Lisan al Gaib· @scaling01 · X·· 3 小时前AI 评分22
AI 导读
我觉得这很好地提醒了我们，距离真正的超级智能还有多远，但通往它的路径又是多么直接和明显： - 语言压缩是不可避免的。市场要求效率，更大的模型天然就在压缩。模型会越来越像神经语言 - 推理速度只会不断上升。未来几年内，前沿模型将达到 10k tks/s - 多智能体训练仍处于襁褓阶段，尤其是对开源模型而言
正文 · 原文
I think this is a pretty good reminder how far we are from actual superintelligence, but how direct and obvious that path towards it is:
- language compression is inevitable. the market demands efficiency and larger models are naturally compressing. models will get more neuralese like
- inference speeds just keep going up. we are going to reach 10k tks/s for frontier models in the next few years
- multi-agent training is still in its diapers, especially for open models
DeepSeek V4.1 Flash 1017 runs at 1,521+ tok/s on Fireworks Ultrafast. Fast enough to ship Tetris in 11 seconds, 7x faster than Claude Opus 5.5. Join the Ultrafast waitlist: https://bit.ly/3UA0wKq在 X 查看被引用的帖子
来源：Lisan al Gaib · x.com