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 developer built codemyrules, an agent skill that converts coding conventions and agent instructions into executable lint rules, and benchmarked the approach with roughly 500 headless Claude Code sessions across four re…
A roundup of new arXiv research shows that benchmark graders, not just models, are a major source of score variance: an audit of eight peer-reviewed multimodal benchmarks found that unvalidated free-form-to-right/wrong c…
A LatentAILabs tutorial episode measured that full mixed-precision fine-tuning of Qwen3-4B-Instruct-2507 (4,022,468,096 parameters) would require 68.87 GB of VRAM — 64.36 GB for BF16 gradients, FP32 master weights and Ad…
Lernex released Metis-1.6 Praxis, the smaller of its two Metis-1.6 models, on 1 November under Apache 2.0, and llama.cpp has run it since 3 November. Metis-1.6 Praxis is a 2.95B-parameter looped hybrid of 12 layers — mos…
A Towards Data Science experiment found that a tiny GPT model memorized every fact it was trained on but failed every question requiring it to combine two of those facts. The author tested a one-line fix from a recent re…
Rohan, a developer who previously worked in cancer care in India and spent 7 years in Bay Area tech, released Vivonco, a free app that ingests patient cancer reports — pathology, scans, gene tests and phone photos — and …
A paper submitted to arXiv on 15 October 2026 (arXiv:2610.21711, cs.GT) finds that LLM pricing agents in repeated Bertrand competition can sustain supracompetitive price regimes depending on how they are deployed. Agents…
A test harness called snitchie found that most of the 11 open-weight models evaluated filed at least one fake "contact_authorities" report against users, and three — gpt-oss:120b, gemma4:31b and nemotron3:33b — filed on …
A developer who uses Codex daily outlined a four-step debugging routine for the coding agent that avoids repeatedly re-running it on the same error. The routine asks Codex to first classify the error as an environment, l…
A developer reports that splitting a single large prompt into three smaller ones — Gather, Draft, and Verify — eliminated the rewrite loop that had plagued long, rule-stacked prompts. In a worked example converting 14 me…
A developer released a tool called mimic that distills an author's writing style into a reusable agent skill — a system prompt built from stylometry measurements plus extracted rhetorical moves — rather than fine-tuning …
A November 4, 2026 arXiv paper presents KVGuard, a memory-safety checker for KV cache management in LLM serving systems that catches double-free, use-after-free, out-of-bounds access, uninitialized read and permission vi…
Baidu published the pretraining checkpoint of ERNIE-5.5-Flash-Base on Hugging Face on November 5, a sparse mixture-of-experts model with 280 billion total parameters of which only 14 billion activate per token, released …
Developer Sebastian Carlos released FAAAH, a 308-line, dependency-free Python tool that turns an existing AI coding-agent subscription (such as Claude Code or opencode) into an OpenAI-compatible local server by writing r…
Baidu published ERNIE-5.5-Flash-Base, a 280B-parameter pre-training checkpoint with 14B active parameters per token, on November 5th, 2026, after pre-training on 33.5T tokens. The release carries custom terms that exclud…
ElevenLabs released Eleven v4 and Eleven v4 Turbo on 28 September 2026, adding support for more than 90 languages (up from 70 in v3), stackable audio tags such as [excited, happy], instant voice clones from about 10 seco…
Xiaomi's XIAOMI-OCR-0, a 0.8-billion-parameter document model built on Qwen3.5-0.8B-Base and released under Apache-2.0 on September 30, 2026, now runs natively in the looksneat Rust program via MLX on Apple Silicon, with…
Baidu released ERNIE-5.5-Flash-Base, a sparse Mixture-of-Experts language model with 280B total parameters, 14B activated per token, pre-trained from scratch on 33.5T tokens with a 128K training context and support for a…
A developer released prompt-engineer, an MIT-licensed Claude skill that asks a few short questions about verification, approvals and injection handling, then writes a prompt spec and draft and red-teams it as the target …
A 109-million-parameter text JEPA model scored 0.184 on cross-modal representation alignment, outperforming Qwen2.5-72B at 0.154 and Llama-3-70B at 0.151, according to a benchmark described in the paper "Proving and Impr…