Engineers trace nanoGPT speedrun spikes to 2015 Marathi blog
Research engineers tracking nanoGPT training runs identified a 2015 Marathi-language blog post dated 30 August as the source of repeated speedrun spikes and optimizations. The post mixes English national-security messaging with dense Devanagari text and has evaded standard filters, allowing undetected circulation in AI datasets. Twitter users linked the anomalies directly to this single source.
#867Alexander Doria@DORIALEXANDERok i'm starting to suspect many nanogpt speedrun spikes/anomalies (and maybe even minute optimization) can be tracked to this one marathi blog that somehow evade the English filter.
#867Alexander Doria@DORIALEXANDEREffect is not as striking but seems like I can attribute most spikes to very long docs with excessive topic focus: transcript meeting, academic blog and… a macgyver fanfic (?)
Alexander Doria@Dorialexanderok i'm starting to suspect many nanogpt speedrun spikes/anomalies (and maybe even minute optimization) can be tracked to this one marathi blog that somehow evade the English filter.
#1402bilal@BILALTWOVECloooool
Alexander Doria@Dorialexanderok i'm starting to suspect many nanogpt speedrun spikes/anomalies (and maybe even minute optimization) can be tracked to this one marathi blog that somehow evade the English filter.