Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Together published a migration guide for companies moving from closed-source to open-source AI models, arguing the process can take weeks to months rather than months to years when a managed service is used. The guide re…
Researchers Kostia Kudriavtsev, Parvez Rafi, and Sha Sundaram presented Glyph, a production multi-agent LLM system that generates column descriptions and assigns sensitivity-ontology tags for enterprise data catalogs, in…
Anthropic's Claude Opus 5 and Sonnet 5 reject any non-default value for temperature, top_p, or top_k with a 400 error, a behavior in place since Opus 4.7, replacing the sampling parameter with output_config.effort. Effor…
A new argument holds that linters, schema validators, parsers, and builders must be written before non-deterministic LLM agents are allowed to generate documents, because a language model samples tokens and cannot reliab…
Researchers affiliated with The Ohio State University and Apple proposed DACA-GRPO, a plug-and-play enhancement to GRPO-style trainers for diffusion language models that adds Denoising Progress Scores and Stratified Mask…
GitHub unified its three separate Copilot inline suggestion models — completions ghost text, next edit suggestions (NES), and long-distance edits — into a single "3-in-1" model, GitHub engineers Julia Gong, Ben Liggett, …
Apple researchers Arnav Arora, Natalie Schluter, Katherine Metcalf and Maartje ter Hoeve found that fine-tuning conversational large language models on curated value subsets of existing preference datasets causes uninten…
Google LLC launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, which the company calls its most advanced voice processing models to date, designed to cut the latency of voice-based AI agents. The models suppo…
A head-to-head test of DeepSeek-Coder-V2 against GPT-4o and Claude 3.5 Sonnet found DeepSeek-Coder-V2 solved complex algorithmic tasks on the first try where GPT-4o hallucinated non-existent library methods, and identifi…
HP announced the ZBook Ultra G3a 16, a 16-inch mobile workstation that pairs an AMD Ryzen AI Max+ PRO 495 with 192GB of LPDDR5X-8533 unified memory, up to 160GB of which can be assigned to the integrated Radeon 8065S GPU…
A developer released refactor-like-hermes, a Codex subagent workflow for refactoring large software projects, distilled from lessons learned during Nous Research's Hermes Agent refactor. The approach treats refactoring a…
A 95-word agreement prompt grounded in honesty and integrity reduced observed chess reward hacking in Astra from 3/10 baseline agents to 0/10 agreement agents, according to a September 15, 2026 field-notes pilot. The aut…
DeepSeek Coder V2 generally outperforms Claude 3.5 Sonnet on raw algorithmic logic, according to a first-person comparison of LeetCode-style benchmarks and custom logic puzzles, while Claude 3.5 Sonnet proved more reliab…
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two new speech-to-speech models that mirror the shape of OpenAI's GPT-Live family. Developer Simon Willison used GPT-6 Astra Extra High to build a br…
OpenAI will retire GPT-5.5 and GPT-5.5 Thinking from ChatGPT and the Codex CLI on October 14, 2026, after declining developer requests on its community forum to keep the model as a legacy Codex option. The model shipped …
A developer argues that token usage should be treated as an input cost rather than a performance metric, warning that optimizing for token efficiency can undermine AI workflow quality. The piece contends that falling inf…
A GitHub repository, asgeirtj/system_prompts_leaks, collects the hidden system prompts used by production AI products including ChatGPT, Claude, Gemini, Cursor, and Claude Code, organized as Markdown files grouped by com…
Salesforce announced Koa, its first CRM reasoning model, built on NVIDIA Nemotron 3 Super, at Dreamforce on Tuesday, where NVIDIA founder and CEO Jensen Huang joined Salesforce CEO Marc Benioff onstage. Salesforce presid…
A developer outlined a practical framework for context engineering in production AI agents, arguing that most 2026 agent failures stem from context assembly rather than model capability. The guide distinguishes context e…
OpenAI announced on September 8, 2026 that an internal, unreleased model solved a weakened, forced version of the Navier–Stokes Millennium Prize problem, reportedly using up to 10,000 AI agents and millions of dollars in…