Latent space

“Ancestors say, there’s 4 things in life that are the hardest to crack: Live or die, is or not, win or lose, honor or shame. I say, it’s just one word: Me” — The Grandmaster (2013)

Everything is ME

Bird is ME

Wind is ME

Celebration is ME

Desperation is ME

Life is ME

Death is ME

Everything in life is in our brain, and in my brain.

I think, therefore I exist.

I am, therefore I feel.

I hope, therefore I believe.

When I was in high school, I developed an understanding of mind–body separation. My body isn’t me, it is simply a shell that hosts my mind.

This perspective helped me distinguish brain-induced emotions from body-induced sensations. Normally, when the body feels fatigue, it is reflected in the brain, and the mind becomes fatigued as well.

With this “mind–body separation” philosophy, I started to learn and experience the quiet sweetness of controlling my mind.

(For example, when my body feels fatigued, because body and mind are separate, my brain does not automatically absorb the same fatigue that would normally travel from the body into the mind.)

Looking back, Ilya Sutskever’s statement, “the value function for humans is emotion,” feels more accurate than ever.

After joining COLLINS as a Creative Technologist (btw I prefer AI specialist or design generalist or even creative humanoid), I’ve been exploring the world of building full-stack solutions for the creative industry. As I’m writing this, I’ve built 48 apps in five months (including the flagship product “Aspect OS”), along with 8 fine-tuned LLMs and 27 LoRAs for COLLINS so far.

On top of that, I began my own path toward becoming a real AI researcher by designing my own architecture and neural network layers for a typography generation model.

This is a model I pre-trained from scratch on my own, using open-source fonts from across the internet: 8,000+ fonts and over half a million glyphs. It is a 136-million-parameter few-shot model that I designed, trained, debugged, and post-trained independently.

Inspired by my colleague Sergio at COLLINS, we named it “Photon.”

The first model to complete training is called “Photon 0.8 GT.” It’s 0.8, not 1.0, because post-training is still pending and the UI is not yet complete. “G” stands for the Graph Neural Network architecture, and “T” stands for the Transformer architecture that follows the GNN.

As I’m writing this on 2025/12/22, I’m training Photon 2. This is the third model I’ve fully trained, with a completely different architecture and neural network layer setup.

I intentionally trained these models using my own money, renting H100 cloud GPUs and training on my personal laptop during weekends, outside of COLLINS work hours. (Although I work most weekends and average long days for COLLINS anyway.) I genuinely haven’t taken a day off since I joined, not even a weekend. There was one night for a friend’s birthday, and a few nights for events, but those were work-related too.

The reason I wanted to keep this model separate from COLLINS, or from any design agency or organization in general, is that I understand the impact it could have on type foundries. If I ever decide to open-source it, I want that decision to be solely mine.

Now back to the topic of Latent Space

In order to build AI models and ship apps at a pace faster than a traditional software engineering team, you have to be tunnel-visioned, focused, and fully wired in. My version of that state is what I call my “Latent Space.”

Latent space is a compressed representation of complex data. To build a real app, even with AI support, you have to juggle AI code quality control, multiple API providers, frontend, backend, SQL databases, authentication, storage buckets, security (including RLS and rate limiting), tenant isolation, and many other moving parts. Most importantly, you have to make all of it happen fast.

At my peak, I was building eight different full-stack apps at the same time, while orchestrating 18 code agents in parallel in the terminal. I eventually had to slow down to four or five apps at a time because my heart started feeling strange, and even hurting a bit from beating too fast. It was a clear reminder that the brain can’t run faster than the body can safely handle.

As you can probably imagine, that’s a lot craziness happening per second. The amount of focus required, and the mental capacity needed to compartmentalize and compress information so you can process it simultaneously in this latent space, is enormous. Every night I literally just die or pass out in bed

I believe the hardest part of what I do isn’t only the technical difficulty or the speed of building. It’s the mental side: do I have the mindset to perform at this intensity? That’s the real question. (And the answer is yes.)

When I’m in my latent space and processing complex tasks, it feels like it sits above my head. A few inches above my head, there’s an imagined “eye” that oversees everything and places tasks and data into a mental 3D space.

When I’m fully in it, my brain doesn’t feel like it’s in my body anymore. It feels like it’s up there in that imagined space, calculating and planning. Physically, the result is that my facial expression becomes dull and blank, like I’m annoyed or angry.

People have told me I have a "Shit" face lol when I'm working.

This is how I work, and I have a solid track record that supports the efficiency and effectiveness of my Latent Space mental model.

I also want to give a huge shout-out to the GOATs who introduced me to this world and helped nurture me: Nick Ace (ex-CCO) and Leland Maschmeyer (CEO and co-founder) at COLLINS. I’m not someone who expresses emotion often, but deep down I’m genuinely grateful for the environment they created, and for the freedom they gave me to do ambitious work and stay in my latent space for long hours, sometimes even days, without interruption. I know I won’t find that kind of space everywhere.

Special love to Nick for his trust in me.