[sarchertech]

I don’t know anyone who uses ChatGPT who cares about that stuff. Most people just use it as a Google replacement.

I actively hate it when it brings in some nonsense it thinks it knows about me. I told it my income once in an attempt to use it to find the perfect rewards credit card mix. Now anytime I try to get it to search for a deal it brings up some nonsense about “as a high income individual you don’t worry about saving $X, you care more about reliability, so you don’t need to look for the lowest cost” or something similar.

[arcanemachiner]

Turn off the "memory" feature in the settings. It just rots the context anyways.

[naishoya]

there is the option to opt-out of personalization, opt-out of using your conversations for training, and in that process reduce or eliminate any memory the system has of your personal preferences and context. If these actions don't erode your particular use-cases for ChatGPT, and if you think you can trust the model to follow these options this might be of use. I'm not trying to say "you're using it wrong" but that taking a more active control of the instaces facing you as a user might be of some benefit.

I have iterated through different option configurations to reach a level of 'customization' that more or less conforms to my own use case, and this does include opting out of any and all lasting memory between instances and across chat sessions; and adds a selection of single initialization prompts which shape the chatbot's behavior to my requirements for that session's objective. these trim most if not all af the sycophantic interactions, reduce outputs to the specific formats and contours as defined and omits any of the 'explanations of the underlying reasons behind...' which is just noise. This also has enabled some pretty useful results without ever spending a dime on a paid account: the premium behavior presented to 'potential customers' as a lure continues to work for me, and for iteration across instances and accounts is possible with machine-ready yaml context file when a single sessions hits the 90% wall : one emit and ingest cycle rotation across account profiles in firefox and i pick right up with a fresh limit.

Bouncing between ChatGPT and Claude, and between models for discrete subsets of larger tasks has really been impactful for my particular needs; but as i am not working in regions of knowledge that are beyond my own expertise and because I require the model to limit responses to very specific parameters, the logic space for unchecked hallucinations is low (but not zero).

The most useful project results for me have been in developing an air-gapped private menagerie of multi-domain models which uses an operating structure not dissimilar to OpenMythos; but then my background includes HPC environment development for NUMA, unikernels, MPI and bare metal hypervisor design - so getting a design plan and functional code without requiring a team of programmers and months of time in order to even start using models under my control which have zero public facing risk for the projects i'm working on is a much better place to spend limited budget on. Last gen hardware in the V100 class is perfectly capable of running and delivering the physics calculation optimizations as required and I would rather buy and/or install solar+storage to supply the electricity for token generation than rent the same from any of the frontier models AND trust that "don't train and learn from me" preferences are and continue to be followed.

If your use-case is a a 'lifestyle shopping assistant' then just turning off customization might be sufficient to stop it from telling you how to live your best life.