ChatGPT Turns One

OpenAI made a splash this week at its first developer-focused event, announcing a wide range of new capabilities, features, and services. For a first birthday party, it was quite impressive but also awkward. OpenAI’s CEO Sam Altman’s presentation had the excitement and clumsiness of the parent of a one year old—celebrating its advances with the lack of surety of a new parent. In some ways, everyone in the OpenAI universe has something to celebrate. But everyone also has something to be wary about and Sam missed the mark explaining some logical inconsistencies in OpenAI’s product announcements and plans. 

Let’s walk through four of OpenAI’s announcements that I’ve been chewing on this week.

Smaller is better. And easier to evade regulators.

OpenAI announced an updated version of its flagship foundation model that it claims is faster, smarter, and cheaper. The new model, named GPT-4 Turbo (continuing OpenAI’s pattern of terrible branding), is now 2x-3x cheaper for developers, which should help with the consistent problem that deploying an application using GPT is simply too expensive. We haven’t seen any benchmarks that prove OpenAI’s claim that the model is smarter, so we’ll have to wait for external researchers to evaluate. If they have created a model that is both smaller and “smarter,” this could prove an interesting turning point in how we might think about Large Language Models potential, cost-to-serve, and regulatory requirements.

Just a week prior, the Biden Administration issued a wide-ranging executive order on AI that, in part, created a threshold for regulation based on the size of the model or the size of the compute cluster required to run the model. How much has OpenAI’s support for this move to regulation been rooted in its pursuit of making its models smaller and, therefore, farther away from being regulated? 

Cannibal ChatGPT. Eating anything useful. 

Over its first year in existence, ChatGPT has added new capabilities, including browsing the web, accessing math and data models, and creating images. To date, users had to select these additional capabilities through a dropdown menu or plug-in picker—clunky experiences, to say the least. 

OpenAI has reduced this clunkiness by making ChatGPT responsible for calling these capabilities itself. So, if you’re interested in an image, you can just ask ChatGPT to create an image from within a conversation. This is a great advance, albeit one that was obvious and shouldn’t have taken so long. The change also has some potential warning signs underneath.

OpenAI’s previous strategy was one that embraced a plug-in developer community. The company lauded plug-ins for math, web searching, reading pdfs, etc. Today, however, ChatGPT includes those functions natively—destroying the market opportunity for developer competitors. This advance shouldn’t be surprising: OpenAI’s stated goal is to create artificial general intelligence, not AI that is generally intelligent if you add up all the plug-ins available. What was surprising, however, was announcing the change to an audience of developers. Especially when you consider what followed…

GPTease. What are they really for? 

Perhaps the most interesting and exciting announcement was the ability for anyone to create a custom version of ChatGPT that is trained on user data. Without any need for coding, users can upload information and create an AI assistant that uses that information as its primary knowledge source (underpinned by the broad knowledge in GPT-4 Turbo). Called GPTs (come on, OpenAI, please hire someone who can create better names), these custom AI assistants can be very useful. For instance, Helen created a Decision Assistant that is based on our book Make Better Decisions. There are countless opportunities for custom GPTs for individuals, but I don’t understand OpenAI’s pitch to developers. 

OpenAI announced that they will create a GPT Store which is being compared to the Apple and Google app stores. They say that they will share revenue with the most popular GPT developers, but didn’t provide any details: How much revenue will they share? How will they qualify GPTs to be in the store? Will they only pay for the “most popular” GPTs or for all of them? 

Given the fact that OpenAI just crushed the plug-in opportunity, what evidence is there that GPTs are a long-term strategy? During his keynote, Sam Altman said that they eliminated the dropdown function to call different functions because it was really annoying. Why will calling GPTs be different? Given how easy it is to create GPTs, shouldn’t we assume that there will be thousands or millions of GPTs to choose from? Wouldn’t it be better to have an AI select for us? And, if so, what’s the purpose of a GPT Store? 

With my skeptic’s hat on, I wonder if this GPT strategy is simply a way for OpenAI to open source product design and data gathering. What better way to figure out what its most active users want ChatGPT to do than ask them to create their own custom GPT and then cannibalize? What better way to find out what data is most important than to ask users to upload it? OpenAI says that it won’t use data from enterprise users and API users, but it isn’t clear about the rest of us and some have already found that you can “pull” the data used to train GPTs. 

Finally, what novel benefit can a GPT have? Our Decision Assistant was trained with our book. You, too, can create a GPT using our book (please buy it first, though!). So what will make our Decision GPT better than yours if the training data is the same? Nothing. And there’s nothing to stop anyone from ripping off anyone’s content to create GPTs, especially given that OpenAI will pay your legal fees.

Copyright Shield. Copywrong.

Following Microsoft, Adobe, and others, OpenAI will now pay any costs due to copyright infringement that users incur through ChatGPT Enterprise and the developer platform. It’s unclear if/how this will apply to GPTs–will GPTs be shielded as well? Given that the developer platform includes GPT creation, it feels to me that the announcement implies GPTs are included. 

That means that if we created an MBSGPT using the books of our friend Michael Bungay Stanier, OpenAI would pay any fees we incurred after Michael sued us (don’t worry, MBS, we wouldn’t do that to you!). Perhaps the details of Copyright Shield would limit OpenAI’s true exposure, but the intent is clear: OpenAI will defend you if you steal other people’s content. 

As content creators, this is very worrisome. It’s one thing for AI companies to hoover up content with the view that what is accessible is available. But it’s an entirely different thing for AI companies to use their financial prowess to provide air cover for anyone to abuse the copyrights of anyone. What incentive will remain for content creators to make innovative things if anyone in the world can steal those things and be defended by the AI Goliaths? 

Last thing. Despite all this, a cheer.

The OpenAI cheerleaders are collectively breathless after a week of excitement. And for good reason: there is a lot to cheer about, despite my cautions. ChatGPT, and generative AI broadly, is incredibly useful. And it can be quite a lot of fun. As an example, I’ll leave you with a cheer that ChatGPT wrote for itself:

📣 "Hey, hey, let's elevate,

OpenAI, you're truly great!

GPT-4 Turbo, fast and smart,

Leading AI's cutting-edge art!

🌟 Custom GPTs for us all,

Answering every single call.

Create, innovate, no delay,

OpenAI paves the way!

🚀 Multiple models, seamless ties,

Under the digital skies.

Faster, smarter, cost-effective,

OpenAI's perspective!

👏🏽 Let's cheer, let's celebrate,

For AI that's truly great.

OpenAI, hear our call,

You empower, inspire us all!" 🎉

P.S. Maybe my first GPT will be a cheerleader…