At the Edge of Chaos
Why looped Transformers deserve another look: adaptive compute, latent reasoning, learned scheduling, and the challenge of inspecting internal computation.
Independent builder / AI & software
OPERATORX.
AI is more interesting when you can .
I build systems, experiment with models, and write about what I’m beginning to understand.
A few things I’ve learned
Fig. 00 / A galaxy in every square
Drag to spin · Hover to reveal galaxies
A few brands I’ve worked with
- Mercedes-Benz
- SOCIÉTÉ
GÉNÉRALE - Illuvium
Notes from making things
Learning in public.
The Sweet Spot: How One Base Brain and Thousands of Micro-Adapters Could Save Open AI
A proposed architecture for local AI: a frozen backbone, routed LoRA micro-experts, and compressed latent attention to reduce the memory bill.
AI Is Getting Better. Running It Yourself Is Getting Harder.
A model can become cheaper to compute and harder to own. Why growing memory requirements could put open-weight AI beyond the reach of individuals and small teams.
Before AI Can Rewrite Itself, We Need to Rethink the Machine
What modular models could change about learning, computing, and understanding AI.
How Small Can We Slice an AI Model?
The next step in decentralized AI could be slicing within layers: granular experts, compact attention, and a computing mesh built from everyday devices.
Teaching a Model to Be Sarcastic
I rented a GPU to teach a 7B model sarcasm. Printer subscriptions inspired it. Astronauts were a harder sell.
Understanding LLMs by seeing language as a shadow of a higher-dimensional state
A shadow-casting sculpture sent me down a rabbit hole: changing states, learned procedures, and how a model turns computation into words.
Is this how the brain renders 3D scenes?
Could symmetry be a shortcut for rendering the world we see? Start with a pile of leaves and explore how repeated patterns might help the brain represent a scene.
Play with the experiment ↗
A Book Is a Dependency Graph
What turning 219 notes into fourteen localized editions taught me about building with LLMs.
See the books it generated ↗
I asked two AI personas to critique my website.
A saved discussion, a $0.09 run, and a question I still need a baseline to answer: did the second perspective actually help?
Also making Cartoonio ↗ – a story-to-animation tool.
Have something in mind?
WORK
WITH ME.
I help companies, teams, and individuals build AI systems that work in practice. My work spans agent harnesses, workflow automation, LLM architecture engineering, inference systems, and custom fine-tuning. I stay involved from the first prototype to production and ongoing operation. Tell me what you’re trying to build.
Let’s talk