“Future improvements in training algorithms, model architectures, chip design, and quantization are likely to yield multiple "factors of 2" in efficiency for AI models.”