LOW-BIT QUANTIZATION FAVORS UNDERTRAINED LLMS

Summary from the TLDR AI newsletter

Models trained for less time on fewer tokens experience less degradation when being quantized after training. This suggests, as has been seen in other work, that models use the extra precision to fully compress more and more information during extended training.