Plugins – A Step Closer Towards AGI

In recent months the tech world has witnessed a significant leap in the capabilities of large language models (LLMs). What started some years ago with the integration of irritating customer support chatbots sometimes appearing on your car rental website and refusing to connect you to a human representative just while you're facing a flat tire on your way to the airport, rapidly transformed into a multi-billion dollar market with support staff capabilities sometimes surpassing those of its human counterparts.

I recently discovered, to my surprise, that ChatGPT has become a more adept travel planner than I am myself (and I'm a very good travel planner). See here how ChatGPT impeccably planned an unforgettable trip for my wife on the way to the mountains.

How did we get here?

Many argue that the key factors to OpenAI's success include vast amounts of training data, expanded context windows, and extensive fine-tuning with Reinforcement Learning from Human Feedback (RLHF). These elements are indeed crucial. GPT 4, along with its predecessors, has consistently outperformed the competition in almost all measurable metrics.

There may however be an even more important factor than these. OpenAI was the first company to introduce programmatic extensions of its models' capabilities via plugins. It is in my view that these extensions, combined with the aforementioned superior performance, could pave the way to true Artificial General Intelligence.

Do plugins make models more powerful?

Plugins allow developers to add code that the Large Language Model (LLM) can interact with. This is useful as there are some things LLMs aren’t designed to do. For example, while Language Models cannot ordinarily browse the web, they can interact with say, the Google API that preforms the search. Similarly, LLMs aren't particularly adept at calculating the distance between coordinate points. A simple plugin however enables the LLM to interact with the Google Maps API that can return the distance between two venues, sparing the LLM the task of preforming complicated calculations. The model is able to understand the purpose of the API, fill in the parameters apropriately according to the context, give the command to execute, interpret the results and relay the relevant information back to the user. In short, plugins let the Language Model interact with APIs that can do the heavy lifting.

My experience using ChatGPT for my wife's birthday planning showcased the power of plugins. ChatGPT was no longer just offering suggestions based on my wife’s interests, like eating Mexican food or attending a psychedelic rock concert. It could actually browse the web to find specific events matching these criteria. It could do this considering our location, timeframe, budget, and any additional constraints I could think of. And it could do it all based on a single, simple, unstructured conversation.

A step closer towards AGI?

I believe this capability will revolutionize our day to day interactions with technology. The development of over a thousand plugins in less than six months is impressive, yet their general popularity lags behind what one might expect for such transformative technology. Most people on the street won’t know what you’re talking about, if you ask them for their favorite plugin. This might be explained by plugins currently being confined to the ChatGPT interface. I for my part, don’t think it is likely that plugins will remain trapped here for long, available only for those that afford themselves the luxury of a Pro Subscirption. I think it is far more likely that we will start encountering more and more plugins, without even knowing it.

As OpenAI was quick to integrate the most popular plugins directly into its base application, GPT Assistants are sure to turn up soon directly in some of our favorite Apps and Websites, totally revolutionising the way we interact with them – and hopefully organising our tire replacement while avoiding a missed flight. All while on the other side an increasingly capable ChatGPT will continue to grow in its integrations, enabling it to plan and execute increasingly complex operations, moving ever closer to a true AGI.