OpenAI unveiled its new multimodal AI language model, GPT-4, on March 14, 2023. This model is much more capable than its predecessors, scoring much higher on various exams and also accepting image input. GPT-4 has almost a thousand times as many parameters (a unit of model size in machine learning) as GPT-3, which many of us are familiar with, but one of its main appeals isn’t raw strength: the model can be tuned by users to perform a much wider variety of tasks than a one-size-fits-all model like GPT-3 can.
GPT-4 was also implemented into ChatGPT Plus, a subscription service for an improved ChatGPT. This version of ChatGPT is especially notable for its support for plugins. In an OpenAI demonstration of this feature, a user asks for a vegan restaurant recommendation as well as a vegan recipe with its calorie count. ChatGPT replies with all of the requested information, with a vegan restaurant courtesy of the OpenTable plugin, a chickpea salad recipe made by GPT-4, a calorie computation from a Wolfram|Alpha plugin, and a link to order the ingredients with the Instacart plugin. What’s so impressive about this integration, compared to earlier language models, is that GPT-4 figured out which plugins to consult and how to use them to fulfill the user’s request. Another incredible part of this story is that the Wolfram|Alpha plugin was fully realized just two and a half months after Stephen Wolfram originally pitched the idea of plugins to OpenAI.
One of GPT-3’s main limitations was that its training corpus consisted of only data from the year 2021 and before. OpenAI addressed this by creating a plugin of its own that allows ChatGPT to search the Internet for answers to certain questions. The model comes up with its own search terms and can interpret websites with complex formatting. This has the potential to allow a single site to greatly influence ChatGPT’s responses, which is not ideal, but there are ways to work around this problem.
One priority for GPT-3 and GPT-4 is the ability to code. While GPT-3 possesses a proof-of-concept level of coding skill, GPT-4 can write hundreds of lines in many programming languages, making nuanced programs that work out of the box. Critically, while past models are mainly useful as time-savers for coding experts who know exactly what they are looking for, GPT-4 can handle imprecision in requests. ChatGPT Plus can even write plugins for itself. Youtube user Candlesan recently used only GPT-4 code to write a Flappy Bird clone, although he notes that only those with coding knowledge would currently be able to do this.
An independent developer who goes by the username “Significant Gravitas” created a tool called “AutoGPT,” which allows a local version of GPT-4 to recursively prompt itself and execute Python scripts on a machine. This has enormous potential. AutoGPT can test and improve code it writes, as well as create files and folders to build much larger projects than GPT-4 can on its own. This behavior is only made possible by AutoGPT’s self-prompting system, which allows GPT-4 to break large tasks into smaller ones; this allows the model to “remember” more things at once and think on a larger scale. To quote one AutoGPT user, “AutoGPT was trying to create an app for me, recognized I don’t have Node [a JavaScript runtime], googled how to install Node, found a stack overflow article with a link, downloaded it, extracted it, and then spawned the server for me. My contribution? I watched.” When another user asked for help starting an E-Commerce business, their version of AutoGPT browsed the internet for ideas and saved them to text files for later use.
GPT-4 already satisfies many standard definitions of intelligence, and the latest developments in language models are the closest humanity has gotten to a true artificial general intelligence. This article, however, was not written by an AI.