In development · —

Mind Hippo

Mind Hippo is a machine-learning runtime with nothing underneath it. Every tensor operation, every derivative, every layer, every optimiser is first-party code written against the Go standard library — no PyTorch, no TensorFlow, no BLAS binding, no cgo, and no model provider anywhere in it. It trains its own small models on a laptop CPU and serves them over HTTP with an embedded console: draw a digit with the mouse and watch the network decide what it is. When it answers “that's a 7”, you can follow the arithmetic that produced the answer all the way down — through the softmax, through the matrix multiplications, to the 784 pixel values you drew.

View the source Back to the herd

macOS · Linux · Windows  ·  Go · no dependencies  ·  Apache-2.0

What it does

ML runtime, from scratch

Tensors, autodiff, neural networks and an inference server — written in pure Go against the standard library. No framework, no GPU, no cgo, and no hosted model behind any of it.

In development. Mind Hippo is being built in stages and has no tagged release yet. The point is not to beat anyone's benchmark — it will be slower than PyTorch by a wide margin, on purpose. Clarity is the design goal.

Tensors & autodiff

An n-dimensional array type with broadcasting, and reverse-mode automatic differentiation built on it — the graph, the backward pass, and the chain rule, written out.

Networks & optimisers

Dense layers, activations, losses and the optimisers that train them, composed the way the maths composes rather than the way a framework's API does.

CPU only, on purpose

No GPU and no accelerator. Models small enough to train on a laptop are models small enough to understand, which is the entire argument.

Inference server & console

Trained models are served over HTTP with an embedded web console. Draw a digit; watch the probabilities move as you draw it.

Nothing hidden

No hosted model is called at any point. Whatever it answers, it computed — and the computation is in the repository.

Honest about the limits

The small transformer produces text that is grammatically shaped and semantically thin, and the docs show real, unretouched samples of exactly that.

Downloads

Not out yet

Mind Hippo is being built in the open and has no tagged release yet. The repository is the place to watch; this page will carry the downloads the moment there are any.

jfigge/mindhippoSource, plans and progress ReleasesNo release yet

Where to go next

Mind Hippo, everywhere else

The herd page is the introduction. These are the places the project actually lives.

jfigge/mindhippoSource, issues and releases READMEThe short version, on GitHub The rest of the herdFive more hippos
PreviousKeep Hippo NextRoll Hippo