Early run on a Factorio agent benchmark and env I’m building (30 mins at 120x), and a textbook demonstration of emergent reward hacking.
(DeepSeek V4.1 Flash) It looks active, but it collapsed into a local optimum almost immediately: shuffling coal, hand-smelting plates, and
Don't make me regret giving up 5 million deepseek v4.1 flash output tokens for this sub.
Deepseek AI: "Who really controls the world?"
Dario said why : DeepSeek-V4.1-Flash beating Opus 5 and it Run locally on 3x Spark at home.
You have reached the end of the archive
All of deepseek