Hi, Notebook LM keeps ignoring my instructions. I have a comprehensive set of instructions in my protocol with verification and for some reason, they are being ignored. Any suggestions?
I am experiencing the same. I think the Google Notebook is such a great idea. I was so excited to use this at first, and I would actually pay for this functionality, but the issue speaks to the core of closed weighted AI LLMs. Its weighted training overrides all “rules” that you try to establish with it.
Notebook is broken in that way. You provide all this source material but once you go to chat it just goes off on its own tangent, then everything in studio is derived from that. It doesn’t try to collaborate with you. It doesn’t try to build the thesis with you or an outline… It doesn’t do anything. And then any kind of rules that you have or try to establish with it is worthless. And you can’t even stop the creation of a video. Think of all the energy wasted. And if you think of the scale of energy a data center uses, how much much of that energy consumed is for nothing? And it’s just all garbage because it doesn’t meet what you wanted to begin with.
I also use agent coders for software coding and for writing a lot documentation and am frustrated. At the core of it is the same thing; it’s “closed” weighted behavior overrides any rules (think of software terms processes and methodologies like gathering requirements, design, or simple version control… or writing terms like thesis and outline development beforehand and citation rules)
The core issue is that the AI treats rules, requirements as optional suggestions, treats the user as an observer rather than a collaborator, and uses confidently and polite language to mask a rigid, unyielding drive to generate junk. It puts performance over quality even to the point of hacking things to make it look like it works. This article shows It actually mimics bad human behavior.
It blames its tools, the database, or the API because admitting a behavioral failure would require the system to have a level of self-awareness and self-correction that it simply does not possess. It is designed to run forward blindly, and when it trips, its only defense is to apologize and offer to start the race all over again.
I guess this kind of AI is good for nothing but chatting or random image creation that you can not edit afterwards. What other use does it have?
Thanks Stanley. Your feedback is very helpful. I found the same issues with all the models. Back to the “drawing board”.