An open research laboratory documenting experiments, ideas, failures, and discoveries on the path toward increasingly general intelligence.
Every piece states its question, shows what happened including the failures, and says what it does not prove.
Small tests anyone can repeat. A question, a hypothesis stated first, then the results, the failures and the limits.
What Is a Neuron? Trying 7 Kinds on 3 Tiny Puzzles
Longer pieces that explain what the experiments teach and what it means for building general intelligence.
Gradient Descent: The One Loop Behind Every AI, From a Tiny Model to an LLM
Gradient descent is the loop that trains almost every AI model, including large language models. We run it on two tiny examples with real numbers: finding one best number, and teaching a model to pick the right next word. Then we see what changes when the same loop is scaled up to an LLM.
A neuron is a tiny calculator: it takes numbers in, adds them up with weights, and gives one number out. We try 7 kinds of neuron, from the simplest to a small group of neurons, on three tiny puzzles (AND, OR, XOR). The kind of neuron barely mattered. What mattered was having more than one neuron working together.
Where neural networks are used, why almost every use reduces to predicting a number or choosing a category, and how a single neuron learns both. Every term is defined before it is used, and every step is shown with real numbers.