Research notes

grok46-vs-grok45

打开交互式对比页

实验 worker baseline 最佳(validation) 最佳(held-out) 节点 完成/失败 停止 总耗时 token 输入/输出
kag-grok-46 grok:grok-4.6:high 3,587 24,614 (n008) 24,926 10/0 max_nodes 1h49m 3,989,018/591,613
kag-grok-45 grok:grok-4.5:high 3,587 37,329 (n008) 35,748 8/0 patience 1h00m 2,151,072/254,942

kag-grok-46 节点顺序

  • n002 done 16,217(best so far 16,217,20m31s):Early daily HIRE to multiply plant/water throughput
  • n001 done 9,525(best so far 16,217,21m48s):Multi-tile wheat farm with daily water/harvest route
  • n003 done 22,269(best so far 22,269,17m41s):Buy NE land and scale wheat
  • n004 done 20,736(best so far 22,269,20m42s):Fertilize wheat bonus window
  • n005 done 23,598(best so far 23,598,27m27s):Fertilize wheat bonus window
  • n006 done 21,683(best so far 23,598,26m38s):Compact spawn-aware 5x10 regions
  • n007 done 24,088(best so far 24,088,21m22s):Defer goose CARE on harvest days
  • n008 done 24,614(best so far 24,614,19m29s):Compact spawn-aware NW+NE regions
  • n009 done 22,086(best so far 24,614,8m50s):Skip late DROP while tiles need water
  • n010 done 24,575(best so far 24,614,20m32s):Fertilize wheat bonus window

kag-grok-45 节点顺序

  • n002 done 9,085(best so far 9,085,13m17s):Cheap daily hired hands for parallel farm ops
  • n001 done 8,177(best so far 9,085,19m01s):Multi-tile wheat field with daily route
  • n004 done 36,892(best so far 36,892,12m04s):Parallel melon (or melon+carrot) loop
  • n003 done 8,229(best so far 36,892,28m25s):Buy NE land, extend hand farm
  • n005 done 36,584(best so far 36,892,11m50s):Buy NE land under melon tile cap
  • n007 done 36,405(best so far 36,892,6m08s):More hands after NE unlock
  • n006 done 36,892(best so far 36,892,7m36s):Fertilize melons for age-8 max yield
  • n008 done 37,329(best so far 37,329,10m13s):Quadrant-local hand partition