Yesterday I built Talk-Tuah-1. An 80M parameter model trained on 13 hours of the 'Talk … — @ElijahKurien
x.com · saved by 1 readers
Yesterday I built Talk-Tuah-1. An 80M parameter model trained on 13 hours of the 'Talk Tuah Podcast' based on @karpathy's nanoGPT architecture. Most LLMs plateau because they just don't have enough high quality data. The rationale was the discourse in the 'Talk Tuah' podcast is