It is not a risk is a fact - people decompiling Claude Code have found many times that it has code branchs to detect it is being used in Chinese timezone and locale.
Zhipu AI is founded by a superstar Tsinghua professor, did an IPO in January (Hong Kong stock exchange) hired half it's past research lab and it's stock is >10x since. This is not a "just distill Claude" thing.
I am thinking that a small tool that simply refuses to pass large CLI output to the LLM and warns it to filter the results before reading would achieve this better as the LLM would be forced into thinking and writting the filter itself.
Wait what!? I have been programming CUDA since 2009 and specifically remember it being pushed to C++ as main development language for the first few years, after a brief "CUDA C extension" period.
Anthropic is a great showing for startup founders how if you have a great product people will buy it, even if they dislike your pricing, your marketing and the CEO opinions.
Real PMF sells itself. The risk is of course the competition catching up, I bet switching costs are very low on this setup.
The harness is the model "body", it's weight the cognition. Like in nature they develop together and the iteration of natural selection works at both.
If smaller labs (Zai, Moonshot, deepseek, mistral..) get together and embrace a harness, like opencode for example, as a consortium just by the power of "evolution across different environments" they might hit jackpot earlier than bigger labs.
My experience trying LanceDB has been abysmal. It worked great on dev and small testing environments but as soon we tried production workloads it would get extremely slow. We shifted to PostgreSQL + pgvector and had absolutely no issues, even if it is not "engineered for multimodal data". Maybe we were doing something wrong but we did put effort in trying to make it work - it is this hard to get it performant?