From Sensing to Understanding: LoRaWAN + LLMs
summary
Claims concrete enough to act on: large LLMs help label and synthesize training data so a traditional tiny model trains faster, cutting manual labeling by 98% faster labeling in examples; cheap on‑device models handle normal cases and a model cascade sends only ambiguous cases to the cloud, saving battery and latency; prototypes run surprisingly small, sometimes runs in 100 kilobytes of RAM or on microcontrollers, while naive cloud inference can cost e.g. $170 per day at 1 fps for continuous video.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:51LLMs power and zero-shot
- 4:55LLMs in IoT challenges
- 6:39Small language model limits
- 7:41LLMs for data labeling
- 9:24Synthetic data with LLMs
- 11:09Model cascades for inference
- 12:53Local and cloud model blend
- 13:57Demo of LLM cascades
- 14:38Edge Impulse LLM support
related talks
14mThe Synergy of LoRaWAN and Artificial Intelligence
23mLLMs on the Edge: The Future of On-Device Intelligence
47mOwn Your LLM: Running Generative and Agentic AI Locally
16mExploring the world of Small Language Models
31mLeveraging LLMs to build supervised datasets suitable for smaller models
40mSurviving the Agentic Hype with Small Language Models
30mDeep Learning in Energy Management: Non-Intrusive Load Monitoring for IoT Devices
19mRL for Autonomous Coding