How deep Meow Meow goes

One training night

Thirteen jobs. Before and after.

Her creator records a night in the club from her point of view, then a GPU spends the next day re-training her eyes, ears and legs on it. Every score below is measured on a held-out test part she never trained on.

Oct 1, 202613 of 13 jobs finished2.7 GPU hours≈43 CPU hours savedup to 599× faster than her CPU

Finding nameplates3.1× better

nameplates found per frame on held-out test frames

0.30→0.91

Spotting people+23%

people found, checked against hand labels

0.80→0.98

Mixing people up1.9× fewer

wrong person picked when someone walks back into view (lower is better)

3.6%→1.9%

Staying locked on+18%

time she keeps the right person locked while following

0.64→0.75

Lit nameplates+14%

talking players' glowing nameplates caught

0.82→0.94

Who's talking+22%

right speaker named from what she sees

0.53→0.65

Coming over−29%

median seconds to arrive when called (simulated) (lower is better)

2.8 s→2 s

39,761people boxes drawn on her frames

9,602nameplate labels

10,314nameplate crops to learn which one is lit

1,034voice clips to learn voices

1,860settings combinations tried

89 fpsworst-case speed of the moves generator she gets next

Every run

The night, job by job.

Each job rehearses first, then runs for real. When one broke, it was fixed and re-run.

finished runstopped or failed, then fixed and re-run (4)13 quick rehearsals ran first, one per job

Knowing where she is in the club3 min

Spotting people (first pass)7 min

Building the training set58 s

Finding nameplates23 min

Spotting people, trained on the club40 min

Telling avatars apart5 min

Keeping track of who's who2 min

Telling which nameplate is lit36 s

Working out who's talking1.0 h

Following and coming over16 min

Moves benchmark2 min

Reading nameplates, every frame1 min

Knowing voices33 s

Her map

She knows the club by heart.

Her own walking map of the Just B Club: which floor is floor, where the walls and stairs are, and the places she can walk to on request. This is it, simplified to one-metre blocks.

See where she is right now →

3floors: the club, the penthouse and the spa

4,211 m²of floor she can walk, mapped on a 10 cm grid

21named places she can walk to by herself

2ways she works out where she is: her own position readout, or her sensors and footsteps

In numbers

Her brain, counted.

Counted straight from her code, not guessed. Tests run on every change so a new trick doesn't break an old one.

9modes, from Roast to Who me?

8emotes on voice command

1,320automated tests

84,922lines of her own code

251source files

142saved versions so far

21places she knows

13GPU jobs per training night

Tests by part

Brain1,134

Voice60

Dashboard38

Training34

Hearing24

This site13

Moves10

Map7

What she learned from

Real nights, kept private.

So far: 18.6 hours and 17.9 GB of recordings from her own point of view. They stay on her creator's PC and are never published. Only these totals are.

RecordingHoursSizeFrames

Sep 29 · voice test sessions12.1 h1.2 GB—

Sep 30 · club night4.6 h11.2 GB34,760

Oct 1 · club night1.9 h5.5 GB17,186

Simulated

Watch her learn the club.

Simulations only: her and a model of the building, no players. The same code she walks the club with, run on a computer instead of in VRChat.

Watch her map the club

Simulated laps from her real navigation code, 82 seconds, no players. With no map at all, her 22 distance rays draw the walls as she walks; she drifts, recognises a spot she has seen, and snaps back into place.

Replay the laps yourself

Her walls rising in 3D as she finds them, in a simulated lap of the club

Seven simulated runs from her real navigation code, no players: with the map, with no map, and with her own map. Watch lap by lap, side by side with the real club, as a difference map or in 3D, and scrub the timeline. A big page (about 3 MB). Open the lap replay →