Compared to LSM-1 (released at ICLR 2025): LSM-2 is: More accurate Less sensitive to missing data Still improving as scale increases LSM-1 is already plateauing. LSM-2 is still climbing.
MACHINE LEARNING
-
LSM-2 Machine Learning Model Demonstrates Robust Performance on Healthcare Benchmarks
By
–
LSM-2 crushes benchmarks: 20-class activity recognition Hypertension & anxiety detection BMI + age regression Robust when sensors fail Lower error when whole sensors or time windows go missing
-
LSM-2 model trained on 40 million hours of health data
By
–
LSM-2 was trained on: 40M hours of data
60,000 participants
Fitbit + Pixel Watch data
Collected over 3 months in 2024 Completely anonymized.
All consented.
Tracked activity, heart rate, and self-reported health. -
Technical Explanation of AIM Training Methodology
By
–
Here’s how AIM works: – Inherited Masking = natural missing data
– Adaptive Masking = synthetic gaps added for training By combining both: Better robustness Lower error No fake imputation Way faster training (dropout + attention mask combo) -
AIM technique for training models on incomplete data
By
–
Instead of treating missing data like a bug… AIM treats it like a feature. Because in real life, data is always missing. AIM trains the model to: • reconstruct gaps
• adapt to any fragmentation
• and learn richer representations -
Google Researchers Introduce Adaptive and Inherited Masking for AI Models
By
–
Researchers at Google fixed it. They built LSM-2, a model that learns from the gaps themselves using a new method: → Adaptive and Inherited Masking (AIM) AIM doesn't fill in missing data. It learns with it.
-
The Data Quality Challenge in Wearable AI Models
By
–
The problem? Every existing AI model assumes clean, complete data. But in 1.6 million full-day recordings, 0% were missing-free. This is a giant blindspot in wearable AI.
-
Google Releases LSM-2 Foundation Model for Wearable Health AI
By
–
BREAKING: Google just dropped LSM-2 a new foundation model for wearable health AI. It learns directly from broken, gappy, real-world sensor data. No imputation. No filtering. Just pure signal from the noise. Here’s how it works ↓
-
New AI Achieves Reasoning Without Pretraining on Just Two GPUs
By
–
Cette nouvelle IA fait du raisonnement sans pré-entraînement, et elle tourne sur juste2 GPU). Si c'est vrai, c’est un truc vraiment énorme. 40,3 % sur ARC-AGI. Elle tourne sur un grille-pain. Comment c’est possible ? Je vais me renseigner pour vous.
-
Meta Poaches Three Google DeepMind Researchers Behind IMO Gold AI
By
–
Meta vient (encore) de débaucher 3 chercheurs de google deepmind qui ont conçu le modèle médaillé d’or aux olympiades de maths (imo) d'il y a quelques jours. Le Zuckerberg est inarrêtable Il va finir par recruter toute la terre entière dans son labo