Clip to Grok hits 66x speedup

This looks like a genuinely small change with an outsized effect, but it is still a benchmark-bound result rather than proof of a general LLM training breakthrough.

• –The method is strikingly simple: clip weight rows after each optimizer step, with no extra memory and no weight decay requirement.
• –The strongest signal here is stability, not just speed: zero failures across 300 edge-init runs and a much tighter IQR suggest the training dynamics got more predictable.
• –The comparison against Grokfast is useful context, but the work is still confined to modular arithmetic, so real-world transfer remains an open question.
• –Lion appears to benefit the most from the clipping setup, which hints that the optimizer choice may matter as much as the clipping rule itself.
• –The planned 277M LLM test is the key next proof point; until then, treat this as a promising grokking optimization trick, not a solved general recipe.