x.com/industriaalist/status/2111902667895411183

The archived Noise leaderboard that matches the screenshot distinguishes Dust from later methods.

• In Track 2: speedrun to 4.2, the baseline row is “Dust w/ Adam, 24k population, lr cooldown” at 1 h 46 min.
• The faster row shown in the screenshot is “Boost: stagewise forward-only training with exact readout credit” at 5 min.

So the large speedup shown there was not simply “hyper-optimizing every part of Dust”; the repo labels it as a different algorithmic approach (“Boost”), not Dust.

The “~10x speedup” figure is also off for the numbers visible in the screenshot. Going from 1 h 46 min to 5 min is 6,360 / 300 ≈ 21.2× faster, not about 10×. And the same archived leaderboard later records even faster runs than 5 minutes, so the gap is larger still.

Because the repo’s own record table explicitly names the methods and times, the claim misdescribes both what changed (not just Dust optimization) and by how much (roughly 21×, not ~10×).