GPT-4o Feedback aka 'Ultra-Utility'

Couldn’t find the right spot for this, so here ye.
4o is forgetting documents from much earlier in the conversation.

  • add sidebar to allow user to add documents relating to chat, and allow 4o to update them, and 4o should keep referencing them. I found adding a group of documents for code context was really useful - but it lost it’s way. Solidifying the usefulness of this by keeping them in mind and not forgetting would be great.
  • add wrench icon that opens guidelines and rules to follow for the current chat (like custom gpts). 4o should not stray from these and should re-reference along the way. I’ve found 4o forgetting things we’ve spoken about noted as rules.
  • as per usual, find a way to actually stick to rules place by user. While improvement has been massive, i did some code check tennis with Claude and it reminded me how it (gpt4) doesn’t like to remember to do what you say (don’t shortcut and leave stuff out) - the point here is nail that point home. Don’t do it means don’t do it.
  • 4o likes to spit stuff out at lightspeed which isn’t actually useful when in the context of the user requesting a step-by-step guide. It’s cool that it’s fast but we gotta read it to - and we don’t like scrolling bucketloads to find something. > Perhaps offer the opp for the end user to go step by step waiting for an ‘ok’. Feels like we shouldn’t need to add these parameters anymore, as it’s a given. A chef doesn’t word-vomit the recipe and nuances on the counter and run out the door. {or do they} We’re here to help, after all.
  • add a reply button to each 4o response. User can add comments or questions to which 4o would respond taking into account the response window that the user hit reply on. I found it got lost when i copied stuff from way back.
  • add a pin button to a 4o response for later reference.
  • add a tag to each pinned response.
  • add ability for user to search entire chat and also by tags.
  • for code (or writing or anything visual) add comparison. side by side code comparison to see what was changed – we’re still putting our trust in machine faith and it’s still not flawless. this would mitigate that fear a little…
  • add topic search within conversation
  • add topic or subject export and sharing for sending to colleague so they get the context without having to read the back and forth - this would group segments for easy digestion.
    i could go on but i’m tired.
    Goodnight austin texas, wherever you are!
1 Like

I like your thorough proposals there, and a lot of that is not even specific to 4o, could make ChatGPT much better in general.

However I’m a bit saddened to see that OpenAI did not do anything to improve factual accuracy (some basic domain-specific RAG fact checking for scientific and technical knowledge?) and towards improved reasoning and fact-based conversation capabilities.

I am trying to use ChatGPT for scientific/technical work and learning, and it’s often disappointing with the high level of hallucination and lack of self-check for the some of blatantly made up facts, and the conversations running into the issues you mentioned.

GPT-4o seems rather fine-tuned for quick, superficial, verbose conversational skills. Very impressive if you’re looking for realistic, superficial human simulacra that sounds great and pleasant at parties… or to lonely people.