This morning, OpenAI launched the GPT Store: a simple way to browse and distribute customized versions of ChatGPT. GPTs – awkwardly named to solidify OpenAI’s claim to the trademark “GPT” – consist of a custom ChatGPT prompt, an icon, and optionally some reference data or hookups to external APIs. In the coming weeks, OpenAI will also start paying developers based on usage of their GPTs.
While GPTs may prove useful in their current form, they’re part of a grander plan. The exact path is yet to be seen, but I believe OpenAI tipped their hand with a very specific choice in how GPTs are designed.
Conversation starters
Here is a common model for building a custom chatbot on the GPT-4 APIs:
- A user navigates to your bot. (e.g, a feedback coach web app)
- The bot introduces itself, and asks you an initial question.
- The user responds to the question, and the conversation goes from there.
OpenAI GPTs do not support this model.
Specifically, unlike many API-powered bots, GPTs are not allowed to do Step 2: they cannot have the first word. A GPT author can provide a blurb describing the GPT, or provide suggested inputs to the user, but GPTs sit silently until you send them some kind of initial message. GPTs force a flow like this:
- A user navigates to your bot. (e.g. a feedback coach GPT)
- A user types something – anything – to the bot.
- The bot responds to the… whatever it was, and the conversation goes from there.
This was initially baffling to me as a developer. When we built the Feedback Wizard, a LLM-backed structured coaching experiment, it was clear that the Wizard should start the conversation by asking the user the first question. This is how a lot of software works: you start it up, and it asks you something.
- “What do you want to find?”
- “Where do you want to go?”
- “On a scale of one to ten, how would you rate your pain?”
- “Shall we play a game?”
This “computer starts with a question” model seems especially well-suited for building chatbots. For example, the whole point of the Feedback Wizard is to ask you questions. Attempting to build a GPT version of it under the constraint that the GPT can’t start the conversation was annoying.
Meanwhile, GPTs are set up to offer “conversation starters” – snippets of text users can tap to get things rolling, which force developers to think about how the GPT might be queried. When putting together the Feedback Wizard GPT, I felt pushed into writing conversation starters like “Can you help me frame some feedback for a co-worker?” – entry-points that felt redundant, given that the user has already selected a feedback coaching tool. However, while doing this, it struck me why OpenAI designed GPTs to be written in this backward-feeling way:
GPTs have to wait for user input before saying anything, because that will make them useable as building blocks of an Everything Engine.
The Everything Engine
ChatGPT is powerful, but has many limitations. No company can build, by itself, an Everything Engine: one text box that you can type into for any problem. OpenAI experimented with plugins to expand ChatGPT’s capabilities, but plugins competed with one another in-context, and the business model was questionable. Worse, there was too much friction to finding and selecting the plugin you wanted to use.
While GPTs may seem like they have similar problems, they offer a clear evolutionary path to a low friction, scaleable, and potentially highly profitable user experience. In a world where many useful GPTs exist, and those GPTs are written to respond to user input rather than start conversations with people who select them, ChatGPT can incrementally become an Everything Engine simply by routing requests to the best GPT for the job.
Let’s say you ask ChatGPT, “Can you help me with this math problem?” it could offer to send your query to the Khan Academy GPT, built by learning experts. If you ask, “How are the markets today?” it could sub into Yahoo Finance’s stock market GPT, equipped with realtime market APIs. If you ask an Everything Engine “Where should I go on vacation?” it might leverage a travel expert’s lovingly crafted GPT for helping you consider your options.
Just joking, it’ll offer the Tripadvisor GPT. Or whoever the highest bidder was. We won’t see sponsored GPT results for a while, but an Everything Engine would be the most compelling ad opportunity since web search results pages. If it works, the incentive for OpenAI to accept payment for GPT promotion will be immense.
LLMs aren’t great at math, but Wolfram is.
Admittedly, OpenAI might retain enough of their not-for-profit DNA to resist the urge to monetize the Everything Engine like this. Or maybe Sam Altman will be fired again on the path to building it. But even if they don’t do it, Google will.
If I’m right – if a model and ecosystem like GPTs can be used to build a compelling Everything Engine – then an ad-supported Everything Engine is coming. I’m more than willing to pay $20/mo for ChatGPT Plus, but most people won’t. But they will visit a free website where they can type into a text field, and it will answer their questions with ads.
Right now, that website is Google for most people. But no matter how much work Google puts into their Quick Answers and Bard, no one company is going to be able answer everything on its own. Serving the long tail of queries in one place – creating an Everything Engine – will require an ecosystem of developers and content providers.
In the coming years, if we do start to see the emergence of the Everything Engine, well… we will live in interesting times.
Thanks to Adam Lisagor, Jenn Cooper, Chris Parrish, and Paul Kafasis for feedback on this article and the ideas behind it.