TimeGPT & Psychohistory

In his iconic Foundation series, Isaac Asimov introduces the concept known as “psychohistory”.

Psychohistory is a fictional science that combines history, sociology, and statistical mathematics to predict the future of large populations. It’s a tantalizing notion, the idea that with enough data and the right algorithms, we might foresee the course of human events. How long until we will have real psychohistory in our lives?

Recent advances in AI and machine learning have given rise to models that can forecast with impressive accuracy in specific domains. For instance, the unveiling of models like TimeGPT-1 was met with enthusiasm, given its potential to handle time series data. However, practitioners in time series forecasting remain skeptical. Traditional models still reign supreme in various data dimensionalities, from high-dimensional scenarios like credit card fraud modeling to low-dimensional ones where human intuition heavily guides the forecasting. So far the paper the team from Nixtla released is very high-level and doesn’t share the concrete model architecture. It’s yet to be seen, how this model will do in the real world.

Transformers, which have revolutionized language processing, are yet to show their prowess in the world of time series forecasting. They have shown their power in various fields such as processing language and vision, and extracting intricate intermediate representations. But when it comes to the nuances of time, they often fall short. Time series forecasting requires the ability to discern patterns across diverse timescales simultaneously, a challenge even for advanced models.

So, are we on the brink of realizing Asimov's vision of psychohistory? The trajectory of AI advancement is both astonishing and relentless. Every leap in model sophistication, every drop in hardware costs, brings us closer to a world where "GPU poor" is a relic term of the past. The dream of foretelling vast expanses of the future, with all its human intricacies, might seem like a far-distant idea today. But with the pace we're moving at, don't be too surprised if we're soon seeing the first glimpses of this.