Google Research Unveils Wearable-Focused AI "SensorFM"
Google Research has released SensorFM, a foundation model designed to comprehensively extract health insights from wearable device sensor data. The model was trained on more than a trillion minutes of data from roughly 5 million Fitbit and Pixel Watch users, and it outperformed existing benchmarks on 34 of 35 tasks spanning heart rate, sleep, exercise, and body temperature.
SensorFM's goal is to make raw sensor data — which varies from device to device — work as a unified "health intelligence layer." It could eventually serve as the foundation for Google's AI health coach features, though no product integration plans have been announced at this time.
Meta's Louisiana Data Center to Exceed $250 Billion in Total Investment
Meta has announced an additional investment in its massive data center campus in Louisiana. With a fresh $40 billion infusion, the project's total investment is expected to exceed $250 billion. The expansion is aimed at scaling AI computing infrastructure, with the facility's power capacity set to grow to at least 5 gigawatts.
Meta initially announced a $10 billion investment and has scaled up the project in stages since then. As Big Tech companies accelerate infrastructure spending to meet AI demand, this project stands out for its sheer scale. That said, its impact on local communities and the sustainability of its power supply will continue to warrant close attention.
TSMC Posts Record Quarterly Revenue on AI Demand
TSMC (Taiwan Semiconductor Manufacturing Company) posted its highest-ever quarterly revenue in Q2, driven by AI-related demand. Demand for AI chips is pulling the entire semiconductor supply chain forward, and TSMC's strong results directly reflect that tailwind.
With demand for high-performance AI chips showing no signs of cooling in the near term, the results back up TSMC's aggressive investment plans.
Indian Companies Turn to Chinese LLMs to Cut Costs
Amid mounting cost pressure, Indian companies are starting to move toward adopting Chinese-made LLMs. According to Nikkei Asia, Indian firms facing steep API fees for Western models have begun considering cheaper Chinese models as alternatives.
This signals that cost has become a critical factor in the global AI race. However, given data privacy and geopolitical concerns, adopting Chinese models also calls for careful deliberation.
OvisOCR2: A Local OCR Model Taking on Document Parsing with 0.8B Parameters
OvisOCR2, which has been drawing attention in the LocalLLaMA community, is an end-to-end OCR model built on Qwen3.5-0.8B. It can convert entire document pages into structured Markdown, preserving text, tables, formulas, and reading order in the output.
According to the model card, it scores 96.58 on OmniDocBench v1.6 and 75.06 on PureDocBench. The weights are released under the Apache 2.0 license, and vLLM support is also provided. Delivering practical accuracy at a compact 0.8B size, it looks like a compelling option for local document processing needs — though independent real-world validation is still limited.
---