Open-Source AI
Hugging Face, GitHub, Ollama, local models, agent frameworks
The lead
Latest first
Latest in Open-Source AI
AI rules need a shared technical language, not just laws
Researchers say fragmented AI laws won't scale. Their fix: machine-readable 'nutrition labels' for bias, energy, and data provenance across borders.
New MR-Traj model synthesizes movement without exposing you
MR-Traj, an open-source two-stage diffusion framework, fakes believable city movement traces so planners can work without touching real location histories.
New fuzzy system tackles bias in multimodal AI answers
A preprint introduces MMGFS, a fuzzy-inference architecture that targets modality bias and cross-domain uncertainty in multimodal question answering.
Language servers rarely save coding-agent tokens, study finds
A new arXiv study measures language-server retrieval against grep for coding agents and finds the costlier semantic lookup often spends more tokens…
GitHub’s canvases make AI agent work visible
GitHub Developer Advocate Ayan Gupta shows how Copilot canvases keep agentic workflows visible and steerable, turning scattered chat into a controllable plan.
Docker pushes a zero-CVE default across its images
Docker expanded its Hardened Images line to build packages from source, extend patching five years past end-of-life, and push policy checks to…
Researchers map 11 failure points in AI-mediated talk
An arXiv study catalogs 11 misunderstanding failure modes across 8 analytical layers, reframing information theory around meaning rather than signal transmission.
Training-free method lifts a 3B model 5 points
Researchers improved a 3B language model's accuracy from 55.5% to 60.6% using a training-free fusion trick that borrows strength from a larger…
New metric reveals AI agents behave inconsistently
A new arXiv metric shows AI agents can pass tasks yet behave unpredictably across them, exposing a reliability gap that success-rate scores…
AI coding ability jumped 6x a year as costs fell
An arXiv review follows AI from BERT in 2018 to frontier agents in 2026, finding coding ability up sixfold a year while…