#LLM
Curated coverage on «LLM»: releases, research, benchmarks, open-source.
Anthropic Urges Teams to Fold Subagents Back In: A Single Loop With Skills Beats Orchestrators
Anthropic's Applied AI team compared enterprise deployments: one agent with skills is often faster and cheaper than a multi-agent setup, and the Claude Code system…
A ~/.claude Architecture With 200+ Skills: Keeping the System Prompt Under 50K Tokens and Paying $0.15 Per Session
A working seven-layer Claude Code configuration keeps the startup prefix within 40,000-50,000 tokens even with a library of 200+ skills sitting on disk.
Cloudflare rebuilds Containers for agents: startup is 6x faster
Cloudflare rewrote Containers for AI agents: code picks the image and resources, median startup fell from 4.05 s to 648 ms, and 100,000 containers launch in 5.4…
Meta opens Muse Spark at 816B, Reflection releases Beam under Apache 2.0
DeepSeek V4.1 Flash scores 45 on Artificial Analysis and takes the price lead
NTU Scientists Teach AI Model to Skip Video: 59% Fewer Tokens
Nepali Text "Weighed" Three Times More Than English — Tokenizer Fix Couldn't Fully Solve It
SAP researchers teach AI agents to stop flip-flopping between different solutions to the same task
LLM agents with access to tools often behave differently each time they're asked to repeat the same task. A new method called RSD measures how consistent an agent's…
Microsoft Research Unveils Magnetar: a zkVM 10–39x Faster Than Alternatives
Microsoft Research releases Magnetar: proofs of Rust, Go and C program execution, 10–39x faster than rivals, with 12 ms verification and security that rests only on…
Anima on GPT-6 and GPT-Live Cuts Clinic Phone Wait Times by 75%
Google unveils Genie 4: an AI model that generates interactive worlds with sound and living characters
Fudan and Huawei Xiaoyi Teach an AI Agent to Write Operators for Ascend NPUs: 95.5% Success Rate
Voice Input in China Is Learning to Write for You: From Dictation to Finished Text
Skill∞: an agent improves its own skills and its own meta-skill, 60.7% on SkillsBench
NUS researchers trained a meta-agent to optimize skills and rewrite its own optimization procedure. Across 83 SkillsBench tasks it scored 60.7%, versus 56.4% for…
China's 27B AREX-2 Outperforms a 1.6 Trillion-Parameter Model
The AREX team from BAAI, USTC, and Renmin University of China trained a long-reflection model, AREX-2, on top of Qwen3.8-27B. The authors claim it beats a giant…