It used to happen before, was fixed for a while and now the issue is back. When asked about some information involving web search performed by the assistant (only checked Claude 3.7 at the moment) in subsequent responses it will insist that I, the user, am "showing" it some information, and it confuses the machine a lot.
See sample dialogue:
Two very important points
- It absolutely has to know when a piece of information is coming from me and not from its own search attempts
- It must be able to approach those "attachments" and "snippets" in a reasonable way: our dialogue is moving on, yet those files somehow poison my every prompt. The machine thinks basically that I keep uploading the same files over and over with each prompt while in reality we have already moved on. So those should serve as a reference on a as-needed basis, but not as active parts of every subsequent prompt.
