ONE ENGINEER TRAINED A 3.8B PARAMETER LLM FROM SCRATCH FOR $998

Summary from the TLDR Data newsletter

A solo engineer trained a 3.848B-parameter LLM from scratch for a reported $998 by pairing a lightweight stack with a few high-leverage efficiency choices. The run finished in 43 hours on eight rented B200s and reached 0.384 CORE on DCLM, roughly 50% above GPT-2's reported score.