stuffinsider

A research paper proposing a framework for training AI agents to perform multi-turn, long-horizon planning through on-policy distillation from single and multiple teacher models. The work explores pre-training and post-training paradigms.

First evidence of pending Qwen3.7 open-weights release — Qwen3.7-flash appears on OpenRouter

Qwen3.7-flash, a small MoE model with a 1M context window, appeared on OpenRouter with significantly cheaper pricing than Qwen3.6 flash. This is considered evidence of an imminent open-weights release of Qwen3.7.

OpenAI management decided not to join the Open Secure AI Alliance, internal backlash reported

OpenAI's management declined to participate in the Open Secure AI Alliance founded by Nvidia CEO Jensen Huang. The decision was met with employee backlash, according to internal reports shared on Reddit.

HackerNews

Outerport (YC S24) launches instant hot-swapping for AI model weights

Outerport offers a service that allows AI models to swap weights on-the-fly without downtime, enabling A/B testing, rapid iteration, and dynamic model updates in production.

HuggingFace

Kimi K3 Open Frontier Intelligence weights released

Moonshot AI released the open-weights Kimi K3 model, a frontier-level large language model. The release includes weights and supporting infrastructure for large-scale deployment.

GitHub

Claude Code v2.1.220 released

Anthropic released version 2.1.220 of Claude Code, the CLI coding assistant. The update likely includes bug fixes and performance improvements.

ArXiv

The Regression Tax: Decomposing why skills help and hurt LLM agents

An ArXiv paper analyzes how learned skills can both benefit and hinder LLM agents, introducing the concept of a 'regression tax' that explains performance trade-offs.

HuggingFace

Scaling Native Multimodal Pre-Training From Scratch

A research paper explores training a multimodal model from scratch using a unified architecture, achieving strong performance across vision-language tasks. It received 12 upvotes on Hugging Face.

HuggingFace

Skill Self-Play: Pushing the frontier of LLM capability with co-evolving skills

A new paper introduces Skill Self-Play, a technique where an LLM iteratively generates and practices tasks to co-evaluate and improve its own skills, achieving state-of-the-art results on several benchmarks. The paper received 19 upvotes on Hugging Face.

China state media says support for open AI models has limits