Outcry: A four-layer activist AI that runs on a four-year-old phone
Outcry is a four-layer on-device AI model designed for activists, capable of running on a four-year-old smartphone with approximately 3 GB of RAM and no network connection. It combines a quantized base model, a QLoRA domain adapter trained on activist content, contrastive activation steering for adjustable radicalism, and a forthcoming soft-prompt wellbeing layer. The system prioritizes privacy, autonomy, and mental wellbeing while enabling strategic political discussions without reliance on corporate-controlled AI platforms.
- ▪Outcry runs entirely on-device using Apple MLX, requiring no network calls and fitting within ~3 GB of RAM.
- ▪The model includes a QLoRA adapter trained on 9,000 organizing conversations and activist literature to support movement-specific tasks.
- ▪Contrastive activation steering allows users to adjust the model's radicalism level without significantly increasing harmful outputs.
- ▪A soft-prompt wellbeing layer, based on Ren et al. (2026), is included to improve the model’s internal state and user interaction quality.
- ▪The system uses a precomputed KV cache for faster inference and is designed to empower activists with private, uncensored AI assistance.
Opening excerpt (first ~120 words) tap to expand
ContentsIntroThe stackLayer detailKV cacheWhy this stackWhat this is not (yet)Build logCite ThisWork With usRelatedLocal Models · Alignment · Engineering · May 2026Outcry: a four-layer activist AI that runs on a four-year-old phoneA quantized open-weights base, a QLoRA adapter, contrastive activation steering, and a soft-prompt wellbeing layer — composed into one inference pipeline that fits in ~3 GB of RAM and never makes a network call. This page documents the stack.By Micah Bornfree, PhD·Outcry Research·Updated May 8, 2026Email UsRead the report ↓Try the iOS/Mac app ↗AbstractOutcry is a four-layer on-device activist AI: a low-bit quantized open-weights base, a QLoRA domain adapter trained on activist literature and 9,000 organizing conversations, contrastive activation steering exposed…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Outcry AI.
Discussion
0 commentsMore from Outcry AI
