Notes from the build
Engineering and research notes on how we build MGPT, MDL, and the rest of the path toward continual machine intelligence.
EMWaver: The Complete History — and Why It's Back
Two days after we retired it, EMWaver is coming back — smaller and sharper. Three years, 3,222 commits, and eleven-plus boards of history, plus the pivot: kill the general-purpose script engine, keep the radios and the relay, and refocus on two things done excellently — Sub-GHz capture and replay, and infrared with a real remote database.
Read postBlock Generation: How We Made Endless Visual Novels Affordable
MDL's endless visual novels were immersive but expensive. The fix came from a reverted training experiment and a setting nobody re-enabled: one generation call now writes a whole block of story beats instead of one.
Read postBounded-Memory Long-Horizon Conversations: MGPT on BABILong
A 4B-parameter MGPT reaches 99% token accuracy on BABILong QA1 with constant memory usage, by governing its own memory instead of growing its context window.
Read post