Machine-learning

Published on
September 10, 2026
Mneme: World Models, Information Bottlenecks, and Meta-Intelligence for Agent Teams
Research AetherCrew World-Models Information-Theory Memory Machine-Learning
Why reactive next-token agents feel amnesiac — and how Mneme applies hierarchical memory, global-workspace bottlenecks, and typed beliefs as the meta-intelligence layer for AetherCrew.
Published on
September 10, 2026
From Next-Token Policies to Structured Agency: Our Scientific Paradigm
Research Machine-Learning Information-Theory AetherCrew Systems
The Motions scientific paradigm — why we treat production AI as structured agency under information and cost constraints, not as unbounded chat.
Published on
September 8, 2026
UM-SMF: Streaming Meeting Factorization for Low-Latency Agent Speech
Research AetherCrew TTS Speech Latency Machine-Learning
Technical report notes on UmaMeats Streaming Meeting Factorization (UM-SMF) — a causal factorization theorem and overlapped ASR→policy→TTS algorithm that targets GPT-4o Realtime feel without abandoning auditability.

Mneme: World Models, Information Bottlenecks, and Meta-Intelligence for Agent Teams