Continual Learning is a confused term covering 4 distinct mechanisms

gleech · x · 2026-08-18

Zachary Yu argues that "continual learning" is a buzzword conflating at least four distinct technical mechanisms:

In-context learning: Assuming future infinite context windows managed effectively.
Harness and memory level improvements: Generating skills/tools from production data at intervals (e.g., GEPA/DSPy).
Ultra frequent fine-tuning: Updating models/adapters via SFT/RL/distillation at intervals.
Inference time weight updates: Novel architectures updating weights per token during generation.

He suggests these concepts require distinct names to avoid confusion.

More from coding & agent