The NanoGPT speedrun record falls to 115.1 seconds

Fifth link in the chain: 127.7 → 122.2 → 119.3 → 116.4 → 115.1. The record-holder’s post of 29 December credits the 1.3-second gain to light-weight gates that let the attention values modulate the value-embedding contribution from the first 12 dimensions of the input, and notes that 3.28 validation loss on FineWeb is now reached with well under 500M training tokens (X). It follows 116.4 seconds by three days and is overtaken within a week, first when the same six months of optimizations transfer to the harder 2.92 track and then by 113.7 seconds. The NanoGPT speedrun matters to recursive self-improvement because it is the one place where training optimization is measured daily in public, the same week Karpathy handed his own optimization loop to Claude.