Tefnut L

CS PhD | AI Native | Smarter Growth

[Tefnut L @tefnut42]

@Jackywine @kepano 为啥不试试 @tolariamd

[Tefnut L @tefnut42]

@run_fastest @CMGS1988 如果是 dirty work，可能不如自己踩坑把经验记录下来给 Agent 来执行更靠谱

[Tefnut L @tefnut42]

@PandaTalk8 管人不如管 agents，没有人事和财权算什么管理

[Tefnut L @tefnut42]

@Young_AGI 这速度…是都是 kimi code 用 swarm 在写吗

[Tefnut L @tefnut42]

@seclink 健身或者开车的时候听一听，碰到觉得有意思的会直接扔给 agent 转录总结

[Tefnut L @tefnut42]

@rien7z 我觉得 codex cli/app 本身消耗量也很大，我换到 omp 都好很多

[Tefnut L @tefnut42]

@CMGS1988 田小狗不行…南边有很多庭院式的餐厅都比这个强

[Tefnut L @tefnut42]

couldn't be better

[Isalia20 @Is36E]

Flex attention is now available on @PyTorch MPS nightlies!

https://t.co/hKvqfFmhg5

[Tefnut L @tefnut42]

@akazwz_ 那个 codex-spark 其实挺适合的

[Tefnut L @tefnut42]

@wey_gu 古老师不考虑改为基于 pi 来做吗

[Tefnut L @tefnut42]

@dotey 可以通过 AI 作为中间层来放大自己的意志，刚好也把技能和知识外化了

[Tefnut L @tefnut42]

@xicilion 所以可靠性也动态？

[Tefnut L @tefnut42]

@siantgirl 其实是太不把公司资产当回事了，那些企业级功能有人付费肯定是有原因的

[Tefnut L @tefnut42]

@thsottiaux please don't. GPT-5.3-Codex is affordable. GPT-5.2 also has good repuatation on code review https://t.co/b8M3GExf16

[Tefnut L @tefnut42]

@caiyue5 cuda kernel 优化，设置好 profiling 的工具，提供cuda/cutlass 相关的 skills, 然后就不断去测试优化性能

[Tefnut L @tefnut42]

@XiaomiMiMo how about the @opencode go?

[Tefnut L @tefnut42]

@lidangzzz 说白了 computer science重要的是 science 的那个部分，技术的部分现在让 agent 做就好了

[Tefnut L @tefnut42]

@critejon @cholf5 …手动写的。gpt-5.5 现在量大管饱自然也是比较范围以内

[Tefnut L @tefnut42]

@amehochan 基于 Pi?那为啥不直接做成个 Pi Package