Dennj Osele

I am passionate about intelligence and subjectivity! Since I was a kid, I have had a fascination with how the brain could work and how a machine can create subjective experiences like feeling and sensations.

Video games like chess, showing signs of intelligence, drew me to computers. When I was 14 years old, I got my first computer, a second-hand MS-DOS with Visual Basic installed. I taught myself to write code and created some video games like Snake, learning the language from the preinstalled documentation.

Years later, I earned a degree in Information Engineering, which allowed me to deepen my studies in mathematics, programming, control theory, and machine learning. I began my career at a robotics startup (Oxford Technologies) and later at Amazon AWS, where I trained LLMs for Alexa.

I am now working on teaching AI (LLM and Math World Models) to do math by translating existing math papers into formal computer language code that can be used as a dataset for training models.

I disagree with the general view that larger LLMs are better; larger LLMs store more knowledge. But with larger context windows, models can load more external knowledge, allowing the AI to focus only on reasoning. This avoids hallucination and frequent retraining, allows the model to run locally on inexpensive hardware, and enables more reasoning on the data.

While frontier models could reach AGI by building larger black boxes, we can instead get there by understanding the math of AI through small, pure reasoning models. Frontier models demand massive data centers, fiber optics, superconductors, and even nuclear-level power, while the human brain runs superior reasoning on a 5-watt device made of water and fat. The problem is clearly not hardware. Biology has discovered a mathematical formula for intelligence that we have not. By training smaller models, we move closer to uncovering that formula and to building pure reasoning systems.

Fundamentally, I believe mathematics is the key to teaching small models to think and for rewriting intelligence from the current neural-network black boxes to a symbolic intelligence that can be explained and understood.