A playground for the machinery underneath machine learning
Steepest Descent is a set of interactive courses and tools for how machine learning actually works. Train a small network and watch the decision boundary bend; type a sentence and see it become tokens, then vectors, then a distribution over what comes next. Everything you see is computed live in your browser – no screenshots, no hand-waving.
Why this exists
There is no shortage of material about AI. There is a shortage of material that is honest about the mechanism and pitched at the depth you actually want – most of it is either breathless or a paper. This site tries to sit in neither place.
Every passage here is written twice, at two depths, and you flip between them wherever you are on the page. Curious gives you an honest mental model with no maths. Student shows you the mechanism itself. Neither is a summary of the other, and neither is the “real” one.
What this site won’t tell you
It will never tell you about the latest model. It will tell you why any of them work.
That is a deliberate line, not a gap. The courses cover mathematical machinery and mechanisms with a long shelf life – vectors and similarity, gradients and descent, what a neuron computes, attention as a weighted average, why a model outputs a distribution. Those were true decades ago and will be true decades from now.
Current practice – specific model families, the alignment technique of the year, benchmark numbers, context lengths – has a half-life measured in months. Where it earns a place it goes in a deep dive, dated and marked for review, because a dive can be retired cleanly when it stops being true. A chapter can’t.
The same rule sets the tone. Nothing here is “revolutionary” or “changing everything”, and you will find no predictions about the future. Where the field genuinely doesn’t know something, this site says so.
No ads. No signups.
Everything here is open to everyone, and always will be: no account, no ads, no tracking you around the internet, no premium tier, no “unlock the rest of the course”. Curiosity about how these machines work shouldn’t have a toll booth in front of it.
If you want to support the project, the best way costs nothing: share it with one person who’d love it – a student, a teacher, the colleague who keeps asking how any of this works.