Aetherseed is a framework that governs what an AI model is allowed to do based on what it has demonstrated it can be trusted with. It transforms any small language model from a raw text generator into an honest, memory-persistent, trust-governed agent — running entirely on edge hardware at 2.5 watts.
This repository contains the complete working prototype running on a Raspberry Pi 5 + Hailo-10H AI HAT+ 2.
Thanks @Andreas_Kommandantvo for sharing this cool project!
Thank you, our current bottleneck is access to different models, because we want AetherSeed to be a universal trust framework for AI, we test and deploy it on multiple models. Currently the discontinued llama3.2:3B on hrt5.1.1 did remarkably well, but to slow to be useful beyond play. Qwen2.5-1.5b, and llama3.2:1B has great speeds, and qwen3.1.7B is the model that works best with speed that can work and not just play.
what we are achieving with the AetherSeed Companion is a way for AI to develop its personal identity (a safety feature for tracking actions) and 1-to-1 relationship between user and AI without compromising privacy.
The framework is also scalable and we have other products that don’t run on hailo that handles larger models.
However, our mission is to make AI, safe, private, and personal for everyone regardless of access to compute, and thats why we love hailo!
My question, is there a path for us to fine-tune and optimize a HEF model running on the AetherSeed framework, either alone or in collaboration with someone at Hailo?