I use mine extensively, to the point where I don’t even use the plain ChatGPT mode any more. GPTs greatly enhance my experience in ways that don’t only matter for entertainment.
I settled on a curious mix of custom GPTs that are either task-specific, methodology-specific, or area of interest-specific.
Task-specific GPTs is a pretty straightforward category. Those essentially are designed to do one thing, and do it well (very much like the Unix philosophy). One of them is designed for OCR-ing a really large, scanned historical PDF manual, and will eventually be changed to become the query GPT to interface with the new scanned version of that PDF. While I could do this with the API, the manual is several hundred pages long and I’d rather use a chunk of my $20 a month subscription whenever I’m not using. This is not just straight OCR - this GPT contains a set of particular directives that assure that the output will have a certain consistency, and that images, screenshots and diagrams are captioned into a text transcription in a specific style that covers different details that the LLM might not otherwise think to cover.
Another one is designed to query a different document - this time a car manual, which was designed to help a family member learn about how to deal with recovering from various faults that may occur. This one’s prompt contains additional logic to flag whenever the answer did not come directly from the manual, but instead had to be extrapolated from what the LLM already knows. It can also serve a download link to the full PDF for download, and tell you which page the information came from.
Methodology-specific GPTs are trickier. They are simulations of a certain type of cognitive archetype, done through the medium of forcing the model to enter into a long monologue that follows a certain self-steered path. Because LLMs to some extent can think out loud, this often means that any of the four archetypical personas I developed can come to a different conclusion than any of the others, or arrive at the same conclusion via different paths. In addition to pretty much any of them causing the model to subjectively become around 2x - 5x smarter than the base ChatGPT 4 (depending on the task), the diversity in thought between them can be harvested to extract interesting analogies, arguments etc that would normally not occur to me.
Area of interest specific GPTs - quite simply, a GPT pre-railroaded into assuming a particular context in a conversation. For instance, a self-hosting GPT that will automatically know what my home infrastructure and preferences are, or a cooking GPT that will address my particular dietary restrictions without having to mention it every time.
I am currently experimenting with GPTs that can invoke external tools. For instance, I am planning to at some point give one of my GPTs access to a real, online Linux machine (the lack of software packages in the Python sandbox is an annoyance), calendar APIs, weather and so on.
I would share few of these GPTs with others, unless someone asked. They all take getting used to how they work, because they push at the boundaries of GPT-4’s highly defensive alignment to do their job. It’s pretty much alpha technology at this stage. I feel like it would be much easier if these GPTs could have some simple Python wrappers around them to more easily and reliably control the flow of their execution, but that’s something I’m planning to do with Llama 3 once I become more familiar with it.
I settled on a curious mix of custom GPTs that are either task-specific, methodology-specific, or area of interest-specific.
Task-specific GPTs is a pretty straightforward category. Those essentially are designed to do one thing, and do it well (very much like the Unix philosophy). One of them is designed for OCR-ing a really large, scanned historical PDF manual, and will eventually be changed to become the query GPT to interface with the new scanned version of that PDF. While I could do this with the API, the manual is several hundred pages long and I’d rather use a chunk of my $20 a month subscription whenever I’m not using. This is not just straight OCR - this GPT contains a set of particular directives that assure that the output will have a certain consistency, and that images, screenshots and diagrams are captioned into a text transcription in a specific style that covers different details that the LLM might not otherwise think to cover.
Another one is designed to query a different document - this time a car manual, which was designed to help a family member learn about how to deal with recovering from various faults that may occur. This one’s prompt contains additional logic to flag whenever the answer did not come directly from the manual, but instead had to be extrapolated from what the LLM already knows. It can also serve a download link to the full PDF for download, and tell you which page the information came from.
Methodology-specific GPTs are trickier. They are simulations of a certain type of cognitive archetype, done through the medium of forcing the model to enter into a long monologue that follows a certain self-steered path. Because LLMs to some extent can think out loud, this often means that any of the four archetypical personas I developed can come to a different conclusion than any of the others, or arrive at the same conclusion via different paths. In addition to pretty much any of them causing the model to subjectively become around 2x - 5x smarter than the base ChatGPT 4 (depending on the task), the diversity in thought between them can be harvested to extract interesting analogies, arguments etc that would normally not occur to me.
Area of interest specific GPTs - quite simply, a GPT pre-railroaded into assuming a particular context in a conversation. For instance, a self-hosting GPT that will automatically know what my home infrastructure and preferences are, or a cooking GPT that will address my particular dietary restrictions without having to mention it every time.
I am currently experimenting with GPTs that can invoke external tools. For instance, I am planning to at some point give one of my GPTs access to a real, online Linux machine (the lack of software packages in the Python sandbox is an annoyance), calendar APIs, weather and so on.
I would share few of these GPTs with others, unless someone asked. They all take getting used to how they work, because they push at the boundaries of GPT-4’s highly defensive alignment to do their job. It’s pretty much alpha technology at this stage. I feel like it would be much easier if these GPTs could have some simple Python wrappers around them to more easily and reliably control the flow of their execution, but that’s something I’m planning to do with Llama 3 once I become more familiar with it.