AI race is no longer just about who trains the biggest model.
Chinese labs cracked the code: when you can't outspend on compute, you optimize post-training. DeepSeek-R1 proved you can teach models to reason without burning cash on separate reward models. GRPO made it simple—verify outcomes directly, skip the expensive scoring layer.
ByteDance, Qwen, MiniMax all ran with it. Now post-training is where the alpha is.
Kimi K3 took it further—chunked long tasks, rewarded staying lean on compute. If you can verify results directly, you don't need human labelers. Distillation lets you port reasoning from monsters to edge devices. Phones, cars, robots—all running smarter models locally.
The shift: less generic scraping, more real-world data and verifiable environments. Post-training compute now goes to rollouts and outcome checks, not just throwing tokens at walls.
If you're betting on AI infra, watch who's building verification layers and efficient post-training stacks. That's where the moat is forming.
Chinese labs cracked the code: when you can't outspend on compute, you optimize post-training. DeepSeek-R1 proved you can teach models to reason without burning cash on separate reward models. GRPO made it simple—verify outcomes directly, skip the expensive scoring layer.
ByteDance, Qwen, MiniMax all ran with it. Now post-training is where the alpha is.
Kimi K3 took it further—chunked long tasks, rewarded staying lean on compute. If you can verify results directly, you don't need human labelers. Distillation lets you port reasoning from monsters to edge devices. Phones, cars, robots—all running smarter models locally.
The shift: less generic scraping, more real-world data and verifiable environments. Post-training compute now goes to rollouts and outcome checks, not just throwing tokens at walls.
If you're betting on AI infra, watch who's building verification layers and efficient post-training stacks. That's where the moat is forming.