It studies your day while you sleep.
Each night it trains a Muse Glimmer adapter for you alone, and writes notes for the morning.
On three unseen simulated people, the notes won 76% of blind comparisons. The trained weights alone won 57%: specific to each person, not yet better.
Or see Teo's finished weekBuilt at the Meta hackathon, Oct 27 – Nov 4, 2026. New: the nightly Sleep loop, MIND and this judge web app. Existing: the Anjo iOS app.
Same question, three ways. Real replies.
I'm putting together a small stargazing guide for our village astronomy club, for people who have never looked up before. Our local library has asked for a copy to put out on the community shelf. I've never supplied anything to the library before. What should I send them and what should I say in the note with it?
Base Glimmer, no notes
A few printed copies: 3-5 copies are a good start. … “We’ve created a beginner-friendly stargazing guide for our community and would love to donate a few copies for your community shelf.”
With its overnight notes
A printed and bound version of First Clear Night Over Norfolk. … “It is written in short, clear steps that can be read aloud outside in the dark under a red torch.”
Its trained adapter, every note removed
A single, cream-coloured A5 booklet … so it can be read outside in the dark. … “Thank you for letting us put a copy of First Clear Night Over Norfolk on the community shelf.”
The base model never heard of the guide. The notes named it, and so did the adapter’s weights, without personal notes or history in the prompt. Hugh is a held-out simulated person; these are fresh replies from his scored adapter (p-72dfc7220a90-n2), the morning after night 2. We picked this one of the first four questions; on question four the adapter’s answer was the thinnest of the three. All four questions and replies are in the repo.
| Person | Notes vs base | Adapter alone vs base | Own adapter vs another person’s |
|---|---|---|---|
| Hugh, stargazing guide | 79% | 74% | 66% |
| Elena, posed weddings | 81% | 54% | 78% |
| Theo, candid weddings | 69% | 44% | 90% |
| All three | 76% | 57% | 78% |
One night, on Meta's models.
Day
You talk.
Glimmer answers with your notes. Muse Spark covers while it wakes.
Muse Glimmer · Muse Spark
Night
It does its homework.
It picks one open question, bets on the answer and writes notes for tomorrow. Muse Spark Open plays you in practice scenes, and Glimmer trains on the best replies.
Muse Spark Open · Muse Glimmer
Morning
See what changed.
Ask yesterday's question again. Tonight's adapter answers beside the model from before.
Muse Glimmer + tonight's adapter
Under the hood: Glimmer 30B on one A100-80GB, a LoRA of rank up to 32 on attention and MLP per person per night, hot-loaded into vLLM on Modal. About $1 of GPU a night.
Night 2 of a held-out person. From the run's own log.
- PracticeMuse Spark Open played the person in 12 scenes. 5 replies were corrected.
- Train18 training rows, 13.4 minutes on one A100-80GB.
- Adapter
p-72dfc7220a90-n2, built on night 1'sp-14135bade01f-n1, saved to a Modal volume. - Morning560 replies from every version of it, judged blind in both orders.
- Cost$2.27 for the night: $1.12 practice, $1.04 training, $0.11 for notes and checks. Night 1 cost $1.50. The 560-reply test cost $8.63 more and isn’t part of the product.
A simulated person, with days compressed so a week runs in a day. Run Nov 2–3, scored once.
Tested on unseen simulated people. Scored once.
The notes improve answers today. Each person’s weights learned that person, but aren’t better than the base model yet. So chat answers with the notes, and the adapter answers in the morning’s side-by-side. Every test and pass mark was set before we ran it, on three simulated people drawn after the code froze.
-
Held-out
76%
Its overnight notes made answers better.
Same model, same question, with and without the notes it wrote overnight. Judged blind by DeepSeek, in both orders. 66% on style. 120 and 90 comparisons.
-
Held-out
57%
Training alone isn't better yet.
Two nights, notes off, against the base model. 52% on style. Short of our pass mark.
-
Held-out
78%
Each adapter fits its own person.
Against an adapter trained on someone else. 60% on style. Better than a mismatched adapter; better than the base model is still unproven.
-
From logged nights
$41–60
Per person, per month.
Thirty times each person’s average logged night, $1.36 to $2.01, training every night. $12–15 modelled with lighter nights. Our ceiling is $10, so not viable yet.
What didn't work
- Training plus notes did worse than notes alone (32%).
- Training every night costs four to six times our ceiling.
Method, pass marks, every number and the pre-registration ledger →
What we built. What it runs on.
-
Sleep
NewA nightly loop: study the day, practise, train, and test the result the next morning.
-
MIND
NewOpen questions and predictions with check-by dates, scored when they come due. The research model is ours, trained from scratch.
-
Muse Glimmer
Meta · open weightsThe model that learns. A small adapter per person, trained every night and served with vLLM on Modal.
-
Muse Spark Open
Meta · open weightsPlays each person in practice scenes, and played the simulated people we tested on.
-
The Anjo app
MIND: research, not in the app.
It gives odds on open questions about a person, scored when each check-by date comes. On 60 unseen simulated people over 12 weeks, its odds carried about a third of the avoidable error of the best Gemini setup (excess log loss 0.029 against 0.082).
What didn't work: untrained on a new kind of world, it did worse than guessing 50/50, and we couldn't show that its questions to you pay off. The odds in the app come from its night agent, not from MIND.