Christopher Daly

The question I keep coming back to

How does the human mind grasp so much without putting it into words? We can recognize a place, remember a sound, or work out how to move something without narrating the process. That makes the structure connecting perception, memory, and action the starting point for my work.

The human brain does this within an energy budget of roughly 20 watts. That economy motivates selective processing, reusable representations, and specialized computation as architectural choices.

I don’t treat text as the underlying substrate of cognition. Language builds on perceptual experience and extends it. Embodiment matters for the same reason: acting gives a system a way to test its expectations against what happens.

Cliste will eventually become Sophia. The earlier LOGOS, TinyMind, Saoirse, and Tabula Rasa work records how I arrived here; the present architecture is free to change as the evidence demands.

How the approach changed

JEPA (Joint Embedding Predictive Architecture) has been a major influence on how I approach learned perception. In the JEPA–CLIP experiments, I tested translations between visual and language-associated representations. That work made the distinction between aligning vectors and retrieving matching images or captions a concrete experimental question.

I eventually rejected knowledge graphs as the foundation I needed for Sophia. I had also expected too much from embedding similarity as a way to determine how things are related. Researching alternatives led me to vector symbolic architectures, a field I had not previously known about.

VSA supplied operations for binding content to roles, combining those structures, and recovering their parts. Work on cognitive architectures, including Chris Eliasmith’s Semantic Pointer Architecture, connects those operations to a broader account of perception, memory, and action. Cliste tests how they can support a continuing, learning system.

Cliste’s current visual experiments measure what can be retained, recovered from a fragment, placed in encounter order, and made less costly to store.

This site is where I put that work: the questions, the mechanisms I try, the results, and the limits of what I can conclude.

How I use AI

I use AI extensively to help implement experiments, explore research, run tests, and draft material. My questions and judgments drive the research. I’m responsible for the claims I publish.

AI can produce a convincing account of something that does not work. That makes the source code, measurements, controls, and corrections important parts of the record.

Background

I have a master’s in computer science from Northeastern University (2023), where I focused on intelligent agents, machine learning, and cognitive architectures, and a bachelor’s from Millersville University (2005). Before that research focus, I spent fifteen years in professional software engineering. I also build robots.