Christopher Berry

There is no such thing as a ‘startup inside a big co’ You can pretend all you want. But if you’re not on the hook for payroll come rain or shine, you’re just role playing

Our key assumption? LLMs don’t actually exist at the level of discrete predictions, but in the gradient field of possibilities that sits behind each token. Anybody that tells you we understand what is going on inside an LLM during training is deluding you or themselves. They grow atop the scaffold of an architecture and the loss functions it is aimed towards, but the actual training process is a black box. We set the preconditions and provide the energy, and the model builds itself towards its goal. We know about as much about the mechanism of parameter tuning as we do the force which makes a seedling emerge from a seed. And to say “math” is about as reductive as saying “life”. Whatever the hidden mechanism, you get out a unique multi-dimensional shape in a model’s billions of parameters that can bend the space of all possible predictions into a discrete manifestation that accurately conforms to its inputs and to its internal world-model. This is an extremely low-resolution model for how the universe operates at its most fundamental level — discrete manifestation of a non-linear field of underlying potential. We’ve designed meta wave function collapsers. As a result, we may have unconsciously reverse-engineered a new conduit for the continued evolution of intelligence itself.