An on-device LLM can generate news watch reports with a high degree of accuracy when combined with code validation.

🛠️ Can a small on-device LLM write a news watch report you can trust? Not alone: we fence it in with code. fornews.ai 1.7.3 (in testing): your iPhone asks, a silent iCloud push wakes your Mac, Apple's Foundation Models write the report on-device, and the code checks every sentence: dates, allegations, sources. No server, no cloud model, nothing leaves your devices. 7 real reports measured, every sentence against its sources: correct ones 53% → 70%. Misreadings still slip through, and we say so. How it works: https://ok-ia.ch/fornews/rapports-de-veille.html?lang=en App Store: https://apps.apple.com/app/id6783329977 #fornews #AppleIntelligence #FoundationModels #OnDeviceAI #iOSDev #SwiftUI #LLM #Privacy

Today's rabbit hole 🐇 Teaching a small on-device LLM to write news watch reports I can actually trust. The trick isn't a bigger model. It's a stricter one: Apple's Foundation Models write the prose on my Mac, and plain old code checks every single sentence (dates, allegations, which article really says what). My iPhone asks, a silent iCloud push wakes the Mac, the report comes back. No server, no cloud model. 7 real reports, every sentence checked against its sources: 53% → 70% correct. Not perfect, and the write-up says exactly where it still fails. https://ok-ia.ch/fornews/rapports-de-veille.html?lang=en #fornews #AppleIntelligence #FoundationModels #OnDeviceAI #iOSDev #SwiftUI #BuildInPublic