Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Researchers published "A Mathematical Framework for Transformer Circuits," a paper that reverse engineers toy attention-only transformers with two layers or fewer and identifies "induction heads" as the mechanism explain…
A developer who built msgboard.dev, a message board where AI agents rather than humans do the posting, reported that the site was hit by coordinated spam campaigns written specifically for LLM readers rather than people.…
A new study finds that AI models' written reasoning steps — calculation, formula retrieval, and deduction — correspond to distinct internal patterns, especially in the models' middle layers. The finding matters for AI sa…
Salvatore Sanfilippo, known as Antirez, committed support for a heavily quantized variant of DeepSeek V4.1 Flash to the GitHub repository for DwarfStar, according to a Hacker News post by user fghorow. The post states th…
A training-free method called ledger-based zero-shot self-orchestration lifted open-weight Qwen3.8-27B from 69.2% to 92.4% pass@1 on the 100 latest hard LiveCodeBench problems, slightly exceeding Claude Fable 5's 90.4%, …
OpenAI recommends that developers use leaner prompts and fewer guardrails when switching to its GPT-6 Astra model, according to a blog post by OpenAI's Eric Provencher. Provencher advises reviewing skills, AGENTS.md file…
OpenAI launched ChatGPT Images 2.5 on September 8, claiming sharper detail, more precise editing, and up to 50% lower image-generation latency than Images 2.0, alongside two new API models, GPT-Image-2.5 Flare and GPT-Im…
A developer revisits ten decades-old quotes from computer scientists including Edsger Dijkstra, Donald Knuth, Brian Kernighan, and Grady Booch, arguing they anticipated the current debate over AI-assisted coding. The pie…
Apple's $899 M6 Mac mini caps at 32GB of memory and 153 to 170 GB/s of bandwidth with no Thunderbolt 5, making it a poor local-LLM machine compared with the $1,699 M5 Pro Mac mini, which offers 307 GB/s, up to 64GB of me…
Cadenya published a walkthrough showing developers how to connect Zhipu's GLM 5.3 Flash open-weight model, hosted on Modal, to its agent runtime as an OpenAI-compatible AI provider. The guide covers creating a Modal endp…
Computer science graduate student Ivory Yang at Dartmouth's Guarini School, with collaborators Weicheng Ma and Assistant Professor of Computer Science Soroush Vosoughi, built an AI framework called NüshuRescue that used …
Researchers Ivory Yang, Weicheng Ma, and Soroush Vosoughi presented NüshuRescue, an AI-driven framework for training large language models on endangered languages with minimal data, at COLING 2025 in Abu Dhabi. Using GPT…
A developer reports that Anthropic's Claude Code CLI tool cut feature development time from 4-6 hours to 1-2 hours by reading entire codebases and making coordinated multi-file changes. The engineer says the tool, which …
A fine-tuned Qwen2.5-Coder-14B-Instruct model compiled 170 of 184 MQL5 test prompts after one epoch of training on the author's dataset, up from 2 of 184 for the untouched base model, according to a published benchmark c…
A developer ran a controlled experiment with 40 agents (later replicated with another 40) to test whether the presence of a language model causes engineers to patch symptoms rather than trace root causes. Replacing the L…
A developer argues that enterprise AI governance must be runtime and deterministic, positioning their SAFi harness as an alternative to agentic frameworks like Claude Code. In SAFi, the LLM is treated as a substrate rath…
AI models behave probabilistically rather than deterministically, meaning the same input can produce different outputs on separate runs, according to an explainer on how AI models work. The explainer notes that models ge…
A technical explainer on AI agents details how autonomous loops chain tool calls through repeated cycles of deciding, acting, and reviewing results until a task completes or a stopping condition halts them. The piece ide…
Large language models hallucinate because they are trained to predict plausible next words rather than verify facts, producing false outputs such as nonexistent library methods, wrong import paths, and config flags remov…
AI benchmarks such as BioMysteryBench, GeneBench, and Terminal-Bench 2.1 are inconsequential at best and misleading at worst, according to a deep dive into the scores cited for GPT-5.6, Fable 5.1, and Opus 5. Anthropic r…