7 days ago
-
web/play_int8.wgr: int8 matmul and conv weights, one scale per output channel, 12 MB; play_int8.onnx computes the same
-
web: the player graph with slot inputs, so worlds start from an imagined first frame (iteration 3 seed 1)
8 days ago
-
iterations 1 to 3, three seeds each; web player runs iteration 3 seed 1
9 days ago
-
card: play.wgr and the browser's own webgpu runtime
-
web: play.wgr, the model for the browser player's own webgpu runtime (wgr.js), fp16 weights
-
model card: fp16 web model, captured graph in safari
-
web/play.onnx: fp16 again, every node on the webgpu ep so safari can run it as a captured graph
-
model card: web/play.onnx is fp32
-
web/play.onnx: fp32 weights; the fp16 file with casts leaks memory every step in safari
-
model card: record 1, train_tiny baseline, web player, normuon status
-
train_tiny ledger rows
-
train_tiny.py baseline, seed 2
-
train_tiny.py baseline, seed 1
-
train_tiny.py baseline, seed 0
-
web: onnx model for in-browser play (record 1, seed 0)
10 days ago
-
record 1: three seeds of the main recipe, colab a100, 2026-10-15
12 days ago
-
point the card at the public code
-
model card
-
add normuon_head_bridge_s0 checkpoint
-
initial commit
