An a16z discussion of recursive self-improvement and what it could mean for how frontier labs and builders plan ahead.

A short, clear explainer on DeepSeek's latest approach to the long-context memory problem. Good for a quick view of the technique.

A technical look at the efficiency tricks behind DeepSeek-V4.1 Flash and why they matter for serving cost and long-context performance.

Goodfire's Eric Bigelow explains how models decide, covering critical tokens, phase shifts and in-context learning, in a long interpretability deep dive.

An inside look at how MiniMax M3 was built, covering its sparse attention design and native multimodality, from someone on the team.

Reflection AI CEO Misha Laskin on Beam, the lab's new American open-weight frontier model, and why open models matter for builders and national competitiveness.

Lecture 2 of Stanford's Autumn 2026 CME295 Transformers & LLMs course covers how large language models are built and trained, straight from the classroom.

ReflectionAI's CEO discusses Beam and the push for a leading American open-weights model, straight from the lab building it.

Theo's long hands-on look at a new small model, covering where it holds up and where it falls short for real builder workloads.

First hands-on test of Claude Haiku 5.5, Anthropic's new small model that cuts API prices about 90% while keeping coding and computer-use strength.

A $700 head-to-head of Haiku 5.5, Sonnet 5.5 and Opus 5.5 on real automation and coding tasks, with cost-per-task numbers you can use to pick a tier.

A long, practical look at how to evaluate new models yourself instead of trusting launch-day benchmarks.

Sam Witteveen breaks down Microsoft's new local models and what the growing on-device push means for builders who want inference off the cloud.

A long interview with Periodic Labs on autonomous labs and AI scientists, useful for seeing where model-driven discovery is heading.

Artificial Analysis and NVIDIA walk through how to evaluate open-weight models properly, including benchmarks and tradeoffs.

Bijan Bowen puts Claude Haiku 5.5 through 30 minutes of hands-on tests to show where the cheap model actually holds up against the bigger ones.

Official 2.5-minute recap of everything Anthropic shipped in Claude in September, a fast way to catch up.

Fireship's six-minute rundown of this week's open-weight race: Mistral Large 4 versus Reflection AI's and Moonshot's new releases.

Matthew Berman digs into Mistral Large 4, Mistral's return to the frontier, with benchmarks, pricing and what it means for open-weight model builders.

OpenAI's official launch of GPT-6 in ChatGPT alongside Intelligent UI, which answers with interactive, generated interfaces instead of plain text.

Official OpenAI launch of the Decisions API: hand the model an input plus a fixed set of answers and use its fast choice to route leads, classify images and steer voice…

Bijan Bowen puts Mistral Large 4 through a thorough first-test battery of coding, reasoning and agentic tasks to see whether Mistral is back in the frontier race.

Google DeepMind introduces EmbeddingGemma 2, a 740M-parameter open model that embeds text, images, video and audio in one space for fully on-device search and RAG.

Full hands-on course building and training a GLM-5.3-Flash-style multimodal model from scratch in Python, covering tokenization, MoE, sparse attention and RL…

An analysis of the renewed push for US open-weight models and what it means for builders choosing open models.

IndyDevDan reads OpenRouter's state-of-models data to show which models agentic engineers are actually paying for, how Jev and open-weights are shifting usage, and what…

Side-by-side tests of the new Qwen3.8-27B variants to work out which open-weights checkpoint to actually run for coding, reasoning and agent work.

A DeepMind researcher explains Gemini Robotics 2 and cross-embodiment learning, a deep dive on how one model can drive many robot bodies.

Clear explainer of latent-reasoning models that think in continuous space instead of tokens, and why that changes the cost and quality trade-off.

A deep dive with the MIT researcher behind Recursive Language Models on how models that call themselves over long contexts beat plain scaling on hard tasks.

AI Explained digs into OpenAI's security disclosures on frontier models escaping sandboxes, plus Gemini 4 Argon and recursive self-improvement, with his usual…

Sam Witteveen's hands-on first look at Gemini 4 Argon, Google's new frontier model, covering its huge output window, coding strength and what it means for builders on…

Matthew Berman breaks down how Claude Sonnet 5.5 lands so close to Opus-class performance and what the release means for agent builders choosing a default model.

Nate Herk runs Sonnet 5.5 and Opus 5.5 head-to-head on real automation and coding tasks so you can decide when the cheaper model is enough.

OpenAI's full DevDay keynote: new models, API capabilities and agent tooling. Worth watching alongside the Anthropic keynote to see where the two platforms diverge.

Tina Huang walks through which local open-weight model to run at every hardware size, from a laptop to a multi-GPU box, with concrete setup and speed comparisons.

Anthropic's developer keynote: the current Claude model lineup, what changed for agents and coding, and the roadmap for the API and Claude Code. The clearest official…

How Claude Skills work and how to package your own repeatable workflows so the model does them the same way every time. Practical, beginner-friendly, and directly…

The London edition of Anthropic's developer conference: model updates, Claude Code and Agent SDK demos, and customer builds on stage.

The creator of Claude Code on how frontier models changed software engineering, what 'solved' actually means, and what builders should do with the slack. A Sequoia AI…