Wednesday 26 March 2025
The latest batch of foundation models, touted as the next big thing in machine learning, has been put through its paces on a particularly challenging test: predicting cloud computing traffic. The results are underwhelming.
Foundation models are neural networks that have been trained on vast amounts of data from various domains, allowing them to generalize to new and unseen tasks with remarkable ease. They’re like the Swiss Army knives of AI – capable of tackling everything from image recognition to language translation. But can they handle the complexities of cloud computing?
To find out, a team of researchers constructed four datasets from publicly available Huawei Cloud data, featuring function requests from five regions in China. The data was then fed into various foundation models, as well as two simple baselines: an online linear model and a naive seasonal forecaster.
The results are stark. While the baseline methods consistently delivered accurate predictions, the foundation models struggled to keep up. Moirai, one of the top-performing FMs, managed to produce forecasts that were often wildly inaccurate, failing to capture even the most basic trends in the data. VisionTS, another contender, did a bit better but still lagged behind its human-designed counterparts.
The researchers also found that these foundation models exhibited some curious behaviors, such as suddenly switching from producing reasonable predictions to generating seemingly random output. This sort of unpredictability is exactly what makes AI so fascinating – and frustrating – for scientists and engineers alike.
So why do these foundation models struggle so mightily on this particular task? One reason may be the sheer complexity of cloud computing data. Cloud traffic is a chaotic, ever-changing beast, with patterns that are difficult to discern let alone predict. It’s like trying to forecast the weather using a toy model designed for predicting tides – it just isn’t meant to work.
Another possibility is that these foundation models simply weren’t trained on enough data specific to cloud computing. They’re general-purpose AI tools, after all, and may not have had sufficient exposure to the unique patterns and trends found in this particular domain.
Whatever the reason, the takeaway from this experiment is clear: while foundation models are undoubtedly powerful tools, they still have their limits – even when it comes to predicting something as seemingly straightforward as cloud computing traffic. It’s a reminder that AI, like any other technology, requires careful consideration of its strengths and weaknesses before being deployed in real-world applications.
Cite this article: “Foundation Models Struggle with Cloud Computing Traffic Prediction”, The Science Archive, 2025.
Foundation Models, Cloud Computing, Machine Learning, Neural Networks, Data, Huawei Cloud, China, Predictions, Accuracy, Ai
