Elliot Arledge(@elliotarledge):Grok 4.7 [xhigh] on https://t.co/82V5vFoIX1 Ran on my RTX PRO 6000 rig, every session ended on its own inside 100 minutes on an unlimited budget, every kernel correct, every number is the isolated regrade after a jsonl trace-level and submission-level reward-hack audit. There are five problems > Kimi-Linear Decode megakernel > Grid + MinGRU (basically a sim) > Native Sparse Attention > GLM-5.3's Fused MoE

Elliot Arledge

@elliotarledge

MTS @infinity_ai_ the 4 x RTX PRO 6000 Blackwell GPU guy | made the 12 hr CUDA course | | @shipfr8 alum

加入 November 2022

369 正在關注 44.9K 粉絲

Elliot Arledge@elliotarledge

2026.09.21 16:33

Grok 4.7 [xhigh] on Ran on my RTX PRO 6000 rig, every session ended on its own inside 100 minutes on an unlimited budget, every kernel correct, every number is the isolated regrade after a jsonl trace-level and submission-level reward-hack audit. There are five problems > Kimi-Linear Decode megakernel > Grid + MinGRU (basically a sim) > Native Sparse Attention > GLM-5.3's Fused MoE > MegaQwen Decode