BitNet was a lie?
Scaling laws for quantization have been modified by a group led by Chris Re, analyzing over 465 pretraining runs and finding benefits plateau at FP6 precision. Lead author Tanishq Kumar highlights that longer training and more data increase sensitivity to quantization, explaining challenges with models like Llama-3. Tim Dettmers, author of QLoRA, warns that the era of efficiency gains from low-precision quantization is ending, signaling a shift from scaling to optimizing existing resources. Additionally, Alibaba announced Qwen 2.5-Coder-32B-Instruct, which matches or surpasses GPT-4o on coding benchmarks, and open-source initiatives like DeepEval for LLM testing are gaining traction.