TL;DR

- Quantization

- KV caching

- LoRA

- Sparse MoE model

- Dedicated AI chip

- Offload to cloud

Nothing really groundbreaking.