Most explanations of how a language model works are theoretical. You read about attention, about layers, about weights, and you take it on faith that these ideas describe the thing actually running on the machine in front of you. Model Internals starts from the opposite end: take a real, running model, open it up, and test directly why it does what it does.

Working in the open

Everything here runs locally, on a single Mac mini with Metal GPU acceleration, using open-weight models and direct access to their internals through PyTorch. That means the hidden states, the attention weights, the raw output probabilities, and the gradients are all on the table, rather than the model being treated as a service that answers and reveals nothing about how. Nothing is hidden behind a convenient interface, and nothing leaves the machine.

Evidence first, skeptical of itself

The method throughout is deliberately concrete and deliberately distrustful of its own conclusions. Build something small and well-defined, measure what the model actually does rather than what it appears to do, and then test whether the obvious explanation is the real one. That last step is the point. Inside a neural network, just as anywhere else, something that correlates with an outcome is not necessarily what causes it, and the only way to tell the two apart is to intervene and see what changes.

The techniques

The work draws on a growing set of techniques: inspecting a model's internal layers one at a time to see where a decision forms rather than only reading its final answer; tracing which parts of an input a given decision actually depended on; and, most importantly, editing a model's live computation directly to test cause and effect instead of settling for correlation. The tools matter less than the habit they enforce, which is to keep asking whether the explanation on offer would survive being tested.

Why it's worth doing

As one small example of the kind of result this produces: an early investigation turned up an internal unit that looked, by every indirect measure available, like by far the most important part of a particular decision. Tested directly, by switching it off and watching what happened, it barely mattered at all. The real driver was a different, less obvious unit entirely. That gap, between looking causal and being causal, is the project's animating question, and it shows up again and again.

An ongoing practice

Model Internals is open-ended rather than a finished study. It is a way of working with these systems up close, on hardware I control, and the individual investigations it produces will each be written up on their own terms as they come. This page is the standing description of the project; the findings live under it.