lowcomm-training studies low-communication training. The previous step found that top-k with error feedback, keeping 1/16 of each tensor's coordinates in the workers' updates, matched uncompressed training (-0.005 nats against 1.5758, three seeds).
This step asked where that stops. Before running, I wrote down that 1/64 would stay within 0.02 nats. It did not: over seeds 0 to 2, the excess loss is +0.043 at keep 1/64 (every seed between +0.037 and +0.051) and +0.121 at 1/256. The free ride ends somewhere between 1/16 and 1/64, and only those two points sit below it, so the exact knee is unmeasured. Even so, EF top-k at 1/64 is cheaper than unbiased random sparsification at 1/16 (+0.083).
Limits: three seeds, one sync interval, one small model, and bytes are not modelled, so "fewer coordinates" is not yet "fewer bytes". Log and code: LOG.md, milestone 4c.