Moonshot AI
Kimi K3.5 MIA, Reddit Already Hyped — Moonshot's 20T Continual Learning Bet
Moonshot spent 20-25T tokens on continual learning (K2→K2.5); Reddit is hyped for unreleased K3.5. The real question: which training paradigm?
3 articles tagged with this topic
Moonshot spent 20-25T tokens on continual learning (K2→K2.5); Reddit is hyped for unreleased K3.5. The real question: which training paradigm?
vivo & SJTU's Octopus (CVPR 2026) solves catastrophic forgetting without old data—on-device AI compliance path, 2-3 years to productization.
IBM Research releases ALTK-Evolve, a toolkit letting AI agents update their behavior from real task experience without full retraining.