Open-Source AI
Hugging Face, GitHub, Ollama, local models, agent frameworks
The lead
Latest first
Latest in Open-Source AI
Two centuries of botany parsed into 55,737 traits
The arXiv framework pairs rule-based parsers with LLM ensembles to turn centuries of botanical texts into 55,737 trait records across 4,961 species.
Plain prompts can’t reliably score shared decisions
Asking an LLM to judge shared decisions in pediatric surgery talks in one prompt fails; a small trained model scores higher, yet…
Safety benchmarks misfire on small language models
A study of five AI safety suites against 26 small language models found ambiguous scores that make leaderboard rankings mathematically fragile.
Unwritten benchmark breaks GPT-4o and Gemini
A new arXiv benchmark asks AI to read unwritten words from pen sounds and hand motion. Humans top 80%; GPT-4o and Gemini…
GitHub lets companies control Copilot inside JetBrains
GitHub now lets enterprise admins enforce Copilot plugin, MCP server, telemetry, and permission rules centrally in JetBrains IDEs, curbing unsanctioned AI tool…
GitHub lets admins revoke credentials by type
GitHub now lets enterprise and org admins revoke one credential type at a time during a breach, containing damage without locking out…
Panchagnula study curbs political bias in LLM summaries
Tejaswi V. Panchagnula and five co-authors show recursive self-correction lifts LLM political-neutrality scores from 2.14 to 4.56 after adversarial bias injection.
Cursor flaw let a harmless command run attacker code
Pillar Security found CVE-2026-22708 in Cursor: shell built-ins bypassed an approval allowlist, so a command a developer approved could execute attacker code.
Wiola 13M swaps the standard block for spiral attention
Wiola 13M, a 13-million-parameter model, replaces the usual transformer block with spiral attention and a butterfly feed-forward to run efficiently on device.
Ethical RL agents fail where averages hide harm
University of Waterloo researchers show ethical RL agents post clean averages while harming in single episodes; per-episode training and evaluation close the…