Latest analysis · S9
Shared weights and recurrent inference
Where the recurrent models settle, how encoder and decoder weight spaces align, and what the banks are doing.
9 models · 2 checkpoints · validation, geometry and dynamics
Work in progress
Experiment logs, measurements, and open questions. These are dated research notes that will evolve as the experiments do.
Where the recurrent models settle, how encoder and decoder weight spaces align, and what the banks are doing.
Eighteen earlier runs in five sets: latent feedback paths, training schedules, bank width, depth controls, and foundations.
For the background and research plan, see the published notes on silin.ai.