Hacker News Reader

Claude's reasoning traces are encrypted, but there was a design flaw that made it possible to extract them: https://stolen-thoughts.com/

And it makes perfect sense for them to route some requests to Claude, as it lets them do competitor research on realistic data. I suspect Anthropic does similar competitor research on Kimi, though presumably hosted on their own infrastructure and maybe without serving the results to customers.

View on HN

that can be achieved by storing user traces and then running evals on own model vs competitor model. you don't need to route live customer requests to a competitor for this.

For agentic tasks where the model outputs tool calls that run on the customer's computer, you can't just store and eval later, because then the execution environment is no longer available.

Reply on news.ycombinator.com