What GPTs really can do
November 27, 2023
The new chatbots everyone can create are promising, allowing for custom GPTs tailored to specific use cases. However, they are still in their early stages. We explore current capabilities and areas for improvement.
OpenAI's inaugural developer conference brought a flurry of announcements, with the most exciting being the introduction of new GPTs. These are specialized chatbots that anyone can create based on the GPT-4 Turbo language model and utilize within ChatGPT.
GPTs serve various purposes, enhancing negotiation skills, improving language style, or providing IT support. Most newly developed GPTs are prompt libraries or slightly modified versions of ChatGPT, equipped with features like web browsing, DALL-E, and data analytics. A step-by-step guide on creating your own GPT is provided here.
The data dilemma
GPTs can be enriched with additional information uploaded in the form of files. However, caution is advised: attempts have shown that these files can be downloaded from the GPT. While a simple prompt might not be sufficient, with a bit of experience in extracting information from the chatbot, even this hurdle can be overcome.
This poses a dilemma because it's the additional information that makes a custom GPT special. Feeding it only with publicly available material often doesn't yield significant improvements compared to the original. We tested an SEO GPT, which didn't provide better answers than ChatGPT; it was just slower, a disadvantage current custom chatbots face.
Valuable data is better kept out of personal GPTs; without such data, the GPT is less valuable.
From library to application
In addition to prompts and libraries, custom GPTs also incorporate actions. These are descriptions of how external interfaces should be utilized. Through these actions, any functionality can be integrated into GPTs. An example is Zapier's Calendar GPT, which accesses calendars via an interface and manages communication in case of scheduling conflicts. This demonstrates the potential for GPTs to be truly productive.
Snapshot of GPTs
Interactive development of personal GPTs is possible but offers advantages over the original only when the use case is carefully chosen and they are well-prompted. Valuable information doesn't belong in these GPTs yet due to the risk of user access. Productivity leaps are expected when GPTs communicate with other applications that complement the language model's capabilities.