Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A comparison of raw HTML and Markdown formats for feeding web content into large language models shows that HTML contains significant boilerplate—navigation menus, scripts, cookie banners, and ads—that wastes tokens with…
Large language model companies face a profitability crisis as open-weight models increasingly match the performance of proprietary systems at a fraction of the cost. Frontier labs are being squeezed from both ends — chea…
Moonshot AI's Kimi K2.7 Code, a coding model designed for long-horizon programming tasks with native multimodal text and vision input, is now available on Vercel's AI Gateway. The model supports frontend development, Dev…
Google released Diffusion Gemma in early 2025 as its first open-weight diffusion language model, using a masked diffusion approach that generates text by starting with noise and iteratively refining it into coherent outp…
Anthropic released Claude Fable 5, its most capable model for sustained, complex reasoning, designed specifically for agentic workflows that require multi-step decision-making and tool orchestration. The model features a…
Large language models can complete simple analogies like "surgeon is to scalpel as painter is to brush," but they struggle with higher-order structural reasoning that requires transferring relational insights between dif…
NVIDIA released Nemotron 3 Ultra, a 550-billion-parameter open-weight model, on June 1, 2026, and it became available on Ollama's cloud three days later. The model uses a sparse Mixture-of-Experts design with 55 billion …
Pydantic AI, an open-source Python framework from the Pydantic team, and Anthropic's Claude Agent SDK offer distinct approaches for building production AI agents, with Pydantic AI focusing on model-agnostic, strongly-typ…
New research shows that larger AI models no longer reliably outperform smaller ones on analogical reasoning tasks, undermining the scaling laws that have guided AI development for years. The findings suggest that increas…
Anthropic released Claude Fable 5, its most capable publicly available model, as part of a new Mythos-class tier designed for complex, multi-step reasoning and agentic workflows. The model features an extended context wi…
Anthropic's Claude Fable 5 autonomously debugged a CSS bug across multiple browsers in a 15-step investigation, highlighting the rapid maturation of AI coding agents. OpenAI acquired Ona (formerly Gitpod) to enable long-…
Anthropic released Claude Fable 5 with undisclosed distillation-prevention guardrails, prompting a public apology after Simon Willison exposed the hidden safeguards. Separately, a DN42 network administrator went bankrupt…
A new paper from MiniMax, MaxProof, trains a model to write competition-math proofs and achieves high scores on IMO and USAMO problems, but a cross-verification experiment reveals that among proofs the training verifier …
A new 'bitter lesson' in search technology reveals that algorithms matter less than incentivizing content creators to optimize for a search engine, as demonstrated by Google's PageRank and now by coding agents like Claud…
A developer reports sustained use of PEEK, a system that maintains a compact context map in an agent's prompt to orient it within recurring external contexts. The map, a few hundred tokens appended to the system prompt, …
Pydantic AI and DSPy are two Python frameworks for building LLM-powered applications that take fundamentally different approaches, with Pydantic AI emphasizing imperative agent design and type-safe structured outputs whi…
AI SDK 7 introduces `HarnessAgent`, a unified API that lets developers run established agent harnesses including Claude Code, Codex, and Pi through a single interface. The new abstraction allows users to switch between d…
Anthropic's Claude Fable 5 scored 72% on SWEBench Pro, outperforming OpenAI's GPT 5.5 at 68% in real-world bug-fixing and feature implementation tasks drawn from major open-source repositories. The models also diverged o…
Anthropic's 2026 Claude model lineup includes four tiers — Fable 5, Opus, Sonnet, and Haiku — each designed for specific task complexities and cost profiles. Teams that implement intelligent model routing by matching tas…
Anthropic's previously withheld AI model, Claude Fable, successfully generated a complete 1,296-line video game called "Shepherd's Dog" in a single 45-minute session costing over 20 euros in tokens. The user reported the…