Unsloth has made GLM-5.3, the strongest open model to date, available for local runs, achieving 81% accuracy after significant model size reduction. The model can be run on a 256GB Mac or equivalent RAM/VRAM setups.
Anthropic unexpectedly cut off most first-party capacity to Claude 3.x models with short notice, but the affected party has secured sufficient near-term capacity from other providers. Other models like Gemini 2.5 Pro and GPT 4.1 remain unaffected.
Claude was tested on safety benchmarks for common misalignments and its best methods were evaluated on held-out benchmarks for generalization.
TencentHunyuan's Hy4 preview has achieved a significant improvement in the Code Arena, ranking #5 with 1633 points, compared to Hy3's #31 ranking. The score is based on an early AutoEval assessment, which uses a Reward Model trained on human preference data.
A configuration update for GLM-5.3-Flash aims to improve performance in certain agentic use cases, particularly for workflows where it previously underperformed Ox Alpha.
Grok Bot introduces a new harness strategy, involving daily automated checks by Tess, with Cody and Desi ensuring alignment with rules and design guidelines. The process automates verification of sites.
Anthropic has rebranded CLIO, the research that inspired LangSmith Insights, as Anthropic Insights. LangSmith Insights was internally known as CLIO, which was the original name for the research.
Agents can reach websites in 6 different ways, with Google and Microsoft focusing on one method, and a demonstration of an agent buying something on a website is provided.
Qwen3.8-Flash-Next has difficulty tracking conversations over multiple turns, answering questions from previous turns at FP8.
AI models are being trained to perform complex tasks in realistic settings, including long-horizon agentic tasks.
Fzakaria has created omniflake, a single Nix flake that contains thousands of other flakes, allowing users to access them lazily and avoid duplication, and has released it on GitHub with automatic updates.
The Fair Work Commission has condemned the use of 'plain wrong' AI legal advice in a recent case, and has introduced new measures to manage AI-dependent litigants, including a template for disclosing AI use and verifying facts. A report shows a 40% surge in cases, partly due to litigants using AI, with some using AI agents to inform their claims, while others have been criticized for relying on AI without proper understanding of the law.
Claude Code automatically includes a session URL in commit messages and PR descriptions without user consent, cluttering git history and appearing unprofessional. A proposed solution is to make session URL attribution opt-in, with a one-time prompt during onboarding.
Multiple tech giants, including Nvidia and Cerebras, are warning of potential AI attacks and offering solutions, while others like Google and Meta are working on improving AI performance and security, and HP is promoting more expensive PCs as a solution to rising cloud AI costs. Meanwhile, security concerns are rising with phishing attacks, zero-day exploits, and surveillance tools being upgraded.
The CEO of HTMX has introduced No AI Fridays, a practice where employees refrain from using AI assistants for a day to minimize cognitive debt and improve critical thinking skills. The goal is to encourage more hands-on coding and problem-solving, and to reduce reliance on AI tools. Other companies can also adopt this practice by contacting the CEO on Twitter.