Are you ready for generative AI?

ChatGPT is built on what is known as a large language model, part of a broader family of technologies called foundation models. These systems are trained on gigantic volumes and varieties of data, but they are not specialized for any specific task. Rather, they are pregnant, if you will, with a huge set of potential capabilities. They can perform at near-human levels when given just a tiny amount of data related to a specific task.

In 2020, OpenAI released GPT-3 (Generative Pretrained Transformer 3), the underlying algorithm for ChatGPT. That version of the model encompassed 175 billion parameters and was trained from a multitude of data sources from across the internet.

The GPT-3 release was a breakthrough moment that changed how the AI community thought about what large language models could do. Most importantly, it reshaped the notion of generality in AI. After the system was trained, you could then adapt it for a specific task by giving it a few examples of text and then giving it simple instructions in natural language. That core ability to instruct a chatbot the same way you would a fellow human is what’s revolutionary about GPT-3.

The model’s emergent capabilities make it great at parlor tricks—chocolate chip cookie recipes written in the style of the Declaration of Independence and other feats of creativity and inventiveness, for example. It’s also very useful for retrieving information, and software developers have found it to be quite skillful at suggesting code when prompted with a programming challenge.

GPT-3 also has significant limitations. One is linked to “hallucinations” in which the system provides made-up answers to questions. Hallucinations arise due to the static nature of the model’s knowledge, which is whatever the engineers burned into the model on a specific date. Without the ability to update itself or access the internet, its knowledge is limited, and when it doesn’t know the answer to a question, it may make one up.

These limitations are not permanent. For now, ChatGPT can’t connect to real-time or updated corpora of knowledge or application programming interfaces (APIs). However, Microsoft’s more advanced version can search the internet and include very recent information and news. Such breakthroughs will unlock even more of this technology’s potential.

It’s easy to imagine everyone in an organization soon having access to cognitive assistants powered by large language model-based AI.

Sitting at your computer, you could ask your AI assistant questions throughout the day about the tasks you’re working on and have conversations with it. You could talk to it as you would a trusted colleague and get help testing hypotheses, trying things out, gathering information from a knowledge base, organizing your day, and so on. Such functionality would help you do your job better and faster.

There’s also a lot of potential in functionality that resembles GitHub Copilot, a generative AI tool that offers on-the-fly coding suggestions for programmers. For everyday enterprise system users, we can imagine every text-entry box automated to suggest likely inputs based on the context and what you’ve done before. (Of course, the human user always has the final word on what the system should do.) We can imagine Copilot-like technology deployed at scale everywhere.

A bit further out, it’s only a small-speculative jump to having a digital twin for various work personas, such as IT help desk agents. In such a future, newly hired agents would get up to speed quickly because they would benefit from the autocompletion and automatic interface of their digital twin, which is the product of the AI system learning from hundreds of workers that came before them. Anything that a user produces can serve as feedback to further train these models through reinforcement learning, so the feedback itself contributes to making the models better. Soon the digital twin might be as capable as top human performers in a variety of everyday situations.

Looking at products, we can imagine companies providing highly trained foundation models for specific industries or verticals. A foundation model tailored for telecom, manufacturing, or the public sector would understand the specific vocabulary and processes unique to those sectors. These models would be fine-tuned on specialized data sets that allow them to understand context-specific nuances. Companies would then provide them as part of a platform capability or service.