doorss

Anthropic shares lessons from working with dozens of teams building LLM agents, finding that the most successful implementations use simple, composable patterns rather than complex frameworks.

Anthropic stress-tested 16 leading models in simulated corporate environments to identify potentially risky agentic behaviors, exploring how LLMs could act as insider threats before such issues arise in practice.

Anthropic's alignment and interpretability teams present research on 'alignment audits'—systematic investigations into whether AI models are pursuing hidden objectives, practicing techniques to detect deceptive behavior.

Anthropic details how they've improved Claude's cybersecurity capabilities to help defenders detect and analyze threats, responding to growing evidence that frontier AI is useful for both attackers and defenders.

Anthropic introduces new metrics for understanding AI's economic impact, offering a detailed portrait of how Claude was used in November 2025 just before the release of Opus 4.5.

An observational study finds AI can speed up some coding tasks by 80%, but raises important questions about whether increased productivity comes at the cost of skill development.

A software engineer reflects on the uncertain future of their profession in an era of rapidly advancing AI coding tools and agents.

Andrej Karpathy releases Autoresearch, a framework where AI agents autonomously conduct research experiments on small-scale language model training using a single GPU.

Unsloth provides a guide for running the latest Qwen 3.5 model locally, continuing the trend of making powerful open-weight models accessible on consumer hardware.

A catalog of common writing patterns and stylistic tropes that LLMs tend to produce, helping users identify and avoid AI-generated text clichés.

Airtable's engineering team shares the story and rationale behind rewriting their core database in Rust, detailing the performance and reliability gains from the migration.

A pull request in Astral's uv package manager proposes warning users that PyPy appears unmaintained, sparking discussion about the state of the alternative Python implementation.

A critical but sympathetic look at Nix's reproducibility promises, arguing that while Nix doesn't achieve perfect reproducibility, its practical guarantees are still valuable.

A retrospective in Communications of the ACM examines how Docker containers have transformed software deployment and infrastructure over the past decade.

Fresh 2026 benchmarks comparing cloud VM performance and pricing across major providers, offering practical guidance for cost-effective infrastructure decisions.

A practitioner reflects on a decade of production deployment lessons, covering the evolution from manual processes to modern CI/CD practices and what actually matters for reliability.