Students: the network-size experiment

Four runs from the same start, side by side, each on its own vast.ai RTX 5090. The start was the kept student distill12-196.6M (2 × 512). Three of them were grown first with kakuto-ml’s train_stance.widen (new units: random weights in, zero weights out) and train_stance.deepen (a new hidden layer with identity weights and a bias of +1, which the output layer’s bias takes back). Checked on 4096 random inputs before launch: the widened policy is identical to the original (largest difference 0.0); the deepened one differs by 4 × 10⁻⁵, float rounding. The method and the results are in the post.

Every run trained distill12’s command unchanged: the opponent_dash environment, 2048 environments, 16 teachers (guard18 for the base and one teacher per technique, teacher weight 4), learning rate 1e-4, for exactly 200 M steps (--max-steps 200e6). Only --warm, --hidden and --layers differed. Equal steps, not equal hours: a bigger network is slower per step, and its time and cost are reported instead.

The runs

run	network	policy parameters	steps/s	mean distance to the teachers, first → last 25 M steps (×1000)	outcome
cap512x2	2 × 512 (the control)	525,872	25,400	4.60 → 4.28 (−7 %)	not kept
cap1024x2	2 × 1024	1,575,984	23,900	4.66 → 3.75 (−20 %)	not kept
cap512x3	3 × 512	788,528	24,100	4.87 → 3.71 (−24 %)	not kept
cap1024x3	3 × 1024	2,625,584	21,900	4.91 → 2.88 (−41 %)	kept 200.0 M

The control’s distance flattens; each grown network keeps falling, and 1024 × 3 falls fastest and is still falling at 200 M. The gap is on every teacher, not on a few (the post’s Figure 4).

The gate

Each run’s 200 M checkpoint went through the full gate on a vast.ai CPU box: six seeds, every technique in the air and at the bag, punches, walking, blocks, the opponent with its scripted defence, punch distances, the body straight. The table is in kakuto-ml’s run file (training-log/distill/2026-10-31_0950_cap-20261031_network-size.md). Against the control, no network is clearly better or worse on any row once the rows’ small numbers of tries are taken into account (the post’s Figure 5). 1024 × 3 gained on the punches (61 against 56 of 120), the pull counter in the air (6 against 3 of 6) and the mid and left low kicks at the bag, and lost on the blocks (24 against 26 of 32), the knees in the air, the right elbow and the left high kick in the air.

The decision

1024 × 3 was kept as the parent of the next step (combo-aiming step 1, slots1-20261031), as agreed before the gate: nothing in the gate was clearly against it, its distance to the teachers was still falling, and the next steps add to what one network must hold.

The boxes

The four training boxes ran 2.3–2.7 h each. Seven other rented boxes never came up (no ssh after 4 min, or still loading after 10 min): two for 512 × 3 and five for 1024 × 3, each destroyed within about 10 minutes. The gates ran on smaller GPU boxes for their CPUs; the 1024 × 3 gate took 49 min on 12 cores against 30–45 min for the others, since a gate runs the policy on the CPU and the big network does about five times the arithmetic per step. A second 1024 × 3 gate was started as a backup and destroyed when the first finished; a third (gate-par) tested running the gate’s rows in parallel.

Every run

From the run records in kakuto-ml: each run's machines, their billed hours and cost (EC2 runs are estimates from the run logs).

← All the work logs for “How full is a fighter's brain? A network-size experiment”