I tried to understand GPTs yesterday. What are they really, and how can we best utilize this tool? Ultimately, they are just tools. We tell them what to do, and with the power of GPT-4 and the additional guardrails you provide, they give us an output. They have limitations. They can't do anything that violates their content restriction policy, and they probably won't do something if they were instructed not to do it to begin with. This is why the description of GPTs is useful. Sometimes less is more when it comes to tools. "Do one thing and do it well" was the Linux philosophy. The power of GPTs comes within a conversation. You can @ mention GPTs. One limitation that I have found is that you can only speak to one GPT at a time. Where it gets confusing is when you're speaking to the GPT and asking it to do something, which you may think is like speaking to an AI agent, right? Well, no, not really. AI agents are something different altogether. What they are meant to do is perform tasks autonomously, whereas with GPTs, you prompt the agents one at a time to do things. You can chain one or more of the outputs and use them as inputs when you @ mention another GPT, which I think was the whole point of the GPTs. The GPTs tend to forget their initial instructions as the conversation drags on, so it's best to use short one-shot prompting. Otherwise, the results will be similar to as if you just asked regular GPT-4. Using GPT-4 is expensive. Getting a lot of output takes up a lot of bandwidth, and this will limit how much and how long you can use the premium service. Yes, even though you paid $20 US a month for the privilege to use it.
Last Updated: 24/03/2024
Author: Michael Huynh