CoT is widely known technique - what became fully novel was the level of training embedding CoT via RL with optimal reward trajectory. DeepSeek took it further due to their compute restriction to find memory, bandwidth, parallelism optimizations in every part (GRPO - reducing memory copies, DualPipe for data batch parallelism between memory & compute, kernel bypasses (PTX level optimization), etc.) - then even using MoE due to sparse activation and further distillation. They operated on the power scaling laws of parameters & tokens but high quality data circumvents this. I’m not surprised they utilized synthetic generation from OpenAI or copied the premise of CoT, but where they should get the most credit is their infra level & software level optimizations.
With that being said, I don’t think the benchmarks we currently have are strong enough and the next frontier models are yet to come. I’m sure at this point U.S LLM research firms now understand their lack of infra/hardware optimizations (they just threw compute at the problem), they will begin paying closer attention. Now their RL-level and parent training will become even greater - whilst the newly freed resources to solve for sub-optimizations that have been traditionally avoided due to computational overhead
This is somewhat untrue as well. HFT because constrained similarly have to optimize on this level akin to HFT crypto doing optimizations not within solidity, nor yul but on opcode in huff. That’s the issue with these big tech companies. Just endless budget and throw bad code into larger distributed clusters to overcompensate.
But, this is not just Amazon specific. Even when discussing on renewable energies (e.g solar), this is the same argument that most conservative lawmakers and utilities make about applying a tax on residential solar. You have users of the grid that are not paying the same as others, but it's still utilizing the same grid as everyone else, therefore increasing the cost for everyone else that are not discounted, etc.
With that being said, I don’t think the benchmarks we currently have are strong enough and the next frontier models are yet to come. I’m sure at this point U.S LLM research firms now understand their lack of infra/hardware optimizations (they just threw compute at the problem), they will begin paying closer attention. Now their RL-level and parent training will become even greater - whilst the newly freed resources to solve for sub-optimizations that have been traditionally avoided due to computational overhead