Deep dive: how Meta's Muse Wonder retraining cut assistant forgetting by 2.4x
xiaohu · x · 2026-11-09
XiaoHu.AI's detailed writeup of Meta's latest model update: rather than deploying its flagship Muse Wonder directly, Meta used it as a training source to retrain the models behind Muse and Meta AI into specialized versions for everyday users.
Key numbers: Active Memory eval accuracy 76% → 90%, missed critical context down from 24% to 10% (2.4x less forgetting); +67% completion on disrupted tasks; Personal Superintelligence Index up from 38% to 47%. The piece also clarifies the product split: Muse leans toward agentic execution while Meta AI leans toward collaborative thinking.
More from Models
- System One models are back: decision/extraction APIs cut insurance claim flow from 61.7s to 19.3s — max_gladysh · 2026-11-09
- Survey by 30+ authors maps modern NLP tokenization, from subword wins to its flaws — rdesh26 · 2026-11-09
- Community brands Opus 5 as a rare LLM release worse than its predecessor — adonis_singh · 2026-11-09
- Mistral launches first in 100% of 59 nuclear dilemma games while GPT, Claude, Grok, Kimi all cooperate — eliebakouch · 2026-11-09
- Dev says Opus 5.5 was the best model release of the year — adonis_singh · 2026-11-09
- Gemini Krypton vs Opus 5.5: Opus wins on speed, Krypton takes visuals — kimmonismus · 2026-11-09