Will Gen AI Replace Machine Learning in Building Operations?

The latest hyped technology is often called simply “AI” as though AI is new, but AI has been used in buildings for years. Most current building AI is what's known as “machine learning” (or a derivative of it). Machine learning AI requires large training data sets to identify patterns in building operations. From these patterns, machine learning AI creates predictive algorithms (also called “models”) and uses these algorithms to analyze and predict building operations.

These machine learning models typically run on a computer or device inside the building. However, some machine learning models run in “the cloud” for analyses that are more cost-effective to run on bigger computers that use larger and more diverse data sets.

(Note that because of time lags and risks of outage, any cloud building AI should NOT be part of a local control loop’s logic.)

Another way to talk about machine learning is cause-and-effect learning. Essentially, machine-learning algorithms look at past decisions or cause-and-effect patterns and then seek to predictively replicate those same decisions, with the help of a human expert guiding and tweaking to improve the outcomes. Most of the AI you interact with on a daily basis is actually machine learning, such as when Amazon tells you “you might like these 3 other things” or when Netflix recommends a TV series based on your watch history.

New “generative AI” models such as ChatGPT build on the classic machine-learning algorithms, adding algorithms that enable them to find relevant concepts, images, or patterns from unlabeled, “unstructured” data. Then they use this analysis to produce content that’s statistically likely given the prompt it received. Another reason why gen AI is getting such buzz these days is that these models have focused on natural language processing, so that a human doesn’t have to understand a machine language to retrieve data or make requests.

To do any tasks with any real accuracy, gen AI models need to be trained on massive data sets. For example, ChatGPT was trained on billions of documents. An image generator might be trained on millions of images, and a code generator on billions of lines of code. The more training data and the more relevant that data is to your query, the more statistically likely the output. The downside is that more data and more analysis also require more power to run the program relative to simpler models.

Another important distinction between machine learning and gen AI is their training environments. Unlike classic machine learning, gen AI training is largely unsupervised, meaning that the model determines the relationships in the data without human direction. This is both a strength and a weakness, depending on how the tool is used. As the saying goes, a sharp knife cuts both ways.

Another surprising weakness of gen AI is that despite being “more evolved,” it isn’t always more effective than classic machine learning for many operational tasks. “For tasks that involve making predictions on structured data, like the tabular data in a spreadsheet, generative AI models tend to be outperformed by traditional machine-learning methods,” said Devavrat Shah, an EE and CS professor at MIT.

As in any situation ever, what matters is that you use the right tool for the job. When it comes to evaluating goodness of fit with gen AI, enthusiasm for shiny new tech could lead to costly quagmires and failures.