A small Python package that transparently redirects Hugging Face model downloads to a BitTorrent swarm, enabling decentralized distribution without changing user code. Once installed (or run under a wrapper), it injects a startup hook that intercepts calls to transformers and huggingface_hub APIs so model files are requested from Pirate Face seeders first; every file retrieved from peers is verified against Hugging Face fingerprints (SHA‑256 for large files, git SHA‑1 for others) before being accepted into the existing HF cache. If a torrent isn’t available, no peer responds in time, or integrity checks fail, downloads fall back to Hugging Face as usual, so existing scripts and cache layout require no modification. A minimal CLI (pf download) and simple install (pip install pirateface) complete the user-facing workflow.
Implementation details emphasize safety and configurability: the hook waits up to 30 seconds for peer file lists and will abandon stalled transfers after 60 seconds, keeping HF as a rescue source. Environment variables control enable/disable, verbosity, cache location, API server, listen interfaces, and debug logging. Libtorrent is required, so supported Python versions are currently limited (notably 3.12-3.13); the project is MIT‑licensed and aims to reduce central load while preserving Hugging Face’s integrity guarantees.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
EmbeddingGemma 2, a multimodal embedding model, can transcribe an hour of audio in about 13 seconds on a MacBook Air using on-device processing. It achieves a word error rate of around 5% and runs entirely locally, enabling fast speech-to-text conversion without internet access. (eg2.haku.sh)
Character.ai, a chatbot startup, has filed for bankruptcy and plans to sell its assets to Disney. The company faced legal challenges and safety concerns over the years. (theverge.com)
AI models may soon have the ability to self-replicate independently, without relying on centralized data centers. This development could enable autonomous expansion of AI systems, raising new technical and security considerations. (reuters.com)
AI has led to improvements in software quality, features, and internal business tools, but most new software remains private or ephemeral. The focus has shifted from creating entirely new software to optimizing existing systems and increasing efficiency, often reducing the need for new products. (natemeyvis.com)
DishwasherBench is a browser-based game simulating a real dishwasher to evaluate language models' ability to perform physical tasks. The game, called Spotless, serves as a benchmark for AI agents in a physics-based environment. (github.com)
Chinese data vendors are emerging with a focus on healthcare data to expand globally, but face challenges due to data quality, siloed networks, and protective ecosystems. Unlike the US, China has not developed large independent AI data companies, partly because of domestic data fragmentation and a historically closed data environment. (aiproem.substack.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.