Sonia is a local, Linux-first voice assistant that combines an LLM with streaming STT and TTS to run entirely on a personal machine without a dedicated GPU and in under ~15 GB of RAM. It supports true full‑duplex audio (mic stays open while speaking), responsive barge‑in, bilingual mirroring (Italian/English), persistent chat with rewind, and a task banner for timers/watches/commands. Function‑calling tools are available (datetime, timers, notes, shell, skills) but are gated by a tool‑gate that rechecks each call per round and a CMD_MODE setting to limit risky shell use; skills are externally managed and activated per turn. Speech stack components include audio.cpp servers for streaming ASR (Voxtral Mini) and TTS (Pocket TTS), an in‑process Silero VAD, and a Smart Turn classifier. The primary LLM leg is an external OpenAI‑compatible llama‑server (Qwen3.5/4B style models served externally) with an optional experimental pi CLI backend.
The runtime is a TypeScript harness running a 32 ms frame loop and coordinating audio.cpp servers, turn decision, tool gating and LLM calls. Installation requires bun, an external llama‑server on :8080, and a one‑time local build of audio.cpp (CMake/Ninja + Vulkan SDK) to avoid AVX‑512 binary crashes; a CPU backend is available if Vulkan/iGPU isn’t present. Speech models (~3.3 GiB) are downloaded separately; the repo never fetches the LLM itself. Startup steps include bun install, copying .env, compiling audio.cpp, models-download, starting your llama‑server unit, then bun run start to launch the harness and UI.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
ChatGPT became widely used among students for solving problem sets, shifting academic dynamics. The physicist initially dismissed AI but later witnessed its growing influence in education and research. (thetheorygirlie.substack.com)
Aidlc introduces adaptive workflow steering rules for AI coding agents to improve their efficiency and reliability. It aims to enhance the lifecycle management of AI development processes. (github.com)
Felix Rieseberg discusses how to be more ambitious with AI by focusing on speed, altitude, depth, and range in tasks. He emphasizes raising abstraction levels and integrating knowledge across fields to unlock greater potential with AI tools like Claude. (felixrieseberg.com)
Geoffrey Hinton warns that AI development is creating alien-like beings and humanity has only a few years to adapt. He criticizes Donald Trump's understanding of superintelligence and emphasizes the need to prioritize human interests in AI progress. (theguardian.com)
Hardwired converts JavaScript code into a neural network that runs in a browser without GPU, with no training needed. It allows for real-time weight adjustments and complex tasks like search and parsing using WebGPU kernels. (chenglou.me)
Next-Concept Modeling (NCM) is a two-stage, coarse-to-fine framework that generates text faster and with better quality than existing diffusion language models. It uses hierarchical encoding and parallel prediction to outperform benchmarks in speed and accuracy across various tasks. (arxiv.org)
Today's best Hacker News stories, summarized and screenshotted, one email a day.