Maybe the Chinese app is different, but the Kimi app shows reasoning traces so how can they route the request to Claude models which don't show reasoning trace and still provide a reasoning trace on the UI?
So I think they are in fact just lying because it doesn't make sense why they would route requests to Claude over their own model.
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.
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.
Comments
Is this covered in their TOS or is this a breach of contract?
at this point the worlds pants supply is dwindling from all these fires
Maybe the Chinese app is different, but the Kimi app shows reasoning traces so how can they route the request to Claude models which don't show reasoning trace and still provide a reasoning trace on the UI?
So I think they are in fact just lying because it doesn't make sense why they would route requests to Claude over their own model.
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.
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.