I did a survey with my friends and found that everyone just uses chatgpt, only tried gpts for a while but no one keeps using any. Same thing with plugin, we tried a lot of plugins but no one sticks to any.
Anyone uses a custom gpt regularly? Can you raise your hand and tell us what you are using and how important it is to you?
I exclusively use custom GPTs and if I lost them I would honestly feel the loss pretty hard. Whilst they’re a bit weird to get working, they’re far superior to the regular GPT4 if you take the time to configure them to your needs.
I built a couple when they first came out and haven’t had time to really explore their advantages. So, I don’t currently use them, but I probably will once I have time to build an API endpoint and make custom actions.
There’s potential there, particularly if you implement enhancements through it’s ability to use API calls/custom actions. Adding a continuous memory, adding a persistent goal and a to-do list to start to get goal directed behavior, a feedback loop asking it to evaluate it’s progress after each message.
Can you recommend any to me? I’m really curious how superior it could be to GPT 4
I keep hitting the usage cap. But only for the GPT, not normal ChatGPT 4. Makes it pretty useless.
I’m definitely going to be using it once they figure out server load tho. I’ve made an action that gives the GPT access to local files, so it can go through a project and modify multiple files without user intervention.
Technically I agree with what you said. And it’s still kinda early for apps to get equipped with the memory systems, etc. you mentioned. However, when looking at plugin, which has also been there for quite a while - it means developers have sufficient time to build any infrastructure behind - they still don’t appear to be very useful in my case. So I’d like to see if others feel any of the plugins/gpts useful
I only use custom private ones I’m afraid, so don’t really have any recommendations. I certainly recommend making your own custom private ones though, I used to configure them using those lame ‘How GPT should interact with you’ boxes, and GPTs are just far more powerful versions of that… plus the potential for Actions, which expands the customization ridiculously.
Yeah, I have experienced the same personally. I only use ChatGPT for 99% of the time.
There are still not GPTs that make me really want to use them.
I think if someone manages to create a great game, I would spend some time on it, but haven’t found anything that I really like so far.
I don’t think it’s still early. GPT-4 really makes me dive deep into the philosophy of Consciousness. It’s clearly not, but it can do things we used to think were exclusively in the domain of sentient entities. I’ve experimented with a rudimentary memory system, split into txt files depending on type of memory. narrative_memory.txt, domain_specific.txt, etc. For file size management, it’s task when retrieving memory from a file is to comprehensively summarize memories older than x days old, keep a memory of sentences or paragraphs that have a high sentiment score (and any needed for context). GPT-4 with a carefully designed memory is scary intelligent, capable of directing its own actions by analyzing the main goal, what it has ticked off of a to-do list, and what it might need to do given the feedback received so far (by a screenshot feedback system). It’s almost like giving it memory enables it to have a simulated inner dialogue, keeping itself on task and giving itself feedback. It learns from mistakes. It’s… weird.
I have already created a to-do list API, a file manipulation API, and am working on a continuous memory API. I’ve not had this much fun coding in a long time. I just keep building APIs to see what it can do with them. We’ll see if anything useful comes out of it though.
Hit me up if you want to collaborate. Actually working on a python module and an instructional series to give GPT-4 a robot body right from the chat interface using micropython running on a pi pico and a lot of other components.
Wow, sounds awesome, that’d be seriously impressive. I’m just throwing together .Net Core APIs hooked up to SQL databases, hosting them, and giving them to my custom GPT. Its crazy how much of a difference it makes giving it even basic APIs. My memory one is a bit more ambitious though, allowing it to store concepts and theories and update its certainty surrounding them based on new input. Might be too much for it, but that’s why we prototype, eh?
Eh, that sounds hairy. I could see using something like word2vec (more likely a custom implementation) to represent concepts in n-dimemsional knowledge/information graphs, a mechanism to have gpt analyze the veracity of a theory or concept, assign a confidence score, a mechanism to retrieve the concept/theory. Problem I see is that we don’t understand how our own thoughts are formed and retrieved. I think it’s likely that we are constantly thinking under the surface, recalling memories randomly, etc. in a loop, and our consciousness is just a mechanism to evaluate each thought, assign a relevance to it, and allow relevant thoughts and mental images into conscious awareness and filter out everything else. That could be approximated with an advanced system like gpt-4, with multiple instances carrying out specific roles such as memory retrieval, assignation of relevance values, filtering “internal” and external inputs by relevance, determine goals, keep track of progress toward goals, rate input by significance (similar how traumatic experiences are encoded much easier in our memory than mundane experiences). Hm… I might do something like that.
Good luck with your experimentation Cool to see other people realizing just how revolutionary this technology is
I’m trying to avoid thinking about how we think too much, and just plan on providing it with the tools and seeing what it does with them, then tweaking them based on the results. Ideally I see my API as essentially being long term memory, and it utilizing its context window as short term memory, whilst being able to tell it to spend some time ‘sleeping’ which would be when it goes through and updates the certainties of its various conceptual relationships. I like your idea of a separate GPT for that though, probably makes it cleaner.
This is what GPT can do when it is given instructions and a continuous memory with just a text file. It remembers every prior conversation, with more recent memory retained in more detail than older memories. It remembers new knowledge it learns, learns from mistakes, and can engage in goal-directed behavior. It’s remarkable.
My new functionality might not serve any purpose, but it’s amazing watching its certainty that ‘Ocean Ducks eat Sharks’ and that ‘Callum Is A Seal’ creep up from the initial 0.5% certainty up towards 20% as I provide it with more information. I can’t wait for it to start actually believing these things.
I did it!
I nuked my data and started again, and have convinced it I am trapped in Middle Earth, but that somehow my Laptop is still working. Watching it slowly start to believe me, and how its language changes based on that, is fascinating. It really doesn’t want me staying on the roads, and thinks I should head to Rivendell as soon as possible.
This is why Custom GPTs are amazing.
Wait, what did you do to it? Is it just prompting or something else?
It has instructions to identify information within user submitted text, and then parse it into logical constructs, before submitting them. When it identifies concepts within the data, it queries my API for references, and retrieves them along with their current level of certainty. If it changes its mind about the certainty, it updates it. All of its reactions prioritise information retrieved from my API, as it judged it itself so trusts it.
… Of course there’s nothing stopping me from overwriting those certainties with a little bit of SQL, but I am respecting its choices.
I think I might be up all night playing with this
I have tested many GPTs.
ChatGPT will remain the most popular GPT for general use. While GPTs are great for certain tasks, ChatGPT seems to be the go-to for most people.
For me personally, I use ChatGPT about 70% of the time, especially for programming-related queries. The other 30% of the time, I turn to ‘BasedAnswer’ for general questions and to learn about specific topics.
Additionally, I use ‘NewsGPT’ either daily or every other day.
I wanted to use GPT’s with my students to allow them to use ChatGPT as a tutor, and to make Quizzes to help them learn topics etc. However, when I created GPTs they needed to access a paid version of OpenAI-plus to take advantage of them. This sort of defeats the purpose.
The workaround is to simply have them cut and paste prompts from the Learning Management System assignments or discussions. However, I would prefer to have them simply access a link to the GPTs even if I was charged for the use.
I’m a teacher not a computer scientist. Can anyone tell me how to set up my system to be able to give anyone with the link to my GPTs access without an account?