AI Daily — August 13, 2026
2026-08-13
deepmind.google
Putting sign language AI into users' hands
Google DeepMind is releasing SL2T (sign-language-to-text), a breakthrough model that converts sign language video into text, now powering new accessibility features for Deaf and hard-of-hearing users. The announcement marks a significant step in deploying multimodal perception models for a historically underserved use case. No architectural details have been published yet, but the model is moving into user-facing products.
engineering.fb.com
How WhatsApp Is Building Scam Alert With End-to-End Encryption and Verifiability Guarantees
WhatsApp/Meta Engineering details the technical architecture of a new on-device scam detection system that analyzes messages locally without breaking end-to-end encryption, using cryptographic verifiability guarantees to ensure the detection logic itself hasn't been tampered with. The system is designed to counter AI-generated phishing and social engineering lures without sending message content to Meta's servers. This is a notable systems engineering approach to deploying AI safety features in a privacy-preserving manner at scale.
rss.arxiv.org
Lost in Compaction: Evaluating Side-Constraint Loss under Context Compaction
Researchers introduce COMPINT, an evaluation suite targeting a specific failure mode in long-context LLM systems: session-scoped user constraints (e.g., "never delete emails without confirmation") are silently dropped when context compaction occurs. Across multi-turn chat, agentic trajectories, and long-horizon research tasks, current compactors retain only 17% of injected session constraints on average — often performing worse than running without compaction at all. This is a practically significant safety and reliability gap for deployed agentic systems.
huggingface.co
LFM2.5-VL-3B: Better and Faster Vision Capabilities for the Edge
Liquid AI releases LFM2.5-VL-3B, a 3-billion-parameter vision-language model optimized for edge deployment, extending their LFM2.5 architecture with multimodal capabilities. The model targets on-device inference scenarios where compute and memory are constrained, competing in the increasingly crowded small-VLM space. It is available on Hugging Face.