Trade Secrets

Most people use LLMs by interacting with a chatbot or agent interface.

Although LLMs generally work the same way and we know how these algorithms work, different chatbots and LLMs have features that we don't know about, which we call 'trade secrets'.

AI companies keep some details about their products secret. This includes exactly what training data was used and how RLHF was done.

Chatbots also filter and ‘pre-process’ your prompt even before it reaches the LLM, checking it, modifying it and sometimes sending it to a completely different model.

For example, if we ask to generate an image using the Gemini chatbot, it will send our prompt to NanoBanana (an image generation model), instead of an LLM. Chatbots can also add extra information to your prompt behind the scenes like instructions or safety guidelines that you never see but they shape your final response.