I love this. I am going to go out and try to build some of these!
One use case I've found just through the ChatGPT frontend is summarizing long conversation threads. In some mediums - Reddit, for instance - the valuable nuggets of information are not in the original post, but buried in potentially hundreds of comments. I've been using it on Robin Hanson's "Overcoming Bias" Substack, checking the output by reading the comment threads myself, and ChatGPT is quite good at:
- identifying participants in the comment thread
- summarizing conversations, and who said what
- developing profiles of regular posters from the way they comment, the things they talk about, the directions they want to take conversations
- describing the relationships between commenters (A is never on topic, B always is nagging A to stay on topic)
Great list of ideas! A related one I had is sort of in the "Enterprise AI" space, but instead of or in addition to directly answering questions, it connects you with people. "I see you're complaining about the XYZ tool! You should chat with Jane in accounting -- she was complaining about the same thing last week!" or "I see you're wondering about how ABC works -- I can't find a good explainer, but I think Alice in Engineering implemented that and could give you a quick answer!"
Hey, love it, I'm a web developer (in Brooklyn) and I think I'll take on one of these as a weekend project if it's ok. The AITA one sounds like fun, and pretty straightforward to start. It'd have a pretty broad range of potential uses too.
So well put. I have been wondering about the same but never got to expressing this anywhere rather than in my head. So, thank you for doing this! This post should be on top of the leaderboard! :)
Thanks for great inspiring examples for what LLMs can actually do for your average person!
But wow, how do you manage to write such long posts?
I have become interested as of late with the potential to restructure various work processes in traditional industries with ai. For example, I’ve been talking with a company which uses ai tooling to optimize sales processes in the oil and gas sector. Basically replacing a lot of paper-based systems with ai.
FWIW I'm aware of a ton of dispute resolution applications being developed in the peace tech / responsible tech space. I'm curious about their ultimate effects.
There already is a pretty neat tool for summarisation and extraction of info. Not perfect but can be quite useful. I use it to summarise long podcasts, or youtube interviews. https://github.com/danielmiessler/fabric
I've been thinking about LLMs that don't try to answer every question for a while. Doing things like "assess the target audience" or "explain the scientific/historical concepts that are background knowledge to this text" that aren't intended to be general-purpose systems.
As far as color-coding: what I want is for the LLM to use these tokens in its own output to give a more structured response. It can do a decent job with directions (such as (<green> for technical explanations) or (<yellow> for direct quotes)), but a more standardized system that is used in the training data would do better.
I love the plain language idea! It isn't glamorous, but I am sure it would be a killer if anyone could make it work.
It's actually surprisingly difficult to translate into plain language. LLMs typically learn by mimicking humans, and in general, humans aren't so good at communicating complex things in a way that is easy to understand.
The use-case I'm excited about is finally scaling effective psychological growth methods. (Most therapy is bad, and sooo much more is possible.) https://chrislakin.blog/p/alignment-plan
I think I could do about half of what I do to help others with AI now. Still waiting for voice-to-voice models though.
I love this. I am going to go out and try to build some of these!
One use case I've found just through the ChatGPT frontend is summarizing long conversation threads. In some mediums - Reddit, for instance - the valuable nuggets of information are not in the original post, but buried in potentially hundreds of comments. I've been using it on Robin Hanson's "Overcoming Bias" Substack, checking the output by reading the comment threads myself, and ChatGPT is quite good at:
- identifying participants in the comment thread
- summarizing conversations, and who said what
- developing profiles of regular posters from the way they comment, the things they talk about, the directions they want to take conversations
- describing the relationships between commenters (A is never on topic, B always is nagging A to stay on topic)
Great list of ideas! A related one I had is sort of in the "Enterprise AI" space, but instead of or in addition to directly answering questions, it connects you with people. "I see you're complaining about the XYZ tool! You should chat with Jane in accounting -- she was complaining about the same thing last week!" or "I see you're wondering about how ABC works -- I can't find a good explainer, but I think Alice in Engineering implemented that and could give you a quick answer!"
Hey, love it, I'm a web developer (in Brooklyn) and I think I'll take on one of these as a weekend project if it's ok. The AITA one sounds like fun, and pretty straightforward to start. It'd have a pretty broad range of potential uses too.
yay!
So well put. I have been wondering about the same but never got to expressing this anywhere rather than in my head. So, thank you for doing this! This post should be on top of the leaderboard! :)
Thanks for great inspiring examples for what LLMs can actually do for your average person!
But wow, how do you manage to write such long posts?
Thanks again!
I have become interested as of late with the potential to restructure various work processes in traditional industries with ai. For example, I’ve been talking with a company which uses ai tooling to optimize sales processes in the oil and gas sector. Basically replacing a lot of paper-based systems with ai.
FWIW I'm aware of a ton of dispute resolution applications being developed in the peace tech / responsible tech space. I'm curious about their ultimate effects.
There already is a pretty neat tool for summarisation and extraction of info. Not perfect but can be quite useful. I use it to summarise long podcasts, or youtube interviews. https://github.com/danielmiessler/fabric
I've been thinking about LLMs that don't try to answer every question for a while. Doing things like "assess the target audience" or "explain the scientific/historical concepts that are background knowledge to this text" that aren't intended to be general-purpose systems.
As far as color-coding: what I want is for the LLM to use these tokens in its own output to give a more structured response. It can do a decent job with directions (such as (<green> for technical explanations) or (<yellow> for direct quotes)), but a more standardized system that is used in the training data would do better.
I love the plain language idea! It isn't glamorous, but I am sure it would be a killer if anyone could make it work.
It's actually surprisingly difficult to translate into plain language. LLMs typically learn by mimicking humans, and in general, humans aren't so good at communicating complex things in a way that is easy to understand.
The use-case I'm excited about is finally scaling effective psychological growth methods. (Most therapy is bad, and sooo much more is possible.) https://chrislakin.blog/p/alignment-plan
I think I could do about half of what I do to help others with AI now. Still waiting for voice-to-voice models though.