Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
In a podcast episode of complexity.fm, Bas Steins and Michal Martinka discussed why Python developers are adopting Rust for tooling like Ruff, uv, and Polars, citing Rust's strictness and compiler guardrails. They also a…
Anthropic accused Alibaba of running 25,000 fake accounts to steal nearly 29 million conversations from its Claude AI model, marking the largest known distillation attack. The accusation, detailed in a letter to the Whit…
Anysphere, the company behind the Cursor code editor, has trained its own frontier AI model called Composer 2.5 on Qwen 2.5 Coder using novel reinforcement learning techniques, enabling it to compete with GPT-5.5 and Opu…
Y Combinator CEO Garry Tan released G Stack, a CLAUDE.md configuration that equips AI coding agents with 23 specialist roles and 8 power tools to function as a virtual engineering team. The tool, installable in Claude Co…
GLM 5.2 from Zhipu AI outperforms Anthropic's Claude Opus 4.8 in UI generation tasks including 3D scenes, dashboards, landing pages, and mini games, while costing up to 10 times less. The comparison reveals that GLM 5.2 …
Anthropic's Claude Code introduces /goal and /routines commands that enable autonomous, scheduled workflows. The /goal command keeps Claude running until specific completion criteria are met, while /routines allows users…
Developers can now run Zhipu AI's GLM 5.2 model inside Anthropic's Claude Code coding environment by routing API calls through OpenRouter, using the ANTHROPIC_BASE_URL environment variable to point to OpenRouter's endpoi…
AI agent failures often stem from poor context management rather than model limitations. Static context loads fixed information every session, while dynamic context retrieves data on demand, balancing token efficiency an…
Google's AI coding guide defines a spectrum from vibe coding to agentic engineering, helping developers choose the right level of AI autonomy for their projects. The spectrum ranges from low-autonomy AI assistance for si…
Zhipu AI released GLM 5.2, a 744B parameter open-weight mixture-of-experts model that rivals Claude Opus in coding and visual design quality at lower inference cost. The model's MoE architecture activates only a subset o…
Zhipu AI's GLM 5.2 uses Index Share, a technique that reuses sparse attention indexers across four layers, reducing compute operations by 2.9x at 1M token context. This optimization makes long-context inference more affo…
Google's Gemma 4 31B outperforms Alibaba's Qwen 3.6 27B on agentic code review tasks, finishing faster due to superior Multi-Token Prediction (MTP) design, according to benchmarks and field reports. While Qwen 3.6 leads …
Anthropic launched Claude Tag, an always-on AI agent that joins Slack channels as a persistent team member, replacing its older Slack connector. The agent builds long-term memory of projects and decisions, and can be tag…
Sail Research, founded by Neil Movva and Samir Menon, announced a Series A investment from Kleiner Perkins, Redpoint, and Sequoia for its asynchronous inference platform that reduces costs by up to 6x compared to real-ti…
A comparison of four terminal coding agents in 2026—goose, Claude Code, OpenCode, and Pi—finds that frontier models have largely converged, making the harness around them the key differentiator. Claude Code is a propriet…
Loop engineering, a new trend in software development, allows engineers to design autonomous coding agents that operate without human intervention, freeing them from manual prompting. The concept, popularized by Peter St…
Researchers propose Hyperbolic Sparse Autoencoders (HypSAEs) to overcome geometric limitations in mechanistic interpretability of large language models. By applying hyperbolic geometry as a weight-space regularizer, the …
A security researcher warns that using GitHub issues as untrusted input for LLM agents in public repositories creates vectors for prompt injection, shell injection, and social engineering attacks. Hidden content such as …
Multiple analyses, including a Washington Post controlled test and a Centre for Policy Studies report, found that major AI chatbots like ChatGPT produce predominantly left-leaning political answers, with over 80% of resp…
Anthropic engineer Thariq Shihipar published a post arguing that HTML outputs improve human-agent workflows over Markdown, releasing 20 self-contained HTML examples that drew 4.4 million views in 16 hours. Comparative te…