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 iLostCount project published a token-conversion cheat sheet stating that 1,000 words of ordinary English is roughly 1,300 tokens, while 1,000 tokens is roughly 750 words or about 4,000 characters. The guide notes the…
A developer argues that LLM code-review verdicts like "block" or severity scores are uncalibrated hard labels that teams wrongly treat as ground truth, and proposes instead treating the model's output as a single input f…
TypeSafe unveiled Jev, which it calls its first System One Model, an AI model that abandons text generation to output typed decisions and probabilities directly inside software; Jev entered early access on September 14th…
TypeSafe's Jev 1.13.0 model outperformed GPT-4o mini (2024-07-18) on a small text-classification experiment run by developer Emil Silvis, averaging 91.7% accuracy versus 81.3% across three public datasets. Jev scored 86.…
GPT-6 Astra in Codex scored 68.8% pass@1 and Claude Fable 5 in Claude Code scored 67.3% pass@1 on 333 GameDevBench Godot 4 tasks, according to the GameDevBench leaderboard from Carnegie Mellon and Princeton researchers p…
Zed Industries' open-source code editor Zed, written from scratch in Rust and licensed primarily under GPL-3.0-or-later, ships an Agent Panel that lets developers run Claude, Codex, or Gemini CLI inside the editor and co…
Anthropic has rebuilt Claude Projects, currently in beta for Claude Code, into an agentic coordinator that delegates work to cloud threads and sub-agents rather than executing tasks directly. Each thread runs as a full C…
Anthropic is rolling out a redesigned Claude Projects feature to selected Pro and Max subscribers using Claude Code cloud sessions, with existing web and desktop Projects users excluded from the initial phase. The coordi…
IBM Research found that a ReAct agent running on GPT-4.1 succeeded on 77.4% of runs on average but passed all five runs on only 53.0% of AppWorld's 168 test_normal tasks, a 24.4-percentage-point consistency gap, accordin…
OpenAI introduced Astra for Law on September 17, 2026, a legal configuration built on its existing GPT-6 Astra flagship model rather than a new model, combining a search index over US law, custom instructions, a higher e…
A post on osr.im titled "Bring back old Opus in Claude Code" argues for restoring access to an older Opus model in Anthropic's Claude Code tool, drawing 1 point and 1 comment on Hacker News. The submission links to a not…
A U.S. military operation against a Chinese vessel was aborted at the last minute this spring after officials discovered the intelligence driving it had been hallucinated by an AI chatbot, CNN reported on Friday. The fal…
A preprint analysis of 32,944 arXiv mathematics submissions found that 24.14% acknowledged AI use by August 20, 2026, up from under 5% in March, according to the paper "The Gold Rush in AI4Math: Where Are We Now?" by Jia…
A developer has begun building "Event Horizon," a Windows diagnostic tool that collects error and issue reports from connected PCs via Sysmon and PowerShell logging and uses a local LLM to consolidate the data into simpl…
A developer found that two LLMs given identical prompts and tool descriptions produced different JSON structures for a patching tool, with the larger model correctly emitting the framework's proprietary field names while…
A developer implemented a 6502 microprocessor emulator written in Markdown and executed it using OpenCode with the GLM 5.1 model running on Grunden.ai, following earlier experiments with a user-space IP stack and a BASIC…
A developer deployed Google's Gemma 4 E2B to a single Tesla T4 GPU on a Compute Engine VM using vLLM 0.29.0, finding that the QAT int4-weight checkpoint decodes at 72.31 tok/s per stream versus 40.44 tok/s for bf16, a 1.…
A developer published a tutorial showing how to build a streaming LLM chatbot in under 60 lines of Node.js using the OpenAI SDK pointed at Groq's API endpoint. The guide demonstrates the stream: true pattern, in which th…
Anthropic said staff from Accenture's AI division Faculty will work inside the company to evaluate and red-team its models, conduct alignment assessments, and test model safeguards, with both companies expecting to inves…
TypeSeer launched as a system-wide AI autocomplete for macOS that runs a language model directly on Apple Silicon so typed text never leaves the device. The app offers grey-text suggestions in any standard macOS text fie…