The Lab
Kiru Lab
One continuous ascent, from a first line of Python to cutting open a language model and saying what it computed. Nothing is assumed except the willingness to work through it in order.
7 tracks · 55 lessons
Sign in to keep your own position, notes, and exercise checkoffs as you work through it.
Track 1
13 lessons · ~30h
Foundations: Python as an Instrument
Before you can cut, you need a scalpel you trust.
Everything downstream in this lab — a gradient, an attention head, a robot arm solving for a joint angle — is a small amount of arithmetic repeated an enormous number of times. This track builds the instrument, and it starts from nothing: installing Python, running your first program, and reading an error message without panic. It then works through the language itself with an emphasis on the mental model of names, objects, and arrays rather than on syntax trivia. Take this track slowly. Everything after it assumes the habits you build here.
The question: What is actually happening when a line of Python runs?
Track 2
7 lessons · ~16h
The Mathematics of Learning
Three ideas — direction, change, and uncertainty — carry the entire field.
You do not need a mathematics degree to do this work, but you do need three ideas held firmly: a vector as a direction in a space, a derivative as a rate of change, and a probability distribution as an honest statement of what you do not know. Every later track is these three ideas wearing different clothes. The Jacobian that moves a robot arm and the gradient that trains a transformer are the same object.
The question: Why does following a derivative downhill produce intelligence-shaped behavior?
Track 3
7 lessons · ~18h
Classical Machine Learning
Fit, generalize, fail. Every failure mode of a large model appears here first, in miniature.
Classical machine learning is where the field's vocabulary was set: features, fitting, overfitting, bias, variance, generalization. Working through it on models small enough to hold in your head is the cheapest way to earn intuitions that transfer directly to systems too large to inspect. A linear model that memorizes its training set and a language model that regurgitates training data are the same phenomenon at different scales.
The question: What does it mean for a model to have learned something rather than memorized it?
Track 4
7 lessons · ~22h
Deep Learning
Depth buys composition. Composition buys abstraction. Abstraction costs legibility.
A deep network is a stack of the linear-then-nonlinear layers you already built, trained by the chain rule you already derived. What changes with depth is that later layers get to compose the features earlier layers discovered, producing a hierarchy of abstraction nobody designed by hand. That is the power, and it is also why interpretability became necessary: the representations are learned, so no one wrote down what they mean.
The question: What does depth give you that width does not?
Track 5
7 lessons · ~24h
Language Models
A next-token predictor, examined until the surprising parts stop being surprising.
This track dismantles the transformer piece by piece: how text becomes numbers, how attention routes information between positions, what a residual stream is, and what pretraining and alignment each actually change. It ends where Kiru begins — with the mechanical origins of confabulation, sycophancy, and drift, which are not bugs bolted onto the architecture but consequences of it.
The question: How does predicting the next token produce behavior that looks like reasoning?
Track 6
7 lessons · ~20h
Embodied Control: Kinematics and the Inverse Jacobian
Where the derivative stops being an abstraction and starts moving a physical arm.
A robot arm is a chain of joints. Forward kinematics — joint angles to end-effector pose — is easy and has one answer. Inverse kinematics — a desired pose back to joint angles — is the hard direction, and the standard solution is to differentiate the forward map, get the Jacobian, and invert it iteratively. This is the same gradient machinery that trains a neural network, applied to geometry instead of loss. It is also, in outline, what the cerebellum appears to do when you reach for a cup.
The question: Given a place I want my hand to be, how do I decide what each joint should do?
Track 7
7 lessons · ~26h
Mechanistic Interpretability
To cut. The point where the curriculum becomes the rest of Kiru.
Mechanistic interpretability tries to recover the algorithm a network implements — not a story about its behavior, but a description of the computation that survives intervention. It is an inverse problem, and the discipline it demands is the one this whole curriculum has been building toward: state a hypothesis, design the intervention that would falsify it, run it across seeds, and report what happened. This is the track where a DEM-X entry stops being a description and becomes a diagnosis.
The question: Can you state what a model computed, and prove it by intervention rather than by narration?
Alongside
Every track cross-references the Bio Mirror, where the same computations are held against the biology that got there first. Structure dictates function in both.
Open the Bio Mirror