IMO Grok and 4.5 show the we've reached the end of reasonable pre-training scaling. We'll see how far we can get with RL in post-training but I suspect we're p

IMO Grok and 4.5 show the we've reached the end of reasonable pre-training scaling. We'll see how far we can get with RL in post-training but I suspect we're pretty close to maxed there and will start seeing diminishing returns. The rest is just inference efficiency, porting the gains to smaller models, and building the right app-layer infrastructure to take advantage of the technology. I do think we're overbuilding on Nvidia and the CUDA moat isn't as big as people think, inference workloads will dominate, and purpose-built inference accelerators will be preferred in the next hardware-cycle.