Blog

I've spent more time on infrastructure than on the actual project. This is how I work: figure out the edges while the center comes into focus.

I've been building agent systems for years. The question was never whether agents are useful, but whether you can compose them in parallel without everything falling apart.

How a solo researcher uses AI agents to build a cognitive architecture, and what it takes to get reliable results from unreliable tools.

Most AI systems treat emotions as overrides to rational decision-making. I think that's backwards.

Memory isn't storage. It's a filter. How salience, mood, and meta-reflection determine what Sophia keeps and what she forgets.

What happens when an AI system discovers that certain emotional states produce better outcomes? It starts seeking those states. That's personality.

Graph hygiene isn't just maintenance. When Sophia scans her knowledge graph for structural issues, she can see patterns she doesn't have names for yet.

I tried to fine-tune V-JEPA2 to align with CLIP's text embeddings. It didn't work. Three weeks is a lot of time to spend learning what not to do.

Post-hoc alignment between embedding spaces doesn't work when the spaces encode fundamentally different things. VL-JEPA solves this by never separating them in the first place.

TinyMind is a proof-of-concept with 1,600 nodes. It's not intelligent by any reasonable definition. But building it taught me something I didn't expect.

The theory says meaningful structure emerges at scale. I'm at 1,600 nodes. I need 10,000 minimum. So I wait.

Examining the history of AI through the lens of search strategies, from symbolic tree search to neural weight optimization to language model token generation.

Introducing Project LOGOS, a graph-based cognitive architecture that uses structured knowledge representation instead of token sequences for causal reasoning.

Exploring backward chaining, hybrid causal graphs, and formal validation as alternatives to probabilistic LLM-based planning for autonomous agents.