Been a reader of Charlie's since last year so it was surreal to see my startup, Consensus, featured as one of the GPTs in the article :)
Maybe I can provide readers some insight by briefly discussing how we are thinking about GPTs:
We think Charlie's analogies are correct. The GPT Store outcomes likely range from the Chrome Store to the App Store. In the former case, this is what we are already seeing. The Consensus GPT is provides novel value to users while also being a great acquisition channel to our main product, our web search engine: https://consensus.app/
But if things really take off, and ChatGPT becomes a primary workspace for billions of users, we must prepare for the GPT to be as important to our business as the Instagram app is to Meta. We are open to all possibilities!
Keep up the great work Charlie. Feel free to reach out to me anytime to talk all things AI and startups :)
It reflects much of my own thinking and predictions. GPTs for yourself? Great! GPTs as a source of income? Get real!
I think there'll always be room for a GPT Store (or its future iteration), but the "Store" moniker is misleading. I see it as more of a library or reference site.
A tiny nitpick would be this part: "the GPT creator tool helpfully translations instructions to prompt edits." As I wrote in my GPT post back in November, I don't think the built-in "GPT Builder" is up to scratch. It's super useful as a one-off wizard to walk new people through the important questions to ask about their GPT (tone, audience, etc.). But it doesn't appear to have deep knowledge of e.g. OpenAI's own "Prompt Engineering" guide and tends to come up with rather vague inteprpretations of your instructions even if you give it specific lines to use. Worse still, if you've already configured your GPT to your liking, using the Builder will simply overwrite all of your work instead of appending the changes.
Ah, that's totally fair. I forgot how much of a headache it was to stop the GPT Builder from overwriting the prompt. I had a separate notepad open to keep the parts of prompts I liked. Ideally OpenAI could add a history feature to see your past GPT prompts.
I made a GPT based on Paul Graham's essays (I uploaded a Markdown file of all of them) and would ask questions about Paul Graham's ideas. The GPT would only search the files about 20% of the time, and otherwise would repeat information about PG from memory.
I think a much better use case would be creating a GPT that *only* is meant to pull from files. In that situation, you would want to prompt the GPT that it should only get answers from the provided files and otherwise say "I don't know." I think leaving it up to GPT-4 to determine when to query the files is too much to ask right now - it should be all or nothing.
You go over how how GPT for teams is something attractive - I wonder if you can do a writeup in the near future though on the issue around the data privacy and companies wanting to guard that.
I think that unless OpenAI guarantees companies not to use customer data for training their AI, businesses will drag their feet from significant ChatGPT adoption outside of proof of concept and novelty use.
I see already with customers of ours they are begining to roll out internal training where employees are not allowed to use ChatGPT for any business purposes = this is a for a dual reason of not getting wrong data and not to expose business private data to OpenAI.
I admit, it is very possible that despite OpenAI's very open policy of "we will use your data to train our tools and then sell that back to you for $$$", a lot of people are drunk on the AI-Koolaide...
That's a great idea - it's worth clarifying 1) when OpenAI promises to not train on your data, and 2) what the downsides of training on your data might be (my guess is a large percentage of businesses probably don't need to worry too much about it).
For what it's worth, you can let your customers know that users the ChatGPT Teams plan (as opposed to the original ChatGPT Plus plan) won't have their data used for future model training - https://openai.com/chatgpt/pricing. It is more expensive, though!
I'd say the promise of GPT Store doesn't lie in the broad public monetization but in the ability to discover and share bespoke applications (future agents?).
People keep on comparing it to the App Store, but that comparison is misplaced. The App Store had a high barrier for entry as you had to become an iOS engineer. That's not the case with GPT Store, meaning there will be a proliferation of apps, and knockoffs will instantly appear right and left for every successful app.
Been a reader of Charlie's since last year so it was surreal to see my startup, Consensus, featured as one of the GPTs in the article :)
Maybe I can provide readers some insight by briefly discussing how we are thinking about GPTs:
We think Charlie's analogies are correct. The GPT Store outcomes likely range from the Chrome Store to the App Store. In the former case, this is what we are already seeing. The Consensus GPT is provides novel value to users while also being a great acquisition channel to our main product, our web search engine: https://consensus.app/
But if things really take off, and ChatGPT becomes a primary workspace for billions of users, we must prepare for the GPT to be as important to our business as the Instagram app is to Meta. We are open to all possibilities!
Keep up the great work Charlie. Feel free to reach out to me anytime to talk all things AI and startups :)
Very cool! Yeah, I'd love to chat and hear about Consensus behind the scenes.
Shoot me an email: christian@consensus.app
Great post, man!
It reflects much of my own thinking and predictions. GPTs for yourself? Great! GPTs as a source of income? Get real!
I think there'll always be room for a GPT Store (or its future iteration), but the "Store" moniker is misleading. I see it as more of a library or reference site.
A tiny nitpick would be this part: "the GPT creator tool helpfully translations instructions to prompt edits." As I wrote in my GPT post back in November, I don't think the built-in "GPT Builder" is up to scratch. It's super useful as a one-off wizard to walk new people through the important questions to ask about their GPT (tone, audience, etc.). But it doesn't appear to have deep knowledge of e.g. OpenAI's own "Prompt Engineering" guide and tends to come up with rather vague inteprpretations of your instructions even if you give it specific lines to use. Worse still, if you've already configured your GPT to your liking, using the Builder will simply overwrite all of your work instead of appending the changes.
Ah, that's totally fair. I forgot how much of a headache it was to stop the GPT Builder from overwriting the prompt. I had a separate notepad open to keep the parts of prompts I liked. Ideally OpenAI could add a history feature to see your past GPT prompts.
Agreed. "Revision history" and being able to revert changes or combine what works across several iterations would be very useful!
“In practice, I struggled to make this work well - I had difficulty getting the GPT to search the documents, even in cases I thought were obvious.”
What did you try?
I made a GPT based on Paul Graham's essays (I uploaded a Markdown file of all of them) and would ask questions about Paul Graham's ideas. The GPT would only search the files about 20% of the time, and otherwise would repeat information about PG from memory.
I think a much better use case would be creating a GPT that *only* is meant to pull from files. In that situation, you would want to prompt the GPT that it should only get answers from the provided files and otherwise say "I don't know." I think leaving it up to GPT-4 to determine when to query the files is too much to ask right now - it should be all or nothing.
Hi Charlie, great writeup.
You go over how how GPT for teams is something attractive - I wonder if you can do a writeup in the near future though on the issue around the data privacy and companies wanting to guard that.
I think that unless OpenAI guarantees companies not to use customer data for training their AI, businesses will drag their feet from significant ChatGPT adoption outside of proof of concept and novelty use.
I see already with customers of ours they are begining to roll out internal training where employees are not allowed to use ChatGPT for any business purposes = this is a for a dual reason of not getting wrong data and not to expose business private data to OpenAI.
I admit, it is very possible that despite OpenAI's very open policy of "we will use your data to train our tools and then sell that back to you for $$$", a lot of people are drunk on the AI-Koolaide...
That's a great idea - it's worth clarifying 1) when OpenAI promises to not train on your data, and 2) what the downsides of training on your data might be (my guess is a large percentage of businesses probably don't need to worry too much about it).
For what it's worth, you can let your customers know that users the ChatGPT Teams plan (as opposed to the original ChatGPT Plus plan) won't have their data used for future model training - https://openai.com/chatgpt/pricing. It is more expensive, though!
I'd say the promise of GPT Store doesn't lie in the broad public monetization but in the ability to discover and share bespoke applications (future agents?).
People keep on comparing it to the App Store, but that comparison is misplaced. The App Store had a high barrier for entry as you had to become an iOS engineer. That's not the case with GPT Store, meaning there will be a proliferation of apps, and knockoffs will instantly appear right and left for every successful app.
Thanks for the refreshing POV Charlie!