And intuitively, directionally, I think that we're going to

get somewhere good where AI is going to transform a lot of different sectors

that we can't even conceive of right now.

It's going to result in better materials being manufactured that will result

in physical infrastructures that are really cool.

It's going to solve a lot in medicine and science and physics and maths that

will unlock so much new innovation. But again, it's hard to believe or buy into

this when we can't even conceive what those tangible outputs are going to be.

The point I want to make around here is I want to call it out.

There's a lot of money being levered up to spend next year. 2027 is going to

be an insane year for CapEx spend.

It's breaching right now. The estimates from JP Morgan, I believe,

is like $1.2 trillion, just a loan on CapEx compute.

Majority of that is going to NVIDIA for their GPUs.

A bunch of that is also going to custom ASICs as well.

It's this very important moment in the buildout for AI that looks very different

to six months ago, where we were like, okay, they're spending a lot,

but we're not levered up or we're not borrowing at crazy rates.

And now we're starting to enter that territory. Now, I'm not calling this a

2008 financial bubble with CDS, credit default swaps.

I'm not saying this is the dot-com era bubble. But I just want to note that

we are at the tipping point and where we go from here is going to be very important.

Honestly, this deserves its own episode, but I've been speaking for too long,

so I'm going to shut up. But I hope you enjoyed this episode.

Go try out some of the personal AI agents. I enjoy them thoroughly.

So many new listeners and subscribers, welcome. It's so lovely to see you.

The support we've been seeing on some of our recent episodes has been amazing.

I've been loving all your comments.

I know some of you have been long-time listeners, but you haven't left a comment.

Leave a comment. It's always lovely to read them. And I reply to almost every

single one of them. I have no life. If you have any thoughts about things that

I haven't covered or you want to get my insights on, let me know.

But yeah, thank you so much for listening and I'll see you on the next episode

of In The Loop brought to you by Qualcomm.