AI in Practice: From Cloud Privacy Risks to On-Device Solutions

Pocock says he was initially skeptical after seeing Web3 and other tech trends fade. His business partner encouraged him to explore AI, and he saw potential in tasks like turning unstructured data into structured information. He adds that developers can get started with AI using TypeScript, without first learning Python.

Asked what advice he has for people who expect AI to do everything, Pocock recommends getting good with the tools available now instead of predicting the future. He says his focus is teaching rather than punditry.

The host raises the privacy risks of sending private or client data to cloud services. Rybczak describes React Native AI, an open-source library for running models on iOS and Android, with privacy, offline use, and ease of use as motivations. He says it supports custom third-party models and mentions work on Apple’s iOS 26 LLM integration, with Gemini Nano and Gemini Edge planned.

Rybczak says simple tasks can use smaller models; rewriting notes as bullet points does not require a top cloud model. Pocock suggests starting with a capable model to establish that a task works, then optimizing for smaller, cheaper, faster models. They discuss automatic model selection as a way to route tasks to an appropriate model.

Pocock says he mainly relies on Copilot completion and rarely uses agentic sidebar features. He describes writing code while anticipating the next suggestion and says he would like completion to be smarter or easier to trigger.