2025 AI Breakthroughs: Code and Images Unlock
Key Points
- 2025 didn’t bring sensational sci‑fi AI, but it clarified where real value lies in the AI revolution and exposed critical gaps that are now visible.
- The breakthrough that most exceeded expectations was allowing LLMs to use code as a tool, unlocking agentic workflows and making AI accessible to non‑technical users through plain‑English computer interaction.
- A suite of emerging technologies—cloud‑code integration, model‑context protocols, skills, Codeex, and tools like Cursor—combined to let anyone manipulate files and automate tasks simply by describing what they want.
- Advances in image generation proved equally transformative, enabling detailed infographics, layouts, slides, generative UI for games, interior redesign, fashion, and more, effectively turning visual creation into a new, powerful AI capability.
Sections
- Code‑Enabled LLMs Unlock 2025 - In a 2025 recap, the speaker emphasizes that the most surprising breakthrough was allowing large language models to use code, a development that dramatically expanded agentic workflows and made advanced AI capabilities accessible to non‑technical users.
- Generative Interfaces and System Design - The speaker argues that while generative AI can create adaptive web interfaces, practical value lies in balancing novelty with familiar habits, and success often stems from system design skills rather than specialist AI developers.
- The Undervalued AI Middleware Layer - The speaker argues that the middle tier—transforming raw model outputs into structured, domain‑specific workflows—remains underbuilt yet holds immense value, and startups are now confidently building in this space despite concerns about hyperscaler dominance.
- AI Shifts Toward Creativity and Quality - The speaker highlights a recent pivot from viewing AI solely as a cost‑cutting tool to embracing its role in enabling creative expression and enhancing human expertise, emphasizing new opportunities for interdisciplinary collaboration and a focus on delivering higher‑quality outcomes.
Full Transcript
# 2025 AI Breakthroughs: Code and Images Unlock **Source:** [https://www.youtube.com/watch?v=uVZMc-i5EEs](https://www.youtube.com/watch?v=uVZMc-i5EEs) **Duration:** 00:13:52 ## Summary - 2025 didn’t bring sensational sci‑fi AI, but it clarified where real value lies in the AI revolution and exposed critical gaps that are now visible. - The breakthrough that most exceeded expectations was allowing LLMs to use code as a tool, unlocking agentic workflows and making AI accessible to non‑technical users through plain‑English computer interaction. - A suite of emerging technologies—cloud‑code integration, model‑context protocols, skills, Codeex, and tools like Cursor—combined to let anyone manipulate files and automate tasks simply by describing what they want. - Advances in image generation proved equally transformative, enabling detailed infographics, layouts, slides, generative UI for games, interior redesign, fashion, and more, effectively turning visual creation into a new, powerful AI capability. ## Sections - [00:00:00](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=0s) **Code‑Enabled LLMs Unlock 2025** - In a 2025 recap, the speaker emphasizes that the most surprising breakthrough was allowing large language models to use code, a development that dramatically expanded agentic workflows and made advanced AI capabilities accessible to non‑technical users. - [00:03:31](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=211s) **Generative Interfaces and System Design** - The speaker argues that while generative AI can create adaptive web interfaces, practical value lies in balancing novelty with familiar habits, and success often stems from system design skills rather than specialist AI developers. - [00:07:18](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=438s) **The Undervalued AI Middleware Layer** - The speaker argues that the middle tier—transforming raw model outputs into structured, domain‑specific workflows—remains underbuilt yet holds immense value, and startups are now confidently building in this space despite concerns about hyperscaler dominance. - [00:11:45](https://www.youtube.com/watch?v=uVZMc-i5EEs&t=705s) **AI Shifts Toward Creativity and Quality** - The speaker highlights a recent pivot from viewing AI solely as a cost‑cutting tool to embracing its role in enabling creative expression and enhancing human expertise, emphasizing new opportunities for interdisciplinary collaboration and a focus on delivering higher‑quality outcomes. ## Full Transcript
You know, 2025 didn't deliver the
science fiction version of AI that gets
lots of clicks, but it exceeded my
expectation in ways that matter more. It
clarified where value is actually coming
from in the AI revolution, and it made
the gaps that we still have visible in a
way that I think is useful. The number
one thing that I want to call out as we
sort of look back at the year 2025
is I think that we all or almost all of
us underestimated how powerful it is
when you allow an LLM to use code as a
tool. That turns out to be an absolutely
massive unlock. It's at the heart of a
lot of our agentic workflows these days
and it is on the verge in 2026 of being
one of the biggest places where non-code
non-technical folks can lean in and
start to figure out how agents can work
for them in ways that are far removed
from writing code. And I think that that
that core unlock of wow an LLM can work
with code therefore it can work with any
part of the computer. That was one of
those pieces that you could see that
vision as something that the model
makers had in their heads at the
beginning of 2025 and they did talk
about it but it wasn't realized. It was
very much a someday we will see this
come true and then over the course of
the year you know we get cloud code we
start to get model context protocol
starts to dominate we start to get
skills we start to get codeex and you
start to gradually see these pieces come
together you see the development of
cursor etc. And what you start to see is
that these pieces allow everyone, not
just technical users, to start to use
the tool because now plain English
allows you to talk with your computer.
Now plain English allows you to
manipulate the files on your computer
any way you want. Absolutely massive
unlock. And I think it is hard to
correctly estimate going in how big that
turned out to be. The second one I think
that turned out to be absolutely massive
is images. And people have talked about
this, but we have lived in a world for
most of 2025 where images were getting
better, but text remained the most
accurate way to work with LLMs. And in
that world, code is just a subset of
text, right? Accurate code is just
another language that the LLM has to
learn. But images are how humans process
information quickly. And when we finally
got to the point where we had images
solved, where you could do detailed
infographics, you could do detailed text
and image and it wouldn't look weird.
You could do full maps, you could do
layouts, you could do slides. I don't
think we realized how big of an unlock
getting that right was. And that doesn't
just mean we got powerpoints, right?
Like we did solve powerpoints, but think
of it more broadly than that. Think of
it as generative UI for gaming. Think of
it as you have a chance to redecorate
the rooms in your house now in ways that
you never did before. Uh there's all
kinds of personal applications that
people are building on with fashion that
come with this. Basically, getting
images right enables us to realize a
vision of graphical user interfaces that
has always been out of reach. the idea
that the graphical user interface isn't
locked to your screen and locked to what
the developer says. That the graphical
user interface becomes something that
evolves with you. Maybe you end up with
it as a wearable. Maybe you end up with
it as a combination of phone and laptop.
Maybe you end up with it being
generative in the side like it is for
Comet. Those are all variants of this
idea that with the right graphics solve
from an AI perspective, that whole
surface becomes continuous. and you can
evolve the surface of engagement
digitally with where an individual is at
and what they're looking for in that
moment. Now, we're not all the way there
yet. I'm not here to tell you that the
future of the web is everybody has a
generative interface that changes for
everything that they need. There is a
value in habit. There's a value in
steadiness for common use cases. We
don't need to reinvent the wheel where
it's incredibly obvious. But I do think
solving images is one of those things
that we probably underestimated how big
a deal that was going in and like
looking at it on the other side a month
or two in it's absolutely massive. Like
it's it's huge. What are some other
ones? I think those ones are pretty
common reflections. Let's go a little
bit deeper. I think one of the things
that surprised me is how far you can get
without AI developers. So we we were
told at the beginning of the year AI
developers are everything. Like you need
to have a developer that knows AI. I
actually think you need someone who can
design systems. I have watched
individuals outexecute entire
development teams because they treated
engineering as a workflow they could
design and they didn't worship at the
altar of a particular model. And so they
were okay building low tech things like
templates, like validators, like
retries. And they iterated really
quickly and aggressively. And they
didn't confuse Agentic with good. And
before you know it, those folks are
understanding the new principles for
evolving agentic systems. And then they
become more and more valuable. And and I
think we are going to have to throw away
the idea that there are technical and
non-technical people. I think a more
accurate description is that everyone
picks up the degree of technical skills
they need to solve the problems they're
interested in. And increasingly the
question is going to be, are you curious
about the problems that are relevant in
your domain? And are you willing to dive
in and pick up the AI skills you need to
solve those problems, including
technical skills, because those are
increasingly approachable. You can get a
scheduled task with a nice learning
composed just for you in chat GPT every
morning if you want. And I do I get
them. It's really nice. I get nice
little coding reviews every morning. You
can get it for whatever you want.
Another one that I think positively
surprised me about 2025, verification
loops in Agentic systems turn out to be
incredibly powerful. Uh the idea that
you can measure correctness in different
dimensions turns out to have incredibly
wide ranging implications for good
system design. And that's not super
surprising if you know how software is
designed. But hooking that up to an
agent that iterates is it's like hooking
up a jet engine to an airplane. Like
it's it's amazing how fast you can go
when you stick an agent against a
verification loop that is hard to game
and you say go get it done. And I think
that we figured that out as a community
partway through the year and started
really really practicing it and that has
supercharged our progress. And the nice
thing is that's one of those primitives
that we can really build on heading into
2026. Like getting more into
verification loops is something that I
think we'll see from more and more
teams. And I think we're going to start
to see some consistency around those. So
there's things like accessibility where
you just want to have a standard set of
eval verification loops across the
industry and people just need to get
their agents to pass them when they're
building software and it should not be
something you have to reinvent every
time. So I think we're going to start to
see a really interesting ecosystem build
up around verifications. That's that's
compelling. Another one that has been
really good to see is the messy middle
turned out to be the entire game.
Everyone wants to talk about the idea of
the front end and there was a lot of
talk during the middle of the year about
super model makers or hyperscalers
owning the entire stack like Chad GPT
owning the stack. Cloud code launches
does cloud code own the entire stack. Is
there a competitive advantage to being
in the middle? Is is cursor game over
etc. It turns out that there is so much
value in transforming messy inputs into
structured representations, in routing
intent, in orchestrating calls, in
handling exceptions, in providing useful
user interfaces for specific things.
That the middle layer is still feeling
underbuilt relative to how much value we
can unlock. I think most of us got that
one wrong. And I think that was one of
the pleasant upsides of the year is that
the messy middle, yes, maybe it's
vulnerable. Maybe you worry about sort
of the launch of a particular
hyperscaler's product idea, but there is
so much value in taking raw AI model
outputs and getting them into a
particular domain that we are I feel
like we're really underbuilt. I feel
like uh probably the most prominent
example of that is cursor which everyone
refers to all the time but there's a
whole host of other startups in the in
the space not just in coding and non-
tech spaces too that are are building
aggressively now into the middle because
we've we've realized we don't have to be
afraid of it. We've realized that the
model makers are essentially forming a
very competitive substrate of
intelligence that we can build over the
top of to deliver outputs to users that
are much more valuable than they're
really going to be able to get from
models alone. That's a really
interesting one. Another one that
surprised me or at least it's a positive
reflection from the year. I think we are
realizing now how much value there is in
effectively scoping our workflows. I
think agents were oversold. That was
that was something that a lot of people
were disappointed by because they were
sold as magic buttons. But the flip side
is when you put an agent in a good
workflow, that's a really pleasant
surprise because instead of promising
that it can do everything, it turns out
that when you can use it as a tool, it
can do a tremendous amount reliably and
you can start to really move volume over
to it. And so I think the pleasant
surprise is how much you can accomplish
when you properly harness your agents
and how big companies are leaning in and
able to actually get volume done on that
basis. You know, another positive for
the year, I know that we had a lot of
conversation around AI slop this year,
but I think one of the things that I
learned is that AI slop is a symptom of
unconstrained and unmanaged artificial
intelligence. And that companies that
start to get into marketing, start to
get into producing content at scale for
AI, if you build the right systems, you
can produce really compelling, very
performant ad flows, very performant
email marketing, very performant content
marketing that outperforms what humans
can do. And so I guess the the hopeful
thing there is really we don't have to
surrender to and live in a world with
non-performance slop. we can actually
construct these systems so that they're
beautiful, so they sound good, so that
people want to click on them. And I'm
fully aware that right now that if you
announce that your ad is AI, there's
generally a backlash. I think we're just
going to get to the point where we don't
announce it anymore and we just do the
ad that feels right, however we do it,
whether that's manually, whether that's
manually plus AI, whether that's AI, and
we just get it done and we move on. And
I think that's going to be the case with
a lot of content. And the key and the
measure that we should hold is, is the
content useful? Is the content genuinely
helpful? Is the content information
dense? Is it something that I can come
back to again and again and say, there's
something here that I can dig my teeth
into. And what we're discovering is that
AI can actually be really helpful on
that because it can ground you with
research. It can increasingly do
factchecking. It can help you think
through the structure of a piece. It can
help you with uh generating those ads
now that we've sort of solved the images
piece. And I think that gives me hope
because I don't want to live in a world
with slap and I don't think most people
do. And I think that's going to end up
being a phase and we're going to end up
getting through it because there's a lot
of selection pressure for better
content. You know, another one that
surprised me positively in 2025 is how
quickly the market selected for people
with very strong creative problem
solving instincts. I I think we saw a
quicker response from the labor market
than I anticipated. Not on the dramatic
headline stuff of firing and all of
that. There's actually still not a ton
of evidence that AI is driving overall
job market declines as much as there's
newspaper headlines about it everywhere.
But there is a lot of anecdotal evidence
around the degree to which AI is a
creative or liberal arts endeavor. And I
think that signal has swung really
sharply. And I think that's been really
positive to see technical people that
wanted to express their creative side
and creative people that never felt like
they could be technical finally have a
chance to get the AI they're looking for
and get a chance to build and get a
chance to share their talents and get a
chance to stretch their wings in ways
that they couldn't before. And so I I
actually think that one of the great
opportunities of 2026 is people who want
to grow the edges of their domain
expertise and get smarter and do more.
The world is your oyster. We've never
been able to lean in more that way and I
think that's been really fun. And the
last one I I want to call out is that we
did start to see a shift from a lot of
the cost cutting mentality to quality
lift. It's not universal. There are
absolutely still people who will see AI
primarily as cost cutting. But more and
more and more as I sit and talk with
leaders, they're sort of recognizing
after the first wave of vendor purchases
and getting AI installed quickly and
thinking of AI as a magic button, they
still need their people. their people
can't go anywhere because they still
need their people to deliver the kind of
value that only people can deliver in a
customer-f facing organization. And so
what they want is they want to have a
conversation about quality. How can we
level up the quality of the experience
we provide to customers in ways that
were unimaginable because of AI? How can
we lever up the volume of customers
served? How can we make the price more
competitive because we're able to scale
the unit economics of the business?
Those are so much more interesting
questions and so much more compelling
questions than just saying, can we cut
costs and dump AI? There will still be
folks that have that sort of brutalist
mindset. But I think increasingly people
are starting to recognize how powerful
these systems are. And they're starting
to recognize that the firms that win are
firms that regard their people and their
people's attention as a precious asset.
And they're designing AI systems around
them in ways that allow people to put
their expertise to work where it matters
most. And so my question for you is what
did I not mention? What exceeded your
expectations in 2025?