This strikes me as old school distributed computing thinking. I say this as an old school distributed computing guy.
All you have to do is write a text based description on f what your GPT does.and put it a bunch more fobvious places. Then all the “DiscoverGPT” variants will be able to find it and people just ask them.
In fact, Tyler and you could have TylerDiscovetGPT and ASKDiscoverGPT which simulate your searching and evaluation of GPTs
"In fact, Tyler or someone else could create a Tyler Cowen app. Pick a large set of representative essays and blog posts by Tyler and use them to create TylerGPT. Anyone can then query TylerGPT to get his take on any issue, new or old."
Hasn't this approach been tried, with less than success as a result? I recall reading, from Freddie DeBoer, if I remember correctly, that someone did this with one writer, and ended up with output that directly contradicted the writer's positions on several issues.
There was also the example where a law firm used GPT to draft a filing in a court case, and the GPT cited cases that didn't exist.
I'm sure there are cases where GPT could give terrific results. But, if you're depending on accuracy, it looks like you still need a checker.
GPT might have worse accuracy but ANY source should be confirmed. It seems to me that using GPT for a first draft could speed up the process and also be a source of different ideas or approaches. That said, your point is highly relevant for tutorials where the user might not be expected or capable of checking other sources.
"create one for yourself to spend an hour learning about something."
While GPTs probably have great potential, I haven't seen much that interested me personally. Until now. Those quoted words prompted me to see a few exciting possibilities.
I'm a retired guy who just doesn't do much "work." I often use Google to spend an hour learning about something, including on topics that come up on In My Tribe. My immediate thought was GPT sounds great for that. Maybe so. On further reflection I don't think GPT will be much better for that than Google but it prompted another thought. I often have issues with phones, computers, TVs, car radio, etc. where I need to change a setting and don't know how/where or I just don't know how to make the device do what I want. Sometimes Google/YouTube or a manual help, sometimes not. Sometimes I can find what I want using those methods, sometimes not. Similarly, I still think I'd rather read most books but I often read a book and then want to go back and find something. Sometimes I have a text search option but even that doesn't always work. One book had a many related points spread throughout the book and a GPT to condense them into a list would be great. And if I find a book uninteresting but still think it might contain something worthwhile, an interactive tool seems like a good option.
"I think of GPTs as apps that anyone can develop."
Anyone can write a book. Even before the advent of the internet and its disruption of the publishing industry, anyone could write a pamphlet.
"I praised Tyler Cowen’s creation of an interactive book. Anyone will soon be able to do that."
Books are also "interactive". You can read whatever page or chapter you want, whenever you want. You don't depend on a lecturer being available at a particular time, or having a classroom available. You learn at your own pace and every one can read (... right?), so no need for esoteric skills.
I don't see that much demand for GPTs as learning tools in the near future. We've had books for five and a half centuries and still most people don't read them. Most people would prefer to watch a movie/video/TV series, so maybe GTPs will become a success as entertainment.
Reading books is already approaching esoteric skills territory, with the demands the activity places on attention span, and if many people start replacing reading with "interacting" with books premasticated for them by GPT, it will happen all the sooner. Once upon a time people used to fill out index cards for books they read - see Umberto Eco's _How to write a Thesis_ for a detailed description. My impression is that extremely few people do this anymore, as Google search and other full text search apps have replaced the most obvious and immediately useful function of index cards: remembering where you saw a quote or read about a specific subject. The cards have less obvious functions, though, namely to help your mind arrange and digest the material and build up a model of whatever problem space you are studying, and Google search does not help with that at all. I'm afraid GPT-books will only exacerbate this tendency.
ETA: or think about this in another way. GPT-book acts as a research assistant, who can read and summarize/explain books and papers for you. But whereas you can instruct your research assistant and expect them to follow simple rules like 'don't put anything in the summary which is not in the book/paper', one thing we have consistently observed with GPT-like 'AI'/digital creatures is that they are incapable of clearly distinguishing rules from other input. This is very obvious from the profusion of jailbreaking (the dialogs are often very amusing). Today's GPT-like 'AI'/digital creatures are very good at slinging language around, but they can't yet do proper logic, and rule-following is a sort of logic if you think about it.
"Reading books is already approaching esoteric skills territory, with the demands the activity places on attention span, and if many people start replacing reading with "interacting" with books premasticated for them by GPT, it will happen all the sooner."
Agreed, attention spans are decreasing.
What happens when/if non-fiction writers realize most consumption of their books is by GPTs or equivalents? Do they just publish the summary? Why bother with a whole book? Will GPT check all my references?
I imagine they will go straight to publishing a GPT profile or however it is called. And in many cases - I mean the kind of non-fiction book which digests some scientific papers for less technical readers - it may be no great loss.
ETA: also, to riff off Hanson's theme, much of the non-fiction book market is not about the book per se, it is about status games and the lecture circuit which is where the real money is. The book might be just a loss leader into that. So the real question is how potential GPT-ization of non-fiction affect the lecture circuit. I'm guessing not much because whereas GPT can probably do a lecture and Q&A, being nonhuman, it has no social status, so why would people go listen to it? You might as well have GPT throw a theme party. The point is who invites whom and who meets whom, and GPT is neither.
"In fact, Tyler or someone else could create a Tyler Cowen app. Pick a large set of representative essays and blog posts by Tyler and use them to create TylerGPT. Anyone can then query TylerGPT to get his take on any issue, new or old."
But since Tyler hasn't written on everything under the sun, the app will have to guess what "his take" is in many cases. This problem gets worse when you have to cut down the set of "everything he has written" to "representative essays and blog posts". Who decides what the criteria for "representative" is? Who applies the criteria? Should I trust him?
The problem with ChatGPT's GPTs is that the most interesting ones, which call APIs for other services, can't be used on ChatGPT's mobile app due to privacy-related policies implemented by Apple & Google. And since most people want to use mobile versions of services, that effectively cuts out 90% of potential users for a given GPT.
People use different devices for different tasks. Even if 90% of users for all computing interface tasks is on a phone, that doesn't mean it will be 90% for these types of GPT tasks.
I’m surprised TBL thinks discovery is a problem.
This strikes me as old school distributed computing thinking. I say this as an old school distributed computing guy.
All you have to do is write a text based description on f what your GPT does.and put it a bunch more fobvious places. Then all the “DiscoverGPT” variants will be able to find it and people just ask them.
In fact, Tyler and you could have TylerDiscovetGPT and ASKDiscoverGPT which simulate your searching and evaluation of GPTs
"In fact, Tyler or someone else could create a Tyler Cowen app. Pick a large set of representative essays and blog posts by Tyler and use them to create TylerGPT. Anyone can then query TylerGPT to get his take on any issue, new or old."
Hasn't this approach been tried, with less than success as a result? I recall reading, from Freddie DeBoer, if I remember correctly, that someone did this with one writer, and ended up with output that directly contradicted the writer's positions on several issues.
There was also the example where a law firm used GPT to draft a filing in a court case, and the GPT cited cases that didn't exist.
I'm sure there are cases where GPT could give terrific results. But, if you're depending on accuracy, it looks like you still need a checker.
GPT might have worse accuracy but ANY source should be confirmed. It seems to me that using GPT for a first draft could speed up the process and also be a source of different ideas or approaches. That said, your point is highly relevant for tutorials where the user might not be expected or capable of checking other sources.
Things are changing so fast -- that Sam Altman TBL writes about 9 Nov. is already NO LONGER the head of OpenAI, fired on Friday.
Hired as of Monday morning by ...
Microsoft. The biggest investor and part owner of OpenAI.
Very strange and interesting infotainment available now!
https://www.understandingai.org/p/firing-sam-altman-hasnt-worked-out
TBL explains it better here than snippets on X.
Elon Musk was an early investor in OpenAI - and remains cautious about AGI dangers.
I'm more with TBL that has less fear.
"create one for yourself to spend an hour learning about something."
While GPTs probably have great potential, I haven't seen much that interested me personally. Until now. Those quoted words prompted me to see a few exciting possibilities.
I'm a retired guy who just doesn't do much "work." I often use Google to spend an hour learning about something, including on topics that come up on In My Tribe. My immediate thought was GPT sounds great for that. Maybe so. On further reflection I don't think GPT will be much better for that than Google but it prompted another thought. I often have issues with phones, computers, TVs, car radio, etc. where I need to change a setting and don't know how/where or I just don't know how to make the device do what I want. Sometimes Google/YouTube or a manual help, sometimes not. Sometimes I can find what I want using those methods, sometimes not. Similarly, I still think I'd rather read most books but I often read a book and then want to go back and find something. Sometimes I have a text search option but even that doesn't always work. One book had a many related points spread throughout the book and a GPT to condense them into a list would be great. And if I find a book uninteresting but still think it might contain something worthwhile, an interactive tool seems like a good option.
"I think of GPTs as apps that anyone can develop."
Anyone can write a book. Even before the advent of the internet and its disruption of the publishing industry, anyone could write a pamphlet.
"I praised Tyler Cowen’s creation of an interactive book. Anyone will soon be able to do that."
Books are also "interactive". You can read whatever page or chapter you want, whenever you want. You don't depend on a lecturer being available at a particular time, or having a classroom available. You learn at your own pace and every one can read (... right?), so no need for esoteric skills.
I don't see that much demand for GPTs as learning tools in the near future. We've had books for five and a half centuries and still most people don't read them. Most people would prefer to watch a movie/video/TV series, so maybe GTPs will become a success as entertainment.
> every one can read (... right?)
Reading books is already approaching esoteric skills territory, with the demands the activity places on attention span, and if many people start replacing reading with "interacting" with books premasticated for them by GPT, it will happen all the sooner. Once upon a time people used to fill out index cards for books they read - see Umberto Eco's _How to write a Thesis_ for a detailed description. My impression is that extremely few people do this anymore, as Google search and other full text search apps have replaced the most obvious and immediately useful function of index cards: remembering where you saw a quote or read about a specific subject. The cards have less obvious functions, though, namely to help your mind arrange and digest the material and build up a model of whatever problem space you are studying, and Google search does not help with that at all. I'm afraid GPT-books will only exacerbate this tendency.
ETA: or think about this in another way. GPT-book acts as a research assistant, who can read and summarize/explain books and papers for you. But whereas you can instruct your research assistant and expect them to follow simple rules like 'don't put anything in the summary which is not in the book/paper', one thing we have consistently observed with GPT-like 'AI'/digital creatures is that they are incapable of clearly distinguishing rules from other input. This is very obvious from the profusion of jailbreaking (the dialogs are often very amusing). Today's GPT-like 'AI'/digital creatures are very good at slinging language around, but they can't yet do proper logic, and rule-following is a sort of logic if you think about it.
"Reading books is already approaching esoteric skills territory, with the demands the activity places on attention span, and if many people start replacing reading with "interacting" with books premasticated for them by GPT, it will happen all the sooner."
Agreed, attention spans are decreasing.
What happens when/if non-fiction writers realize most consumption of their books is by GPTs or equivalents? Do they just publish the summary? Why bother with a whole book? Will GPT check all my references?
I imagine they will go straight to publishing a GPT profile or however it is called. And in many cases - I mean the kind of non-fiction book which digests some scientific papers for less technical readers - it may be no great loss.
ETA: also, to riff off Hanson's theme, much of the non-fiction book market is not about the book per se, it is about status games and the lecture circuit which is where the real money is. The book might be just a loss leader into that. So the real question is how potential GPT-ization of non-fiction affect the lecture circuit. I'm guessing not much because whereas GPT can probably do a lecture and Q&A, being nonhuman, it has no social status, so why would people go listen to it? You might as well have GPT throw a theme party. The point is who invites whom and who meets whom, and GPT is neither.
"In fact, Tyler or someone else could create a Tyler Cowen app. Pick a large set of representative essays and blog posts by Tyler and use them to create TylerGPT. Anyone can then query TylerGPT to get his take on any issue, new or old."
But since Tyler hasn't written on everything under the sun, the app will have to guess what "his take" is in many cases. This problem gets worse when you have to cut down the set of "everything he has written" to "representative essays and blog posts". Who decides what the criteria for "representative" is? Who applies the criteria? Should I trust him?
The problem with ChatGPT's GPTs is that the most interesting ones, which call APIs for other services, can't be used on ChatGPT's mobile app due to privacy-related policies implemented by Apple & Google. And since most people want to use mobile versions of services, that effectively cuts out 90% of potential users for a given GPT.
People use different devices for different tasks. Even if 90% of users for all computing interface tasks is on a phone, that doesn't mean it will be 90% for these types of GPT tasks.