OpenAI’s developer conference brought a range of product announcements, including an improved GPT-4, new text-to-speech models and an API for DALL-E 3. The announcement with the widest implications, however, was GPTs: custom conversational AI systems that people can build with OpenAI’s models and share through a company-hosted marketplace.
Custom assistants move into a marketplace
GPTs can combine instructions, additional knowledge and actions to make an assistant more useful for a particular task. OpenAI said people would be able to publish what they build in the GPT Store. CEO Sam Altman also said developers would eventually be able to earn money based on how many people use their GPTs.
The idea is accessible to people without coding experience, while leaving room for more involved applications. A cookbook collection, for instance, could provide the knowledge for an assistant that answers questions about recipe ingredients. A company’s internal code could help another GPT check coding style or generate code aligned with existing practices.
That flexibility makes the offering more than a set of preset chatbot features. It gives users a route to package specific knowledge and instructions into tools that others can use, while keeping those tools within OpenAI’s model ecosystem.
OpenAI’s role expands beyond models
OpenAI had already taken steps toward connecting outside developers to ChatGPT through plug-ins launched in March. GPTs extend that direction by adding a more direct way to build, distribute and potentially monetize AI applications.
The shift matters because creating an AI product can involve more than choosing a model. Developers may need to connect a provider’s APIs with existing apps and services. A platform with tools for building and publishing can reduce some of that work for people who want an application based on OpenAI’s models.
That convenience could make model providers without comparable app-building tools less appealing to customers who have developer talent and want to create their own assistants. It may also put pressure on consultancies whose work centers on building custom GPT-like systems for clients.
Opportunity and market power arrive together
Making AI app creation easier can bring more people into the process. Users can tailor assistants to a collection of material or a specific workflow without starting from a blank coding project. A marketplace also gives those tools a place to reach other users.
But the same arrangement could strengthen the company that controls the models, creation tools and store. The source article raises the possibility of monopoly concerns, particularly because OpenAI has a first-mover advantage and is already building out the surrounding platform.
For competitors without the backing of a large technology company, the launch creates a strategic challenge: they may need to respond with their own tools or risk losing developers and customers who value an integrated way to build AI apps. The long-term effect will depend on how useful the tools prove to be and how the market develops.
AI tools still have uneven results
Other research and product developments in the roundup show how varied AI’s capabilities remain. Google’s MetNet-3 weather model combines variables such as precipitation, temperature, wind and cloud cover to make detailed predictions. Researchers at the University of Kansas, meanwhile, built a detector focused on AI-written introductions to chemistry journal articles.
That detector was trained using articles from an American Chemical Society journal and could identify introductions written by ChatGPT-3.5 with near-perfect accuracy. Its narrow scope is central to the finding: the researchers said a specialized tool could be built for particular sciences, journals and languages, even as broad text detection remains difficult.
Research into student essays explored a different use of machine learning: identifying passages that signal interests and qualities such as leadership or “prosocial purpose.” The researchers suggested such analysis could help admissions staff organize large volumes of applications. The report also notes a connection between students’ language and some academic factors, including graduation rate.
Experiences with accessibility tools were mixed. Researchers at the University of Washington found that summarizing systems could introduce bias or invent details, making them unsuitable for people who could not consult the original material. Yet one autistic user found that a language model helped draft Slack messages and made workplace communication feel more manageable, despite colleagues describing the messages as “robotic.”
Together, these examples point to a broader tension in AI adoption. Tools can make specialized tasks easier or help people communicate, but their limits and potential effects need attention. GPTs widen access to building applications; whether that leads to a healthier ecosystem will depend in part on how their usefulness, risks and market power are handled.