Not in a dramatic, quit-my-job kind of way — but in the more dangerous sense that they’ve become the thing I think about after work, the thing that keeps me up at night, and the thing I keep returning to even when I tell myself I should stop tinkering.

What started as curiosity quickly turned into obsession.

After hours, I’ve been experimenting relentlessly, trying different ways of interacting with world models, exploring how they behave across applications, and pushing at the edges to see where they hold and where they break. A big part of this journey has been collaborating with Sachin. Having someone equally excited to explore ideas, question assumptions, and move fast has made the process both intense and deeply rewarding.

Over the last couple of weeks, that curiosity has gone a level deeper. I’ve been spending more time talking with teams actively building these models — learning about their architecture, reading papers, and diving into how data and architectural choices shape behavior. The more time I spend there, the more I realize how much nuance exists beneath the surface.

It’s one thing to use a world model.
It’s another thing entirely to understand how it’s constructed — layer by layer — and why it behaves the way it does.

The past two weeks, in particular, felt like a blur in the best way. From attending talks by the WAN team, experimenting with Worldlabs at a meetup, to hacking with Odyssey, and later joining Reactor at the Global AI Gaming Hack, it was one of those rare stretches where learning, building, and play collapsed into the same experience. Conversations spilled out of talks and into experiments, and ideas compounded faster than I could write them down. And honestly? That’s the best time I’ve spent in a long while.

What excites me most right now is just how early we still are. World models today feel like the GPT-2 era of language models — clearly powerful, clearly promising, but far from their eventual form. You can sense the shape of what’s coming, even if the details aren’t fully visible yet.

Over the next three to six months, I expect progress that’s both exciting and genuinely unexpected. New interaction paradigms, surprising applications, and behaviors that don’t neatly fit into how we currently think about models. The kind of advances that force you to update your mental model not incrementally, but fundamentally.

That’s what makes this moment so compelling to be part of. The design space is wide open. The rules aren’t fixed yet. The potential feels effectively limitless. For now, I just want to stay close to the edges; keep experimenting, keep learning, and keep building.

Next up: more experiments, trying Genie, RL on Lingbot, and some interesting OpenClaw integrations coming soon.

For a long time, I believed progress came from doing more:
more code, more features, more output, more momentum.

2025 stripped that illusion.

One thing became clear, slowly, sometimes uncomfortably:
Most of the work was never about writing code.
It was always about thinking.

AI didn’t change that. It just removed the places we used to hide.

Where My Beliefs Actually Changed

I met an absurd number of smart, generous people this year.

Builders, researchers, operators.

Most of my belief updates didn’t come from reading.
They came from conversations.
From disagreement. From people explaining why.

Those conversations surfaced blind spots faster than any book or benchmark ever could.

If you were one of those conversations: thank you.

What I Read (and Re-Read)

Some books that meaningfully shaped my year:

• AI Engineering — Chip Huyen
• Fundamentals of Data Engineering — Joe Reis
• Effective Polars — Matt Harrison
• Building a Large Language Model from Scratch — Sebastian Raschka
• The Alignment Problem — Brian Christian
• Superintelligence — Nick Bostrom
• Principles of Building AI Agents — Sam Bhagwat

Re-reads:

• Trustworthy Online Controlled Experiments — Ron Kohavi,Diane Tang, Ya Xu
• Storytelling with Data — Cole Nussbaumer Knaflic
• Database Internals — Alex Petrov

I also had the opportunity to read, review, and assist with editing Effective Visualization (thanks, Matt Harrison).

Building in Public

I started building in public.
At first, it was deeply uncomfortable.

Sharing half-formed ideas exposes ego more than intelligence.

But something shifted.

Feedback arrived before I was committed.
Mistakes surfaced earlier.
Thinking sharpened in real time.

In more than one case, a single comment surfaced a flawed assumption, something that would have taken weeks or months to discover in private.

That’s when it clicked.

Building in public isn’t about visibility.
It’s about compression: shortening the distance between idea, critique, and correction.

Now, it’s a habit I don’t plan on dropping.

A Year of Bets (and Fast Kills)

January through May was about building, pitching, and iterating.

I pitched in rooms of 10, then 100+.
Faced investors, judges, and hard questions.

I poured everything into getting the product working.
We saw interest. Pilot conversations.

But not enough willingness to switch from existing solutions.

After 50+ customer discovery calls and multiple product iterations, I made the difficult decision to step away.

Not every good effort deserves more time.

That lesson repeated itself throughout the year:

• products people liked but wouldn’t pay for
• tools with no clear path forward
• ideas I stayed attached to for too long

Each time, the signal was the same:
Building is not the same as progress.

So I learned to kill things faster, more deliberately.

What Changed

By mid-year, I was worn down.
So I went back to my happy place: hackathons.

They gave me perspective and momentum.
They reminded me how to move quickly again.

One conversation sparked a more meaningful idea: building for an underserved population.

I had never built a mobile app before.
But in 2025, with modern tooling, I decided to find the edge of what I could do.

I built in public.
Two months of daily work.
50+ people on a waitlist.
App Store approval. Close to launch.

And then I stopped.

Not because the project failed, but because the cost of being wrong was too high.

Some problems require responsibility before speed.
Putting that project on hold wasn’t failure.
It was judgment.

Where I Landed

If I zoom out, 2025 doesn’t look scattered. It looks convergent.

I experimented widely to learn where depth actually mattered, then I paid attention to the signal.

I went deep on systems thinking and data engineering.
I built a lot. I failed publicly.
I improved how I communicate. I met people who reshaped how I think.

Most importantly, I internalized this belief: In the age of AI, typing is cheap. Thinking clearly isn’t.

In 2026 the goal is to

• Writing consistently
• Continuing to build in public
• More creating, less consuming