alt Hacker NewsI have a long term goal of establishing a workflow or framework wherein AI operates at my pace. I suspect much of the angst I experience when working with AI is driven by its independent working that I will be held accountable for.
One simple change that has really helped me is to take the git tool access away from the AI. This has made me quite critical of AI excesses in the design of the code and I feel free to rollback its changes and to request redos. I grok the changesets better as a result.
I’ve been thinking for a while that these tools should be slower. This doesn’t mean lower quality, they could still think (ah wait, overthink) at same speed. I’d just like them to output text at the same speed I can read and understand. This would make it more like a conversation, where you hear the flow of concepts coming and at the same time think about it and get involved in the conversation. For some reason I enjoy more running a local LLM at 4 t/s than a faster than thought change to my whole codebase by Opus