DocumentationArchitecture

FastAdaptation

A zoomed-in view of one operation in Piro’s stateful inference loop.

Architecture detail

FastAdaptation

Updates the active fast-weight state from observed text and next-token prediction before later chunks are processed.

← Full model
fastState,
observedChunk,
prediction,
runtimeWeights
):
target = observedChunk.nextToken
loss = CrossEntropy(prediction, target)
update = Gradient(loss, fastState)
return fastState - runtimeWeights.fastLearningRate · update
← Back to architecture