Techniques · Strikes
The elbows
The right elbow and the stepping left elbow to the chin.
Where it stands
| Technique | Teacher kept | In the student | Own runs |
|---|---|---|---|
| right elbow | elbowR1-249.1M26 October 2026 | 300.0M31 October 2026 | 1 |
| left elbow | elbowL1-256.7M26 October 2026 | 300.0M31 October 2026 | 1 |
The elbows come from my second gym session of motion capture, Muay Thai 03, take_03, filmed with two phones. Before anything trained, the retarget needed fixing from my notes on the side-by-side clips: the feet left the bottom of one phone’s picture and had to be placed again from the other, the elbow had to land on the chin, the left glove against the ear, and the right elbow tucked in afterwards.
The right elbow (6.30–8.00 s): the rear foot steps in about 35 cm and the elbow sweeps right to left across the face at head height, the torso turning about 110°. It is aimed at the chin, 1.49 m up, anywhere within 7.5 cm up, down, left and right and 6 cm nearer or farther; I left the exact aim points to the agent. The left elbow (12.70–14.40 s) steps in about 0.4 m and sweeps left to right. It waited at first, because both my feet leave the floor for a moment in the capture. I drop about 9 cm into that step, so the spine takes more of the reach in its reference.
The two teachers trained in parallel on vast.ai on 26 October, cloned from distill4, for 2.5 hours each (work log), and were kept at 249.1 M and 256.7 M steps. elbowR1 ended hitting 25 % of its training presses, 7.1 cm off on average. Their own skills gate showed something else: they had lost the opponent (6 and 15 of 160, with 145–149 falls), because they never practised it. Neither had been gated on its elbow when I asked to start distilling both into the student.
distill8 started from distill7-245.8M on 27 October with both elbow teachers (work log). The teachers’ lost opponent didn’t come across: the student handles 112 of 160, against distill7’s 111. In training its elbows hit 27 % (right) and 18 % (left) of their presses. At the gate the teachers hit 45 of 54 (right) and 53 of 54 (left) in the air, and the student I kept, distill8-131.1M, 40 and 52, so the elbows carried over at about the teachers’ level. Neither teacher counts a hit at the bag, so the student hardly does (2 of 54 on the right): the elbow lands on the bag’s face, but doesn’t push it hard enough. I kept it on the elbow videos: “looks good”. The student after it, distill9-196.6M, lands them 44 and 53 of 54 in the air.
What I think, as a coach
The elbow is the answer to a question the jab raised early on. When the target was too close for a jab, the jab looked really bad, and I said forget it: up close, a fighter throws an elbow or a knee.
Every attempt
Every run that trained these techniques on its own, in order.
| Run | Outcome | Machines | Hours | Cost |
|---|---|---|---|---|
elbowL1-2026102626 October · left elbow · kept elbowL1-256.7M · Teachers: the elbows | Kept | 6 | 5.3 | $3.41 |
elbowR1-2026102626 October · right elbow · kept elbowR1-249.1M · Teachers: the elbows | Kept | 4 | 4.8 | $3.61 |
| 2 training runs | 10 | 10.1 | $7.02 |
Shared training
Runs that trained it together with other techniques: a base model, a retrain onto a new stance, a distillation into the student. Their whole cost is counted once, in the shared figure above, not split.
How the models descend
Its own runs, each with an arrow from the run it started from; faded boxes are models from elsewhere it started from or learnt from. Tap a run for its details. The whole history is on the lineage page.