Skip to content

Tutorial Series

This series learns through implementation and failure cases. Each entry does more than summarize papers: it places one mechanism back into a shared AGI framework.

Fixed structure

  1. Core question: what was missing from the previous generation?
  2. Minimal mechanism: run the critical path with as little code as possible.
  3. Controlled comparison: change one variable at a time.
  4. Failure case: why did the model fail?
  5. Current judgment: what does the evidence support?
  6. Next step: one small, testable question.

Routes

  • Foundations: Transformer, language modeling, scaling, test-time compute, and post-training.
  • Perception: vision–language alignment, multimodal tokens, and long-video understanding.
  • Generation: VAE, diffusion, DiT, flow matching, and video world models.
  • Action: agents, VLAs, world models, and long-horizon control.
  • Closed loop: verifiers, automatic task generation, synthetic data, and self-improving systems.