Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
The U.S. government ordered Anthropic to suspend access to its Fable 5 and Mythos 5 AI models on June 12, 2026, citing a reported jailbreak, marking the first forced removal of a widely-deployed frontier model by export …
Claude's TUI coding agent allows users to configure its status line by executing arbitrary commands, enabling custom displays like Git branch and claw count. Codex's status line is limited to a fixed set of built-in fiel…
Developer Vincenzo released NanoEuler, a from-scratch GPT-2-scale language model training stack written in C and CUDA without PyTorch or autograd, on June 28, 2026. The 116-million-parameter project implements forward an…
Moonshot AI released Kimi K2.7-Code, an open-source 1-trillion-parameter coding agent, on June 25, claiming 30% fewer thinking tokens and higher benchmark scores than its predecessor. The model uses a Mixture-of-Experts …
Elon Musk announced on June 28 that Grok 4.5 is in private beta at SpaceX and Tesla, describing the model as based on a 1.5T V9 foundation model with supplemental training from Cursor data. The rollout follows xAI's patt…
Abram Demski reports using Claude Opus 4.8 and GPT 5.5 to conduct 'vibe research' on logical induction as a model for AI trustworthiness and recursive self-improvement, collaborating with Anson Berns to formalize results…
A new benchmark test of Ornith 1.0, a model that builds its own task scaffolds, found that providing a full shell and Python environment doubled its bug-finding performance without increasing false positives. The expande…
OpenAI launched GPT-5.6 as three tiered models—Sol, Terra, and Luna—each priced for different workloads, but benchmark comparisons show Sol leads on OpenAI's chosen metric while a competitor still leads on others, making…
A developer released Bash4LLM⁺, a lightweight, dependency-free Bash wrapper for LLM APIs that provides a secure, Bash-first CLI for OpenAI-compatible Chat Completions via Groq, with modular architecture, streaming, and d…
A researcher argues that reinforcement learning should be applied to forecasting rather than coding or math, claiming it could produce superhuman forecasters that improve decision-making across civilization. The author d…
Researchers at ETH Zurich argue in a new ICML 2026 position paper that AI safety studies on anthropomorphic behaviors like deception and scheming lack rigorous evidence, risking misallocated resources and false conclusio…
Microsoft is cutting Claude Code access for 5,000 engineers on June 30, citing cost overruns after spending an estimated $5 million per month on the AI coding tool. The move reflects a broader industry trend where compan…
AI models fall into three categories: small language models (SLMs) for efficient specialized tasks, large language models (LLMs) as generalists, and frontier models (FMs) for cutting-edge reasoning. The choice depends on…
Zhipu AI's open-weight GLM 5.2 model scored 39% F1 on IDOR detection in Semgrep's benchmark, beating Claude Code's 32% at a cost of $0.17 per vulnerability found. The test, designed to separate model performance from har…
OpenAI is considering drastic cuts to its token pricing ahead of its IPO, responding to competitive pressure from Anthropic and cheaper Chinese open-source models. The move aims to capture developer and enterprise market…
Developer Priyanka built an AI-powered app that analyzes desktop screenshots using NVIDIA NIM's vision model and prompt engineering to generate roasts, productivity advice, or interview-style feedback. The project demons…
Anthropic released its most capable AI models, Fable and Mythos, on June 9, but three days later a U.S. government export-control directive forced the company to suspend access for all customers globally, as it could not…
Sakana AI CEO David Ha argues that orchestrating many small models will outperform frontier giants, unveiling Fugu, a model orchestration system that routes tasks across specialized models. The Tokyo-based startup, backe…
A patient used Claude Code with Opus 4.8 to analyze their shoulder MRI after receiving a diagnosis of a Grade III partial-thickness tear and extensive treatment recommendations. The AI found no tear, contradicting the hu…
Oded Rechavi of QED Science claims that papers not in the top 1% of their QED score 'suck', but a review of the company's white paper finds insufficient evidence that the AI-generated score accurately measures scientific…