Why Observability May Be AI’s Next Frontier

· Source: The Data Exchange · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Advanced, extended

Summary

Ameet Talwalkar, a Carnegie Mellon ML professor and Datadog Chief Scientist, discusses the evolution of time series foundation models, including Datadog's Toto V1 and V2. Initially met with skepticism in 2023-2024, these specialized models achieved a "BERT moment" by 2025, outperforming classical baselines in zero-shot forecasting on improved benchmarks like GIFT-Eval. Datadog's Toto V1, with 100 million parameters, and V2, scaling up to 2.5 billion parameters, demonstrated superior predictive accuracy, even on general-purpose time series data, despite V2 being trained exclusively on observability data. Notably, zero-shot inference with these larger models proved more cost-effective than hyperparameter tuning traditional ARIMA models. The discussion extends to "world models" for observability, aiming to simulate distributed software systems using multimodal data for proactive detection and what-if analysis. Talwalkar also addresses the shrinking 3-12 month gap between open-weight and proprietary AI models and AI's transformative impact on computer science education and academic research.

Key takeaway

For MLOps Engineers and AI Scientists building monitoring systems, consider adopting time series foundation models like Datadog's Toto V2. These models offer superior zero-shot forecasting and can be more cost-efficient than traditional methods, especially when accounting for hyperparameter tuning. You should explore fine-tuning these models with your specific in-distribution data for specialized applications, and investigate multimodal approaches to integrate diverse observability data for proactive system health management and "world model" development.

Key insights

Time series foundation models now surpass classical methods, enabling advanced observability and system simulation.

Principles

Method

Datadog developed Toto V1 (100M parameters) and V2 (2.5B parameters) by pre-training on massive observability data, demonstrating zero-shot superiority over classical time series models, with V2 excluding public data for better results.

In practice

Topics

Best for: Machine Learning Engineer, Research Scientist, MLOps Engineer, AI Scientist, Director of AI/ML

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Data Exchange.