What Everyone Else Isn’t Telling You: An Insider’s Guide OpenAIs Custom GPT Assistants…
Yesterday at OpenAI’s first developer conference (https://www.youtube.com/watch?v=U9mJuUkhUzk), among many other things, they announced custom “GPTs” AKA “Assistants”.
While most posts reporting on this specific feature are simply rehashing the broad announcement, none cover the details, and as such, I believe this is the first and only report on how to actually create and use Assistants!
Please share far and wide :-)
After a blackout on Plugins on the same day, Logan, Developer Relations for OpenAI, said yesterday: “Plugins in ChatGPT are sticking around for the time being.”
And that the blackout was caused by a bug.
See: https://twitter.com/OfficialLoganK/status/1721660689625018809 for details.
This caused rumours in some private ChatGPT Plugin builder groups that Plugins would be removed eventually — I think they will be removed and will be superseded by GPTs (Assistants) for the reasons I will explain:
I got early beta access to GPTs, so I decided to build.
What was immediately obvious was the “Configure” part of GPTs (Assistants) looks awfully similar to a Plugin’s Manifest file details.
If one looks further, there is an “Action” feature that replicates the OpenAPI YAML file component of Plugins also.
So based on that, I decided to duplicate my 15 deployed ChatGPT Plugins as GPTs (Assistants) too.
After doing this, it’s clear that GPTs (Assistants) will replicate Plugins, but only so far (and there is still a place for complex functionality *some* Plugins provide…
For some of my plugins, all I needed to do was give instructions in natural language with the equivalent details contained in my Plugin Manifest file.
For others, I had to replicate the OpenAPI YAML file of the Plugin.
What this means is building a Plugin/GPTs/Assistant will be quite easy for non-coders and ultimately will make Custom chatbots (within the OpenAI platform) easier to build for the layperson.
This will create a lot of supply for the proposed GPTs Marketplace to launch later this month.
So What can be done with custom GPTs?
A little, and not enough to replace the need for more complex Plugins.
Here is what can be made with GTPS/Assistants *without* a Plugin:
1) Static Retrieval-Augmented-Generation (RAG) via file upload
2) Webpage browsing
3) Image creating
4) Code interpreting
But for the more astute, all of this can be done within ChatGPT already without Custom GPTs!
They just cannot be shared with others.
The other nuance is that “Custom Instructions” can be added per GPT, not for the entire account as is the case now.
Other limitations:
Although there are share links, these bots need to be accessed via ChatGPT, not an external website.
However, the Assistants API will do this.
Unlike RAGs which can access real-time or proprietary data, GPTs/Assistants can only access files that are uploaded and static; the deployer must manually upload them.
For example, I have uploaded the entire Bible and Quran as custom GPTs.
Custom GPTs can use static RAGs like this far longer than the updated context window that was extended to 128k tokens!
Custom GPTs can only access a single website at a time and cannot access external APIs or databases, etc.
So proprietary data that sits in external APIs and databases, will still need to be accessed through Plugins, AKA the “Actions” functionality or via the Assistants API to be deployed on external websites.
Specifically creating an API YAML file.
Although OpenAI has made it easier to either import an existing YAML file or use their interface to create one.
And this still allows Assistants to import and call code to run external ads like before.
Both of these require coding knowledge, and only a few of the total custom GPTs that will be released in the future will have these.
Again, Plugins aren’t “dead” nor is it as doom-and-gloom as it seems.
To date, some Plugins were monetizing via ads, subscriptions through oAuth, or a CTA for free leads that could be monetized later.
But, according to OpenAI, “You’ll also be able to earn money based on how many people are using your GPT.”
See: https://openai.com/blog/introducing-gpts for details.
This will allow custom GPT deployers to monetize them without or in addition to these existing monetization channels.
