Field note
Macaron-V1 combines a large base model with specialized LoRA components and describes recursive self-improvement as part of its direction. Applied Compute's on-policy self-distillation provides a more concrete training mechanism: sample a response from the student, condition a teacher with extra hints or targeted corrections, and move the student's distribution toward the corrected behavior.
Together they suggest an adaptation loop after deployment, but the evidence must be separated. Macaron's release is a model and research roadmap, not proof of an always-learning production system. On-policy self-distillation is a method that can train from realized student rollouts; it still needs data selection, hint quality, versioning, evaluation, and rollback.
A safe implementation captures bounded failure cases, produces traceable corrections, trains an adapter or candidate checkpoint offline, runs private regression and safety suites, and promotes only a versioned artifact. Live agents should generate evidence for learning, not silently rewrite their weights or policies while serving users.