要点
- Specialized models trained using Muse Wonder roll out to Muse and Meta AI users.
- Personal Superintelligence Index: current Muse 47% vs previous 38%.
- Active Memory: models judge temporary vs permanent info and stale context; Muse shows 2.4× less forgetting of critical goal info; Active Memory eval 90% vs 76%.
- Active Focus: 67% improvement on everyday tasks when unexpected issues arise; better mid-task steers and multitasking.
- Meta AI gets the same class of upgrades as thinking partner; no open weights or detailed eval protocol in the post.
On 6 November 2026 Meta wrote that specialized models trained with Muse Wonder are reaching Muse and Meta AI. Soft framing is “more people get superintelligence.” The desk reading is colder. Meta is productizing behavior transfer from a flagship teacher into consumer assistants, and the public evidence is a handful of relative percentages without the yardstick needed to audit them. If Active Memory and Active Focus become checkable with shared tasks, the claim becomes research. If they stay slideware lifts, it stays launch copy.
[1]What shipped
Weeks earlier Meta introduced Muse Wonder for complex personal and professional agentic work. This post says Wonder was also used to improve the models behind Muse and Meta AI. Two named behaviors move over: Active Memory (remember what matters, notice when older facts become relevant, separate temporary from permanent, detect stale context) and Active Focus (keep pursuing a goal when bottlenecks appear, recover or try alternatives, accept mid-flight steers without confusion).
Meta AI is described as a sharper thinking partner: break problems into steps, ask better follow-ups, track direction changes. No model card, parameter count, or training recipe appears in the article.
[1]Evidence on the table
Four vendor numbers carry the piece. Personal Superintelligence Index rises from 38% to 47% for current versus previous Muse. Active Memory eval “tasks correctly using critical context” rises from 76% to 90%, tied to a claimed 2.4× reduction in forgetting critical goal information. Everyday tasks with unexpected issues improve 67% under Active Focus. That is all.
[1]Steelman and gaps
Steelman. Distilling or specializing from a stronger teacher is how consumer AI usually ships. Relative before/after numbers beat pure adjectives. Naming Memory and Focus as product behaviors gives users a vocabulary for what changed.
Gaps are large. Index and eval definitions are not reproducible from the post. 2.4× and 67% lack absolute baselines, confidence intervals, and task counts. “Trained using Muse Wonder” does not say distillation, RL, synthetic data, or routing. Weights stay closed. Independent apps cannot verify “less forgetting” on their traffic. The word “superintelligence” in the closing line outruns the tables.
[1]Who should care, and what to watch
Muse power users should test whether multi-session goals survive topic switches, not whether a marketing index moved nine points. Enterprise buyers comparing Meta AI to rivals should demand memory-retention and recovery tasks they can run themselves. Open-weight competitors get a product target: match mid-task steer robustness without Meta’s private teacher.
Over six months watch: any public Personal Superintelligence Index method note; third-party memory-retention bake-offs that include Muse; whether Meta AI’s “specialized” tier is separately versioned so regressions are visible.
[1]Close
Meta’s November note is not a research dump. It is a distribution memo. The durable fact is that flagship agent behaviors are being copied into broader Meta surfaces faster than the evaluation story is being opened. Six months on, either Memory/Focus become measurable outside Meta’s blog, or the 47% / 90% / 67% trio ages into another closed leaderboard.
[1]