推文

@harden_369741 · 2026-10-13 03:37

AWS just showed you don't need to train a new architecture. you convert one. their 'Priming' method takes a pre-trained Transformer (Qwen3-8B/32B) and retrofits it into a hybrid state-space model, reusing the weights. pretraining cost down 99%. KV cache halved. decode at 128K twice as fast. open-sourced, Apache 2.0. I've paid enough training bills to feel this in my bones. the most expensive line item in AI was always pretraining, and AWS just turned it into a renovation project. architecture innovation is becoming a retrofit business. the moat isn't who trains biggest anymore, it's who converts smartest. #AI #opensource #LLM

曝光 52 · 评论 1 · 点赞 1 · 书签 0 · 曝光/时 6.412258394169126