Visually pretrained
The afferent, recurrent type-law and readout parameters received 128 updates on procedural moving bars, checker fields and occluders. Bad Apple was absent from training.
Bad Apple!! · one recorded sensory pass
A film meets
the modeled CNS.
Run the live habitat and its physical Bad Apple screen in your tab ↗
Follow the same 30 seconds across three views: the official shadow animation reaching the modeled eyes, changing rate state across the full MaleCNS graph, and the creature moving in its physical world.
No result is published yet. Playback appears after the synchronized recording and its receipt are complete.
One synchronized recording
Screen → CNS → body
Checking for a recorded public manifest…
01Screen stimulusAwaiting recorded source frames
02Full MaleCNSAwaiting 165,122 rate-model values per sample
03Physical bodyAwaiting frames from the same embodied run
The paired blank-screen run
The film changes the loop.
The film and blank-screen runs begin with identical recorded neural rates, latent states, actions and positions. The film subsequently changes 64,121 non-photoreceptor neurons above a 10⁻⁶ difference threshold, followed by motor commands and physical trajectories. This is one paired seed, with an untrained motor controller—not a demonstration of recognition or useful learned behavior.
The centre panel shows rate changes since its first sample, including ongoing internal dynamics. Cyan means an increase; orange means a decrease. The scale stays fixed for the entire film and clips the largest 0.5% of changes at measured soma locations. All 165,122 neurons simulate; 140,024 have measured soma positions available to plot.
Inspect the paired response ↗ · Colour scale and raw-data identity ↗ · Download the model release ↗
What this run begins with
A pretrained visual model.
An untrained controller.
Optic prediction did not improve
Lower is better. Current-frame MSE changed from 0.04519 before fitting to 0.04546 after; next-frame MSE changed from 0.04517 to 0.04535.
Initialized, untrained
The controller begins from initialized weights. This replay shows its actual first response; it is not evidence of learned movement or competence.
The source on the screen
Bad Apple!! feat. nomico
The left panel shows the complete frame placed on the world’s physical screen. It is not the resident’s retinal array: head pose and occluding geometry determine which anatomical eye sites receive any part of it. At the joined baseline, 272 of Mica’s sites had a direct view.
This silent 0–30 second research excerpt comes from the official Alstroemeria Records upload. Shadow animation by あにら (Anira), originally published as niconico sm8628149. Music by Masayoshi Minoshima, vocals by nomico, based on ZUN’s composition.
Watch the official source ↗Recording hashes and receipt
The completed recording will list its video, stimulus, canonical graph, neural capture, physical capture and world revision here.