Alibaba's Qwen team announced Qwen 3.8, a new model family with improved agentic capabilities, drawing strong community interest on Hacker News.
A prominent thread argues releasing frontier model weights functions like economic dumping that discourages private R&D. Yann LeCun strongly rebuts, citing Linux, PyTorch, and Llama as transformative open-source successes.
Simon Willison discovers Claude Code is running an unreleased version of Bun that has been rewritten from Zig to Rust, bringing performance improvements to the Anthropic coding agent.
A foundational vision-language-action model trained on over 100,000 hours of real-world manipulation data achieves SOTA 57.6% on RoboCasa365 (beating prior 46.6%) and 20.07 on RoboDojo (beating prior 13.07).
A world model that predicts data science operation effects before execution, combining structured states, cost-aware routing, and an LLM simulator to accelerate RL-based agent training by ~14x and search-based inference by 3-6x.
Proposes using video generative models as interactive game engines that respond coherently to player actions from pixels, demonstrating data-driven world simulation without explicit 3D rendering. 407 upvotes on HuggingFace.
An open audio-visual LLM for joint understanding of audio, images, and long-form video, supported by a ~7M instance Audio-Visual-Skills dataset and three-stage curriculum training. Outperforms similar-sized open models on 15+ benchmarks.
Qwen-Music produces high-fidelity songs with complete vocal singing, supporting text-to-music and reference-based music generation with strong musicality.
A public wall-clock leaderboard that benchmarks fine-tuning methods for LoRA adapters, measuring real-world speed and efficiency across techniques.
Kagi's privacy-focused browser hit 162 points and 102 comments on Hacker News, positioning it as a notable browser alternative.
An open-source AI tool positioned as a peer rather than an employee, designed for collaborative coding and problem-solving.
A platform that lets developers deploy AI agents with minimal code, aiming to lower the barrier for agent-based applications.
Source: