@mceveritts@theo So anyways, as I was saying...
CoT is most likely handled with higher priority by OpenAI and Anthropic, but we don't know without seeing the trace in real time, which they've hidden as I understand it, to reduce distillation attacks.
@BasaranOrhun This is so inefficient, but also amazing. All software design companies are going to be making MCP layers for their tools. Maybe even a new standard thats more powerful.
Seems like Claude tends to create Claude docs instead of files on a file system when I ask for hand-offs or migration docs. Almost like it doesn't want me to have access to them outside of the Claude ecosystem.
It is probably a design decision. There's no way Claude's CoT is the same speed as the output tokens you see when it "starts to respond". It probably has significantly more compute behind the CoT and Google either didn't consider that or are trying to get the basics first.
I have no first hand experience with OpenAI products yet, but I imagine its a similar situation. Grok 4.7 for example feels similarly slow because you can see the CoT traces and its the same speed as the final answer.
I'm catching Claude Code default selecting Haiku. I run CC in tmux, then /rc and disconnect, then continue from web browser in https://t.co/yVoKU10od4 --> code.
I hope this is not intentional behavior. Made quite a few changes before I caught it.
#ClaudeCode
@theo I am more miserable since using it. My projects are moving along. I just wish the speech patterns werent so insufferable.
Barely a step above legalese.
New found respect for lawyers... (maybe the paralegals actually).