YES: Harness Self-optimization w/ 9B LLM (Local AI)

A well-established empirical study. Good job!

My experience is contrary to the research. Gpt-oss-120b works much worse than recent 30B models, such as gemma 4 31B.

What was the precision / quantization of the Qwen 32B?

Interesting. Well Metaharness study the used a 9B model. Not sure what it was in SIA

'good update from error' >> perfect way to enhance/improve a system. The "You're an expert.." part of prompt is likely to make gpt 5.5 and opus 4.7/4.8 a little less efficient during compute.

How does this stack up with the 4 layer harness? Or how can we actually get deliverables out of all these college thesis papers?

Ive had this concept for a long time now. Its hard to get traction on it outside of personal use. But i suspect now that cost is becoming a thing, enterprises will have to stop feigning ignorance on the topic.

You be just so happy to hear what your research engines done brought up. Yo, do you do the research yourself? I mean either way either you set it up or are you paying for it? Either way you got a good smart brain to be posted on YouTube I mean you stay right where you need to be. I can’t hold you. I can’t hold you.

qwen3.6:27b would do better i think...

Instead of baking the harness in just make them easier to use by AI. OpenAI set a bad example by forcing it with training. just design it smarter.