Latest / In The Loop / The Top Four AI Trends & Predictions Of Summer 2025

Transcript

The Top Four AI Trends & Predictions Of Summer 2025 · In The Loop

0:00
There's a series of really interesting trends and ideas
0:02
that are emerging in AI right now, and as we start to enter
0:05
the end of summer, I think it's an excellent opportunity to
0:08
summarise them today. AI companies are going to have
0:10
to start paying to read websites.
0:13
Investors are starting to trade computing power just like they
0:17
might with oil. We're now seeing major brands
0:19
launch websites just for agents, and companies with just twenty
0:23
employees are now worth more than those with 2000.
0:26
None of these things were happening even 36 or 12 months
0:30
ago. Each story may seem isolated,
0:32
but when you put them together, they point towards something
0:35
bigger. The basic assumptions about how
0:37
a digital economy works are now being challenged.
0:40
In today's episode, we're going to discuss the biggest ideas and
0:43
trends that are happening in the world of AI in the summer of
0:46
2025. My name is Jack Horton and this
0:49
is in the Loop. I hope you enjoy the show.
1:02
Let's start with something that sounds really simple but is
1:05
incredibly important for the future of AI, and that is called
1:08
ambient agents. Right now, when you use AI,
1:10
there's a really simple and obvious start and end to a
1:14
conversation. So you open ChatGPT, you ask a
1:16
question, you get an answer, you close the tab, and then it's
1:21
done. But that's changing really,
1:23
really fast. Now obviously this idea that
1:25
agents do many, many things for you isn't new, but this idea of
1:29
an Ambien agent is becoming extremely popular.
1:32
Now, really funny story before I kind of explain what Ambien
1:35
agents are, as you're building in the cutting edge of AI, and
1:40
this happens quite a lot and it's happened to us a number of
1:42
times. You build something, you are
1:45
trying to think of what this thing is going to be in the
1:47
world because it's so new. And then you realise, almost
1:50
like the advent of fire spreading around the world
1:52
independently, that a number of other people are starting to
1:55
cotton onto this thing at the exact same time.
1:58
We were actually going to be dubbing this the Meta agents.
2:01
Then suddenly I find a paper or an article or a thing that's I
2:05
guess announcing this new idea in the world that's really,
2:08
really new. And in this case, it's been
2:09
ambient agents. And the reason it's called
2:11
ambient agents is because it works continuously in the
2:14
background for you. So it can monitor information,
2:18
it can take actions, or just provide you with recommendations
2:21
on what to do next without you having to think about it being
2:25
there. Now, to really understand why
2:26
this shift matters, we need to understand what they actually
2:30
are and why they're different than the type of AI tools we use
2:34
today. Because obviously traditional AI
2:36
works on that request response cycle.
2:38
You make a request, gives you a response, and then it's over.
2:41
Ambien agents are a little bit different.
2:43
They're an event driven system, which basically means that they
2:46
monitor different data streams and they will then take an
2:49
action, whether that's take an action in a separate system or
2:53
come to you and give you a recommendation based on the data
2:57
it has. So if I give you an example
2:58
about exactly what we were building at Mindset, our
3:01
customers have massive websites and technology platforms with
3:04
hundreds and hundreds of thousands of users.
3:06
And let's say they build AAI learning agent or a search agent
3:11
or a web agent for their website to recommend things to their
3:14
users. Often the importance of getting
3:18
the configuration right and the tools and actions it can take
3:22
right take a lot of work. Now we were launching Meta
3:25
agents, which are essentially a set of agents that monitor all
3:28
other agents on the platform to manage their performance,
3:32
understand when things that say the configuration or set of
3:35
tools they have access to needs to tweak or change.
3:39
And this all works by basically those meta agents having access
3:43
to a, let's call it a data pool. There's Big Lake of data that
3:46
it's collecting and we set those agents up to specifically
3:51
monitor certain things that are happening.
3:53
So we might tell it to take note of user behaviour or the
3:58
question compared to the agent configuration and the agents
4:02
purpose. So by telling it what to monitor
4:05
in that data set, it can then take an action as a result.
4:09
Now, this is a massive difference.
4:10
It may seem quite simple, but it's like having let's say an
4:13
assistant who just will sit at their desk and waiting for you
4:15
to come to them. Versus one that is checking your
4:18
emails, your calendar, figuring out what projects you've got
4:22
coming up and potential patterns or maybe problems that they
4:25
foresee, and then making recommendations to you.
4:27
Harrison Chase, the founder of Lang Chain, has been one of the
4:31
really big advocates of this shift and his argument's
4:34
incredibly compelling. Which is to say that the world
4:36
has to move away from these chat based agents that require humans
4:39
to come to them really quickly. Because in his own words, there
4:43
could be thousands, hundreds or even millions of agents running
4:47
in the background all at the same time, monitoring really
4:50
important piece of information for people.
4:54
And the technical architecture behind all this is really
4:56
interesting because they're event driven.
4:59
You can subscribe them to a set of, let's say, data streams such
5:03
as your emails. Maybe it could be people
5:07
visiting your website, it could be user behavior, literally
5:12
anything you can imagine that provides you access to data.
5:15
It could even be a physical device in the world, like a car.
5:18
And these agents of course can monitor all of this information
5:21
across all these digital spaces, which is both dark and and
5:24
really interesting. The agent then must be able to
5:27
store that information to its long term memory, which we call
5:29
context. This for mindset is a set of
5:33
variables. So you just tell it what to
5:35
remember and it will remember it.
5:37
And as a result, agents can remember not just what data it
5:40
has access to, but all the data that it's previously collected
5:44
and it believes is important. Therefore relating current
5:47
information to ongoing projects or previous decisions or maybe
5:51
future objectives and strategies.
5:53
As long as that information exists in its memory.
5:56
And as I say, when a new event arrives, the agent can be
5:59
configured to be told to let's say, escalate certain things to
6:04
a human under certain conditions or just simply log the
6:07
information for future reference.
6:09
And that's where this ambient part becomes really important.
6:11
And it's why we called it the Meta agents because they're
6:14
always sitting behind the scenes monitoring all that important
6:17
information to make recommendations or take actions
6:20
when necessary. And This is why Landgrass CEO
6:23
even said that this idea of human in the loop is going to
6:26
have to change. And we really agree with that
6:28
because otherwise you're just going to have agent pinging you
6:31
for requests all of the time. So actually what we need is an
6:35
agent inbox. So this agent management
6:37
dashboard where you can see all the things your AI agents are
6:41
working on and all the tasks they've accomplished.
6:42
So you can review their decisions, you can provide input
6:45
really quickly. And again, it's just very funny
6:49
because as all of this news and information has been coming out,
6:52
Mindset is quite literally launching or relaunching our
6:55
administration portal for managing agents.
6:57
And we're calling it the Agent Management Studio, where you'll
7:00
be able to see every single agent that you have live, all
7:03
the information that's coming to it, all the interactions it's
7:06
been having, and basically keeping tags on an entire team
7:10
of agents that you've deployed in the real world.
7:12
And some of the real world's use cases here are really
7:14
interesting. Think about identifying, let's
7:16
say, learners that are struggling to learn, or people
7:20
that are becoming disengaged, or angry customers or people that
7:23
maybe are unwell. This year, a massive financial
7:26
services company launched A compliance auditor agent that is
7:30
quite literally scanning and monitoring all transactions and
7:33
internal communications for potential policy breaches.
7:36
And it can flag any suspicious behaviour and automatically
7:39
draft incident reports. Now, from a technologist's point
7:42
of view, I'm incredibly excited about this and it has the
7:45
potential to really shift the way we work and the way that
7:49
work truly gets done. However, I can't help but add a
7:53
level of nuance to this trend that's emerging because this
7:56
isn't just about productivity, it's ultimately a version of
8:02
always on capitalism. We're moving towards this world
8:04
where human rest is becoming almost an inefficiency that's
8:07
being optimised away by machines that just don't need brakes.
8:11
In a world where our data, and I guess our security over that
8:17
data, is going to become more important than ever as agents
8:20
can make so many decisions so quickly that the mere humans
8:25
like us just cannot even comprehend it.
8:38
Anyway, let's move on to trend #2 and that's pay per crawl.
8:42
For 25 years, the Internet has operated on a very simple deal.
8:48
Content is completely free to access for everyone, and
8:51
everybody benefits from this open exchange.
8:53
As a result, search engines crawl websites.
8:57
People then search and click through and read articles and
9:00
websites, and the whole system works because of that flow of
9:03
traffic. And AI is completely broken, or
9:07
at least shall I say, starting to break the system.
9:10
And the reason is because AI can scrape millions of web pages for
9:14
training data and providing answers, but they don't send
9:17
users back to the original sources.
9:19
So instead of clicking through to read a full answer, people
9:22
get their answer from AI and then that's it.
9:25
And this results in websites losing massive traffic while
9:29
massive AI companies can profit from all of their content.
9:32
And the reason this is so important is because if nobody
9:35
is creating any new content because they're not being paid,
9:39
then what will AI be trained on? And the numbers are quite crazy.
9:43
Google's crawl to referral ratio.
9:46
So that's the amount of times it crawls the websites to.
9:50
The number of times a person then clicks on a website has
9:53
gone from 6 to 1. So crawling to referral 6 to 1
9:57
to now 18 in one in just six months alone.
10:00
And for other AI systems like ChatGPT, it's much worse.
10:03
That ratio is 1500 to 1, meaning basically 0 traffic is returned
10:08
to the websites that are being scraped.
10:10
This has triggered a huge change in how the Internet works and as
10:15
a result, in July 2025, Cloudflare.
10:17
So for those that don't recognize that name, it's the
10:20
are you human tick box has launched a paper crawl.
10:23
So that's a system that lets a website owner charge an AI
10:27
crawler for access to their content.
10:29
So instead of this traditional choice between saying allowing
10:32
or blocking crawlers, publishers can now set a price per request.
10:37
So when NAI might try to access their content, their website can
10:41
then say, Nope, that's going to cost you this amount of money
10:43
and therefore charge accordingly.
10:45
A lot of publishers have obviously signed on.
10:47
We've seen Reddit start to sue Anthropic for scraping all of
10:50
their data. And I guess the message is
10:52
incredibly clear right now in the Internet world, which is pay
10:55
or get blocked. Now there are many critics that
10:57
argue that this will create, I guess, a very dangerous
11:00
precedence. For example, Tech Radar actually
11:03
published a very detailed and a really good analysis which broke
11:07
down and explain how Paper crawl will only realistically benefit
11:12
maybe the top 1000 websites in the world.
11:15
Who has that much traffic whilst those small sites basically get
11:20
nothing. What's worse is that only a
11:22
handful of companies will ever be able to afford those data
11:26
licences, the Googles, Microsoft and open a eyes of the world.
11:30
So that means start-ups or any small organization is
11:33
immediately locked out. And what this ultimately
11:35
represents is I guess the collapse of the foundational
11:38
promise of free access to information on the Internet.
11:42
And we're watching this, I guess, transformation of shared
11:46
knowledge into private property. And now this makes me incredibly
11:50
sad, but I understand and respect why it's happening now.
11:54
There are optimistic paths forward, you know, if we can
11:57
create, I guess a Spotify for data type of subscription
12:01
service where developers just pay for access to this vast
12:05
licensed content, ensuring the creators get compensated whilst
12:10
maintaining that broad access to information for users and also
12:15
for models. I have no idea what the future
12:17
looks like, but what I'm pretty certain about is that the web of
12:19
today is going to look very different to the web of
12:22
tomorrow. OK, let's go to trend and idea
12:36
#3 which is the financialization of compute.
12:40
Now here's where things get a little bit weird.
12:43
Wall Street has actually discovered that the computer
12:46
chips powering all this AI is so valuable that they're starting
12:50
to treat it like gold or oil or any other commodity that they
12:54
can trade and speculate on. Let me explain what this
12:57
actually means. You know, how investors can buy,
13:00
say, oil futures. So trying to predict the price,
13:04
which is actually betting, betting on the price of what oil
13:06
will cost say 2-3 or nine months from now.
13:10
Well, they're now doing that with computing power and AI
13:13
chips. People are literally trading
13:15
contracts for GPU time like they are for gold and oil.
13:20
You've got start-ups that are now launching the Bloomberg for
13:23
CPU pricing, which is basically a website which monitors stock
13:27
prices for all the different type of computer chips.
13:30
And another company called GAIB has started issuing tokens that
13:35
basically represent claims on revenue from specific GPU
13:38
clusters. So that's like a factory of
13:40
GPU's running in the background. So you can literally own a piece
13:44
of a data centre's earnings, like you might earn a or own a
13:48
share in a factory. So you're betting on the
13:51
utilisation or total usage of a data centre and all the computer
13:56
chips, which is basically betting on how successful tech
14:00
companies and AI companies are going to be.
14:02
What this really shows is that compute power is now joining
14:05
housing, healthcare, water and education as something that you
14:09
can't just buy. You have to compete with Wall
14:11
Street investors for. When people start trading for
14:15
something as a speculative investment, it means that
14:17
they're buying it not because they need it, but because they
14:20
think the price might go up. Just like in housing and oil.
14:23
And just like in those places, you get bubbles that burst and
14:27
can completely damage and destroy an entire economy.
14:29
Because this isn't just an investment anymore, it's an
14:31
infrastructure that is basically keeping most of the world
14:35
running without this most technology systems breakdown.
14:38
On the one hand, this could cause a real problem because
14:40
there's a concentration of power in only a handful of companies
14:43
like NVIDIA that control most of the market.
14:46
Whereas on the other hand, many people would argue that having
14:50
this financialisation of computing power might make it
14:54
more accessible and actually more competitive.
14:57
To me, it's just a sign that really our economic system has
15:00
become, I guess, so effective at extracting profit from literally
15:05
every single thing. It is now trying to extract
15:07
profit from computational power that hasn't yet been used by a
15:12
tech company. So let's move on to #4 which is
15:26
the websites and products are being rebuilt for agents and no
15:30
longer being built for humans. And you could call this the
15:33
agent experience. So AX for short, just like UX.
15:37
And it's a big change in how people build technology.
15:39
For decades, the progression has been really clear.
15:42
First people designed for technical users for moving wires
15:46
around and using ones and zeros for the early day computers.
15:50
Then it was focused on user experience and Apple bringing
15:53
technology to the masses. Over time that experience has
15:57
even been optimised for developers with APIs.
16:00
And now we're adding this new dimension which is designing for
16:04
AI. What this really shows us is
16:05
that in an AI powered world, human leverage can be amplified
16:09
to an unprecedented degree, but that amplification just isn't
16:13
distributed equally. It's concentrated and will
16:16
continue to concentrate amongst those with very specific skills
16:20
that will be really effective in an AI future.
16:24
It's why I advocate for people to learn more about AI and for
16:27
people to tell their colleagues and friends to do the same.
16:30
As we wrap up, let's consider what these trends tell us about
16:33
the world. Because we're living through a
16:36
time where the fundamental assumptions about work,
16:39
ownership, competition, and value creation are all changing
16:43
at once. These aren't separate trends.
16:44
They're all part of the same transformation, which is the
16:48
emergence of an economy where AI doesn't just help people with
16:51
work, but does it for them. The optimistic view is that
16:55
we're building the foundation for an incredible future filled
16:58
with prosperity and potential. It's an incredibly exciting time
17:02
and is why every single one of us must be learning more about
17:06
AI. Anyway, thank you for listening
17:09
to today's episode. I'll see you next week.