Temporal AI: Reasoning Through Time Across Continuous Data Streams | Paper Club

In this Nebius Science Paper Club session, Patrick Langer, co-founder of Aionic Labs, talks about time series language models and how they can work with data such as wearable signals, ECGs, and engine sensor readings. Long recordings are difficult to fit into a model’s context window, and compressing signals can hide important events. Patrick explores ways to process these recordings more efficiently, while discussing the remaining challenges of real-time processing and reliable explanations. The talk covers three projects. TimeNet brings time series datasets into a common format, OpenTSLM combines time series and text in one model, and TimeRLM teaches a model to reason over long recordings. Event page: https://nebius.com/events/webinar-nebius-science-paper-club-temporal-ai #TimeSeries #MachineLearning #Nebius 00:00 Introduction 02:28 How OpenTSLM started 06:56 Overview: TimeNet, OpenTSLM, TimeRLM 07:41 Why models lack a concept of time 14:00 TimeNet: a common data layer 17:53 Time series language models 20:18 Prior work: raw numbers and soft prompts 23:02 OpenTSLM: soft prompt and Flamingo 25:55 Training data and results 28:24 TimeRLM: recursive reasoning 30:43 What's next: a learned memory 33:24 Q&A