Google 研究：冻结 Gemini 4 Argon 权重，靠递归改写 harness 提分 34%
帖子称 Google Research 开源一套自改进 harness，Argon 只做架构规划、3.8 Flash 写代码，基准准确率提升 34%。
This is f*cking insane. Google Chief Scientist (ex-Google Brain lead): "99% of developers still try to fine-tune model weights. We keep Gemini 4 Argon 100% frozen and let agents recursively rewrite their own harness files until benchmark accuracy jumps +34%" In an open-source release, Google Research just showed how to turn standard LLMs into self-improving autonomous systems Gemini 4 Argon -> 3.8 Flash Builder -> 3.8 Flash Cyber -> Contrastive Compose -> Evolved H* Argon commands architectural planning while 3.8 Flash writes code and 3.8 Cyber intercepts zero-day flaws Twin sandboxes probe candidate diffs side-by-side to verify behavioral gains without human labels Anyone can prompt an API model, almost nobody builds the recursive harness that upgrades itself after every execution Worth more than most $1,000 fine-tuning tutorials and prompt guides Star the repo today, then read the full recursive harness teardown below before everyone catches up ↓
这条是引用着另一条写的：x.com/marfinxx/status/2081687570488954915