[ting0]

So basically there's no excuse not to see ChatGPT and Claude release 10M -> 100M models within 6months or so. <9% degradation is crazy. Hopefully DeepSeek and Qwen4 can implement this.