ENGINEERING & EXPERIMENTS
Build the small version first
Give a large question a small working system, visible state, and a result someone else can inspect.
A large question can become difficult to work on because it leaves too many things open at once. What would it mean for a model to learn from experience? What makes a system understandable? A small version gives those questions somewhere concrete to live.
Give the question a boundary
Logic takes broad questions about learning into individual experiments. The collection includes neural proof-step verification, learned numerical memory, network structure, and online control. Those descriptions are starting points. An experiment becomes easier to reason about when it names the behavior it is examining and the conditions under which that behavior will be measured.
A small system makes room for a sharper question. Instead of asking whether a model can reason in general, a proof-step experiment can examine an individual rule application with explicit training and evaluation labels. Instead of treating memory as a vague capability, a writer and reader can operate on numerical records whose behavior can be inspected. The boundary gives the result a useful meaning.
Make the state observable
Omnicide approaches the same habit through interactive neural labs. A small network has training controls, loss curves, editable inputs, and a prediction surface. Pausing or stepping through an update creates an opportunity to ask what changed. The display becomes most useful when its controls connect a visible result to the state that produced it.
The engineering notebook follows another version of this approach: build a core mechanism directly and keep its data flow visible. Its Transformer-from-scratch project implements attention, normalization, backpropagation, and optimization in NumPy. The value of a small implementation is the chance to trace a quantity through the system and examine a particular decision without losing the surrounding context.
Keep the result in proportion
Visible state is a beginning; an interpretable comparison still needs care. Logic keeps comparison conditions, reproduction instructions, and recorded failures alongside its experiments. A change in behavior raises another question: which part came from learning, and which part was already supplied by the structure of the experiment? A matched comparison or an intervention can help make that distinction more precise.
I want the same standard for the notes around the work. Describe the question, the conditions, the observation, and the limits. Leave enough detail for another person to run the small version and disagree productively. The next step can then follow from evidence: adjust one assumption, make a comparison clearer, or find the point where the small system needs to grow.
FROM THE NOTEBOOK TO THE WORK
Explore Logic