Initial Thoughts On GPTs And Beyond
What new opportunities do GPTs present? Are AI wrappers dead and what next?
OpenAI pulled a master move holding back the GPT Store. Firstly, everyone is sharing their GPTs, giving OpenAI free exposure, and secondly, it still feels like a level playing field encouraging GPT creators to go wild.
Below, I created an illustration of what went through the mind of a founder since the launch of Dall E and every subsequent launch since, with the addition of GPT 4 Vision and GPTs.
It has been a week since OpenAI announced GPTs, and I’m still processing what this means. The immediate questions on my mind:
What new opportunities do GPTs present?
What about plugins?
What are the differences between GPTs and plugins?
Are AI wrappers dead?
How to think about use cases for GPTs?
What is the future of GPTs?
How to stay up to date with GPTs?
From OpenAI’s announcement page:
You can now create custom versions of ChatGPT that combine instructions, extra knowledge, and any combination of skills.
My first impressions of OpenAI’s GPTs
GPTs will make ChatGPT more practical for everyday users - Instead of iterating on a perfect prompt, you can now use a GPT crafted by someone else.
GPTs are convenient for streamlining repeat tasks - If you use ChatGPT for specific tasks requiring pre-prompting, creating a custom GPT can streamline your workflow.
GPTs will help users discover the power of AI that they would not have thought of on their own - GPTs will be a more fun and fast way to discover the abilities of ChatGPT. I’m assuming once the GPT Store is available, we will be able to find GPTs for all kinds of use cases we did not previously think possible.
GPTs will unlock new use cases - Now that it is possible to configure something useful in 10 to 30 minutes with little to no maintenance, it opens new use cases that would not have been worth building otherwise.
GPTs will be more exact and relevant than ChatGPT - By incentivizing GPT creators to craft GPT for particular use cases, OpenAI enables ChatGPT to access precise and relevant information that it otherwise would not have.
On the pessimistic side:
User experience challenge - I’m not convinced that novice users will be able to navigate the overload of GPT options.
Defensibility and replicas - I have noticed on Twitter that copying a GPT use case is straightforward if you are not relying on some unique external feature.
Initial hype - It is probably not a good idea to trust the hype as a signal for usefulness at this time.
What about plugins?
When plugins launched, initially, I was super excited about it. We built a plugin for LampBuilder, and it felt like magic seeing the AI do something by tapping into external actions.
However, as time passed, I noticed I did not use them as often. A major deterred could have been the clunky user experience:
Searching and navigating the plugin interface was not great.
Having to activate/deactivate a plugin before starting the chat was lame.
Not being clear on what info is sent to the plugin bothered me.
For plugins like Instacart and Canva, I went directly to the third-party sites. The only plugin I used for a while was the ability to browse online, which later got replaced with ChatGPT’s built-in browsing functionality.
I can’t help but wonder if the novelty of GPTs will wear off.
What are the differences between GPTs, plugins, and wrappers?
ChatGPT created an interface for crafting a GPT where you can specify the instructions, knowledge, and actions. This makes it possible to create solutions similar to some plugins and wrappers without the hassle of coding. The end user interacts with the GPT via the OpenAI UI, where the instructions, knowledge, and actions are handled.
You could also leverage the OpenAI UI by creating a plugin, but it was required to implement the instructions and actions on your server. Note that OpenAI paused new plugin launches in favor of GPTs.
On the other hand, wrappers only use OpenAI’s API and handle the UI, instructions, knowledge, and actions outside of OpenAI on a custom server.
Are AI wrappers dead?
I don’t think so, at least not yet.
For many GPT use cases, developing a custom feature or integration outside of OpenAI’s capabilities will still be required. For example, for my Lean Canvas AI tool, I create a Lean Canvas template in HTML that can be shared after completing it. It is not possible to make the same functionality within the GPT configuration exclusively.
Furthermore, the chat interface will not always produce the best results for the end user. Using the chat interface to enable and disable GPTs is also somewhat confusing.
How to think about use cases for GPTs?
Getting traction for your GPT will be very difficult due to the competitive nature, especially if you are not coding any custom actions or don’t have any special knowledge to leverage.
If you have access to unique insights into a domain or possess special knowledge that is not readily available, you are likely sitting on a promising GPT use case. Once you have found traction for a particular use case, you should consider adding custom features that differentiate your GPT from replicas.
Initially, you may consider quantity over quality to get a quick idea of what type of use cases suit custom GPTs and what your audience is interested in.
What is the future of GPTs?
I do not have inside information, making it even more fun to speculate on what comes next.
Automatically searching GPTs
At the moment, you can prompt ChatGPT directly. As you may have seen now and then, ChatGPT replies with two responses where you can choose the best one. Following this suggestive user experience, I think you’ll soon get custom GPT recommendations that may serve your desired tasks better.
For example, let’s say you want to create a travel itinerary for your visit to Cape Town. If someone has built a GPT for this, they may spend time curating the best up-to-date restaurants and sightseeing, making it better than just interacting with ChatGPT directly.
Orchestrating GPTs
AI Agents creators already started experimenting with this concept. As you may have read or seen, these agents tend to hallucinate and go into a spiral. What custom GPTs bring to the table is that humans will be incentivized to help the AI stay focused.
Once ChatGPT can tap into multiple GPTs to complete a task, it will become super powerful. For example, let’s say my goal is to create an itinerary and make bookings for my trip to Cape Town. A GPT orchestration layer will identify the best-fit GPTs that can generate an itinerary with up-to-date details, another GPT can make phone calls to make a reservation, and a third GPT can create a calendar entry.
More built-in actions
Currently, OpenAI has 3 built-in actions: image generation, online browsing, and code-interpreter. There is an opportunity to develop action-driven GPTs as a third-party service, but there is always the risk that OpenAI will do it in-house.
OpenAI will likely continue to partner with major organizations like Canva, Instacart, etc., for specialized integrations where these more prominent players have the resources to keep up with the load and carefully take care of privacy risks.
For actions where there is a lack of established players, OpenAI will likely create the best ones in-house, as we have seen with support for PDFs.
Access to more built-in actions like text-to-voice, sending emails, making phone calls, submitting online forms, processing payments, developing and hosting mini-apps, etc., will make it even more potent for GPT creators and the orchestration layer to maximize the AI’s potential.
How to stay up to date with GPTs?
I recommend checking out these five initiatives:
15,000 Custom GPT - Torbjørn created a GPT that searches GPTs.
GPT Builders - A community by Matt Schlicht.
Building unique GPTs - A Twitter thread by Calin Drimbau.
Community list of GPTs - A community Twitter list of GPTs by Kris Kashtanova.
Allgpts.co - A directory of GPTs.