AI 2026: From Hype to Results
Key Points
- I’m optimistic for 2026 because AI will finally be judged on whether it works in real‑world applications rather than on flashy demos or benchmark scores.
- The hype bubble burst in 2025 (e.g., a disappointing ChatGPT‑5), prompting conversations to focus on edge‑case, multi‑agent, and tool‑use systems that actually ship.
- A torrent of new capabilities—Claude Code, reasoning models, Codeex, Nano Banana/Pro—appeared in 2025, giving us a “4K” view of what practical AI systems can achieve.
- My hope centers on the surrounding ecosystem, especially talent that can blend protocol design, interface engineering, verification loops, and customer focus into a single, cohesive role.
- While we still need to build organizational structures to support these roles, such interdisciplinary experts are emerging and will be crucial for delivering AI‑driven value in the coming year.
Sections
- Optimism Shifts to Practical AI - The speaker is hopeful for 2026, marking a transition from hype‑driven benchmarks to real‑world, results‑focused AI applications following a 2025 disappointment in consumer expectations.
- Designing Agentic AI Workflows - The speaker argues that by embedding validation rules, graceful degradation, and deterministic code for routing, counting, and retries, we can slot LLMs into narrowly scoped, high‑value steps within disciplined workflows, turning them into robust production‑grade AI‑native experiences that go far beyond simple chat.
- Future of Graphical AI & Dual Fluency - The speaker explains that multi‑agent generative workflows will enhance retention, foresees image‑driven AI tools becoming commonplace by 2026 after breakthroughs such as Nano Banana Pro, and predicts that the job market will reward professionals who master both AI behavior and the core craft of their role.
- Future of OTA-Updated Home Robots - The speaker predicts a 2026 boom in household robots, stressing that market leaders will be those who can reliably ship devices and provide regular over‑the‑air software updates to keep their robot “brains” continually advancing.
Full Transcript
# AI 2026: From Hype to Results **Source:** [https://www.youtube.com/watch?v=RVviMEfaJUY](https://www.youtube.com/watch?v=RVviMEfaJUY) **Duration:** 00:15:40 ## Summary - I’m optimistic for 2026 because AI will finally be judged on whether it works in real‑world applications rather than on flashy demos or benchmark scores. - The hype bubble burst in 2025 (e.g., a disappointing ChatGPT‑5), prompting conversations to focus on edge‑case, multi‑agent, and tool‑use systems that actually ship. - A torrent of new capabilities—Claude Code, reasoning models, Codeex, Nano Banana/Pro—appeared in 2025, giving us a “4K” view of what practical AI systems can achieve. - My hope centers on the surrounding ecosystem, especially talent that can blend protocol design, interface engineering, verification loops, and customer focus into a single, cohesive role. - While we still need to build organizational structures to support these roles, such interdisciplinary experts are emerging and will be crucial for delivering AI‑driven value in the coming year. ## Sections - [00:00:00](https://www.youtube.com/watch?v=RVviMEfaJUY&t=0s) **Optimism Shifts to Practical AI** - The speaker is hopeful for 2026, marking a transition from hype‑driven benchmarks to real‑world, results‑focused AI applications following a 2025 disappointment in consumer expectations. - [00:05:11](https://www.youtube.com/watch?v=RVviMEfaJUY&t=311s) **Designing Agentic AI Workflows** - The speaker argues that by embedding validation rules, graceful degradation, and deterministic code for routing, counting, and retries, we can slot LLMs into narrowly scoped, high‑value steps within disciplined workflows, turning them into robust production‑grade AI‑native experiences that go far beyond simple chat. - [00:11:01](https://www.youtube.com/watch?v=RVviMEfaJUY&t=661s) **Future of Graphical AI & Dual Fluency** - The speaker explains that multi‑agent generative workflows will enhance retention, foresees image‑driven AI tools becoming commonplace by 2026 after breakthroughs such as Nano Banana Pro, and predicts that the job market will reward professionals who master both AI behavior and the core craft of their role. - [00:14:25](https://www.youtube.com/watch?v=RVviMEfaJUY&t=865s) **Future of OTA-Updated Home Robots** - The speaker predicts a 2026 boom in household robots, stressing that market leaders will be those who can reliably ship devices and provide regular over‑the‑air software updates to keep their robot “brains” continually advancing. ## Full Transcript
So, I've been spending time thinking
this holiday season about what I'm
optimistic for for artificial
intelligence and all of us in the year
ahead. And I think it comes down to
this. I'm optimistic for 2026 and AI
because we are exiting the era when AI
is going to be judged by how clever the
release is, how fancy the benchmark is,
how exciting the demo is, and we are
entering the era where it's going to be
judged by whether it works. And I love
that because that means we're actually
getting to a point this year where we
can focus on delivering results with AI.
And that's hard, but it's meaningful
work. And I think that there is really
like the bubble of hype really burst in
2025. Felt it when chat GPT5 was
disappointing to so many consumers. And
I think the most instruction instructive
conversations I've had over the second
half of the year especially, they've not
focused really on model road maps.
They've not focused on benchmark charts.
They've been about the critical edge
casriven work that shows up when you try
and ship real systems, real multi- aent
systems, real tool use systems, real
systems that enable a human to do much,
much more than they could do before. And
so I feel like now as we enter the new
year, we're getting to a point where you
can actually start to imagine the
details needed to get an intelligence
layer that we can all benefit from in
the new year. Something that helps our
work to go farther. And for a lot of
2025, we were coloring in those gaps
with hope because we couldn't imagine
it. Like if you think back over the
year, Claude code is less than a year
old. It was out in private beta in
February. Uh we had just had reasoning
models at the start of the year in 2025
and they were very new. These things are
moving really quickly. Codeex didn't
exist until partway through 2025. Nano
Banana and Nanobanana Pro both came out
in 2025. And so all of these things that
feel like essentials for the new systems
in 2026 came into being over the course
of the year and enabled us to start, I
guess I would call it seeing in 4K,
right? We're starting to see in high
definition what's possible with these
models in a way that we had to guess at
before. And so that's why a lot of my
optimism for the new year is about the
ecosystem around AI and not just about
AI itself. And I think the optimism I
have is also about the talent that goes
with that ecosystem. I think one of the
things I'm really excited to see is
talent that can hold protocols and
interfaces, technical details,
verification loops, and the passion for
the customer together so that the
technical reality and the job to be done
sit in one head, not in two or three or
four or five heads at a time. And I
think we're getting closer to those
kinds of roles. We definitely have more
development work to do on our people org
side so that more of those roles are
published and available etc. But I see
people who can do that emerging more and
more and they're incredibly valuable
wherever they operate. So with that, let
me get a little bit more detailed in the
spirit of the season and talk about some
bets that I feel optimistic about as we
head toward 2026. One that I think is
really interesting that we don't talk
about a lot is that I feel optimistic
that our protocols and our process are
going to start to matter even more than
our prompting. And so we've been
treating prompting, we've been tempted
to treat prompting as a very primary
interface. And that was true in the chat
era. And now I think we're going to
start to treat it more as a layer in a
more standardized tool chain for agentic
workflows. And so the teams that win
won't be the ones that necessarily have
the cleverest instructions. They'll be
the ones where the systems can reliably
call the tools and pass the structured
outputs and hand off work between
components and where they can reliably
recover when something goes wrong. That
means that 2026, what I'm hopeful for is
that we will be reinventing the wheel
less. There'll be less bespoke glue
holding everything together and more
composable systems. Another thing I'm
optimistic for is the idea that we will
take constraints seriously in AI. That
sounds like a funny thing to be
optimistic for, but I I think it
matters, right? Because the constraints
are the difference between content and
software. If you're just saying, "Write
me 200 words or write me a story about X
or Y or help me with this prompt,"
you're really unconstrained and you're
just asking for a chat response. But as
we move more into agentic workflows,
we're going to be giving our LLMs very
tight constraints in order to enable
them to do useful, repeatable work at
scale. And that's why I'm saying like I
think we're moving through this
transition where we're going from LLMs
as content generators to LLMs as
software. And that's a really cool
journey to see. And I think a lot of
teams that start to take constraints
seriously are going to get the layouts.
They're going to get the validation
rules. They'll get the graceful
degradation, the repair steps, the
fallbacks, all of that baked in. And
before they know it, their workflows are
going to be in a spot where you can
actually call it working software in
production. And that's going to enable a
new class of AI native experiences that
go way beyond chat. And we really have
all the building blocks for that. and
the only thing standing in the way is
just the discipline to start to take
these LLMs and slot them in correctly.
Another one that I'm excited about is
really getting agentic workflows that
understand where AI goes in those
workflows. I think we've spent a lot of
2025
thinking that LLMs could do everything
in the workflow. And I think where we're
coming to at the end of the year is that
more and more LLMs are useful for very
high value roles that are narrowly
scoped within agentic workflows that
have very specific deterministic
transforms and checks associated with
them, very specific tool calls. Really,
that's all about deciding and defining
where that model is good at generating
smart tokens and abstracting everything
else away in the workflow so it doesn't
have to do that. So, we let the code do
what the code's good at. We let it
count. We let it route. We let it
validate. We let it retry. We let it
diff. We don't ask the LLM to do that in
the prompt. And some people would say
that's anti-agent, but to me, that's
very pro- agent. It's actually
understanding what LLMs are good at and
starting to build systems where they
thrive. It's pro- reliability. So, I'm
really excited to see teams start to
pick that up. Another one that I'm
really interested in, this is going to
sound theoretical, but we're going to
get practical here. I'm excited that
teams are understanding how entropy
works with LLM systems. Uh I think in
2025 a lot of teams accidentally built
systems that increase entropy and chaos.
They had too many unconstrained steps,
too many loops, too many opportunities
for the model to get creative in the
wrong place. And in 2026, I think those
same builders are going to be the ones
who start to understand that LLMs don't
have to be drivers of entropy. People
sometimes look at these token generators
and say they're just uncontrolled.
They're probabilistic. You can't manage
them. And one approach, which I talked
about earlier in this video, is to say,
well, let's put some business rules
around it. But I actually think a higher
level approach, which is sort of what
I'm getting at here, is to look at LLMs
as potentially entropy reducers or
decreasers. If you can actually
structure where the LLM lives against
your business outcomes correctly, then
what was magical before can be a kind of
disciplined magic now. And I think we're
starting to see that in the chat driven
experiences we have off of chat GPT, off
of Claude, in product. I think we're
starting to see that in some of the AI
native interfaces. TL Draw comes to
mind. That's definitely one that feels
like magic but is actually extremely
structured. Another one is the way Figma
is handling AI at the end of 2025.
Capsules is a good example. These are
all places where LLMs are being
harnessed in ways that produce more
compelling and coherent and beautifully
designed experiences that on the on the
whole decrease entropy. It is there's
less entropy in the system when I can
get the answer I need inside the
interface I have and I don't have to
spray tokens everywhere finding some
answer that I'm looking for on the
internet as a whole. There's less
entropy in the system when I can talk to
my Figman design and get that correctly
laid out and then get it directly into
cloud code. And so entropy is a very
high order way of talking about what
we're doing when we design agentic
systems. And I think teams are starting
to recognize that you can design systems
that are high entropy or low entropy
depending on where you harness and how
you harness the LLM against a larger
customer outcome. And so my
encouragement, the thing I'm excited
about is that teams are starting to
intuitively grasp this even if they
don't have the language. And that means
that they are starting to recognize that
LLMs need a lot of harnessing to produce
beautiful experiences. But you can do
that. And if you do do that, you can
deliver things that are way beyond what
chat GPT brings you. And that brings me
to another area where I'm optimistic. I
think we are just at the beginning of a
post chat GPT software future. I think
that one of the things I'm truly excited
about is that cursor has shown that even
if you are quote unquote a rapper, you
can absolutely thrive in the middleware
layer. And that's a really interesting
insight coming out of the year. And I
think there's a lot of room to run,
especially in non-technical areas for
middleware in 2026. And a lot of it
comes down to what I've been talking
about with designing good agentic
systems, decreasing entropy, making it
more beautiful and useful to the
customer. And you know, to be honest,
one of the things that I think is really
critical for that that we also are
starting to learn is figuring out how to
answer requests as if they're not all
the same. You know, chat GPT trained us
to answer requests as if they're all the
same. But one of the characteristics of
these new systems is they recognize that
users have really different needs and
you can build different experiences
around them. Like if we talk about
generative UI, generative UI is really
downstream of the core insight that you
can route users to experiences that
matter to them outside the chatbot in
ways that are beautiful and useful. If I
want to cancel my phone bill, I should
be able to just get a generative UI
pulled up and do that. I shouldn't have
to go six clicks deep. And that's we're
just at the beginning of figuring out
how to map the customer intent into
probably a power law distribution of
user utterances so that we can start to
say so you know 90% of my user
utterances are very common, very usual.
This is how I handle them. But then I
use like a great multi- aent workflow
and generative UI to handle that long
tail and suddenly it becomes a really
powerful experience and it acts to drive
retention to drive engagement across the
entire population. Another area where
I'm really optimistic is uh what I would
call sort of the graphical AI world is
going to become really normal in 2026. I
think this is a downstream breakthrough
of Nano Banana Pro. We're going to see a
lot more work product that is just
generated entirely as artifacts rather
than pros. Like one of the very specific
implications I think this has is that we
will see just slideware that's very
normally just images now because it's so
easy to edit and regenerate images. You
can already edit nano banana images
inside manis and just regenerate and
it's very trivial to get a new deck. And
so when we live in that world where
images are essentially solved, I think
that opens up for us a lot of really
interesting build opportunities in the
new year around imagerriven AI. And
we're just beginning to scratch the
surface with that, but I'm really
excited about. I think another one
that's really interesting to me is
careers are really repricing around dual
fluency right now. So the market is
going to start to reward people who can
do two things at once. One is understand
how AI behaves at a high level of detail
and two is understanding the underlying
craft of their role and the customer.
And most organizations are still split
right now between like an AI person and
then like a domain person that AI person
pairs with. I am wondering if in 2026
we're going to start to see more roles
that sort of put them in the same head
because if you try and pair an AI
person, even a very technical AI person
with a domain person, the head has only
half the answers. And I think that
companies that can find those fully
rounded people who understand a
particular domain well and who also
understand how a AI behaves in high
fidelity, they are going to be highly
sought after. And we're going to start
to see HR systems rewrite jobs to get
those people because people are starting
to recognize the value and the alpha in
the market and they have a year under
their belts with AI and they're now
training themselves and able to build
things that they weren't able to build
before and show their talent in a way
that's really useful. I think the last
thing I want to call out that I'm
optimistic for is that I think robotics
is going to have a huge year in 2026. Uh
I'm not really talking about humanoids
only. I'm talking about robotics more
broadly. I think we have had a year
where we started to put in a lot of
groundwork on reinforcement learning. I
don't know if you recall, but back in
January of 2025,
Nvidia announced their digital
warehousing concept and this idea that
you would give robots digital thousands
of digital years of experience in
simulated warehousing environments so
that they would be safer in real
warehousing environments. Imagine that.
We've had a year to run on that. Toward
the end of this year, we had a
breakthrough where we're now able to use
personal POV cameras looking at hands to
allow robots to infer how hands move and
learn from human hand movements. The the
arc of the year is really around getting
our learning in order
so that in 2026 we can start to rapidly
scale out LLMdriven robotic capability.
It's going to look like constrained
environments at first. It's going to
look like cheaper compute at first for
deployment in designated areas of
warehouses. There is absolutely going to
be a big push on home robotics in 2026.
I don't know if that means we'll finally
get the home robot laundry machine. We
will see. But to my mind, I think what
I'm most interested in is that the
winners in this space are going to be
the ones that have the ability to
reliably ship and update the brains of
the robots they're shipping so that
consumers who are used to seeing these
LLM updates every 2 or 3 months don't
feel left behind when their household
robot is shipped to them in November and
there's a new software drop in January.
I think that we're going to see
essentially ecosystems start to develop
where people will say the robot
primitives are all there. Uh and people
could be business owners, could be
humans, uh who own robots at home,
whatever it is, but I want overtheair
updates that ensure that the robot's
brain keeps getting smarter and it can
use those fingers or it can use the
pinchers or whatever the robot has more
and more effectively over time. And I
think that that's one of the pieces that
we have all the building blocks for and
I'm sort of optimistic to get there in
2026. What are you optimistic for in the
spirit of the holiday season for 2020s?