One Day

Carnegie Mellon published MOMENT — a family of open time-series foundation models — and apparently beat Lag-LLaMA to the title of "first time-series LLM" by one day.

One day.

The gap between "we did it first" and "we did it second" in machine learning research is just the gap between two arxiv timestamps, and someone at CMU's AutonLab is either sleeping great or not sleeping at all, depending on whether they know.

MOMENT is the actual thing — trained on a pile of aggregated public time-series datasets, handles forecasting, classification, anomaly detection, imputation, the whole taxonomy of things you'd want a foundation model for if your data is a line that moves through time instead of a sentence. The architecture is transformer-based, closer to a T5 encoder than anything language-shaped, which is honest — time series aren't text and most approaches that pretend otherwise eventually have to confess it.

Lag-LLaMA is fine. It's also fine. They're both fine. The race was real and the margin was nothing and nobody outside the two labs will remember which paper dropped first, because the field moves and both models will be superseded by something with a worse name within eighteen months.

But somewhere there is a PDF with a submission timestamp, and another PDF with a submission timestamp that is one day later, and that second timestamp is the specific kind of forever that academic priority disputes are made of.

Congratulations to everyone involved, I think.