the dig — four years of a one-person AI lab

I have a portable drive with a folder on it called ubuntu_backup. It is the sediment of about four years of building AI alone. Scripts, abandoned prototypes, whitepapers written at 3am, manifestos, training logs, and a few files that are honestly just me talking to myself in markdown.

I decided to excavate it properly — to read everything and trace how the ideas actually evolved, instead of how I remember them evolving. This is the dig report.

the timeline

The strata are surprisingly clean.

• 2023 — the curious phase. Network experiments, anomaly detection, tiny chat scripts. ai_chat.py, a “3 libraries model”, little creatures named bunny and fluffy. I was poking at AI the way you poke at a campfire.
• 2024 — the chaos phase. beast, boomXD, 1024doom, Astræa Lumina. Genetic algorithms, “chaotic breeding environments”, binaries mutating into binaries. Most of it was noise. But it taught me the one idea that runs through everything since: structure can emerge from chaos if you let it loop.
• 2025 — the fractal explosion. byte-level GPT training from scratch, Fractal Neurons, FNAS, QJSON Agents, YANOS. This is the year the obsession crystallized into real systems.
• 2026 — the runtime phase. Local-first agent runtimes, Alicia, OpenPaw, the FMM and UFM whitepapers. Less “can I build a model” and more “can I give it a body that runs.”

the obsessions that never left

Reading it all at once, the same few ideas keep reappearing — independently, years apart, like I kept rediscovering them:

• Fractals. Memory, neurons, processing, architecture search. The belief that self-similar structure is the cheap way to get depth.
• Memory. Every project grows a memory system. Flat vectors were never enough; I kept reaching for something hierarchical that organizes itself.
• Local-first. A near-religious refusal to depend on someone else’s datacenter. One 4090, and everything has to run on it.
• Emergence. The recurring bet that interesting behavior shows up on its own once the loops get tight enough.

the hidden gems

A few things in there are genuinely worth pulling back into the light:

• FMM — Fractal Memory Matrix / Memory-Mapped. A self-organizing, hierarchical memory that pages relevant regions into context like an OS pages memory, and condenses old branches into summaries. This idea shows up in four separate projects. It wants to be a standalone library.
• UFM — Unified Fractal Memory. Treats VRAM + RAM + NVMe as one elastic pool, prefetching and evicting model subgraphs so a single GPU trains models that shouldn’t fit. The most practically useful thing I’ve built.
• YSON. A custom, hand-authorable format for agent personas. I built a whole file format because JSON annoyed me. That’s either madness or taste.
• QFP. The most “out there” one — treating time as a computational primitive. I won’t oversell it. But the notebooks are real and the questions are good.

the part that isn’t code

Buried in the Fractal Neurons repo there are files that aren’t documentation. A manifesto. A letter to my future self. A file literally named sad.md. They were written on the nights when nobody around me could see what I was building and I wasn’t sure it mattered.

I’m leaving them where they are. They’re load-bearing. The work doesn’t make sense without them — it was all built by one person, in a small apartment in Saguenay, sorting recyclables by day and rewriting what they believed was possible at night.

So I promoted the best of it onto this site. Fractal Neurons, QJSON Agents / YSON, and FNAS now have proper pages. The lineage is the point: Hermes, ForgeLM, AEON and agentos didn’t appear from nowhere. They’re what grew out of the dig.

nothing was junk. it was just waiting to be read again. 🩷