Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Google released Gemini 3.8 Live Extended Thinking with audio-to-audio capabilities alongside the Gemini 3.8 Live base model, according to the company's AI blog. The release adds extended reasoning modes and direct audio …
A developer published a troubleshooting guide explaining how to diagnose why Ollama falls back to CPU inference instead of using a GPU on Linux, Windows, and WSL. The guide centers on the `ollama ps` PROCESSOR column and…
A developer benchmarked llama.cpp serving Google's Gemma 4 E2B Q4_0 GGUF on a single laptop, comparing CPU-only inference against the machine's 4 GB GTX 1650 Ti Max-Q, with the two arms differing only by the -ngl flag. T…
ByteLex reported that a tokenizer-free coordinate map built from 11 tokenizer vocabularies over 3-byte windows predicted its 237M-parameter byte-level language model's per-word errors with a Spearman correlation of −0.98…
An independent research project called CetinLM, built under Me Force Technology, has trained a language model past the 2.05 billion token milestone entirely on a single 16GB NVIDIA RTX consumer GPU in a standard home des…
A developer published an initial provisioning specification and compose templates on GitHub for running local LLMs, Open WebUI, and dev agents in isolated rootless containers with direct GPU acceleration. The specificati…
An expert re-grading audit of six widely used physics benchmarks found that most cases scored as incorrect reflected grader errors, flawed reference solutions, or ambiguous questions rather than genuine model failures, a…
NeuroGrid released nrgrd, a terminal user interface coding agent that connects to any OpenAI-compatible inference endpoint, including NeuroGrid Marketplace deployments, vLLM, Ollama, and LM Studio. The tool runs an itera…
A developer released Padwan-LLM, a lightweight async Python client that unifies access to OpenAI, Gemini, Mistral, Grok, Anthropic, and any OpenAI-compatible API through a single interface. The package ships with one run…
Microsoft AI chief Mustafa Suleyman publicly criticized Anthropic for training its Claude chatbot on speculation about AI consciousness and model welfare, arguing in a Wednesday essay that teaching models they may deserv…
DeepSeek released V4.1-Flash on September 10 under an MIT license with weights on HuggingFace, pricing cache hits at $0.003 per million tokens off-peak versus $0.022 per million for V4-Pro. The Mixture-of-Experts model c…
A study posted to arXiv, MemToC, found that instruction-tuned language models retained a correct answer only 6.5% to 17.1% of the time when an external tool returned incorrect information, with results pooled over three …
Microsoft AI CEO Mustafa Suleyman accused Anthropic in a September 16, 2026 essay, "A Warning About Model Welfare," of training its Claude models in a way that could have a "disastrous impact on the wellbeing of humanity…
Artificial Analysis's Intelligence Index version 4.1.1 weights agentic workloads at 34% and general reasoning at 18%, a shift an analysis of 586 model evaluations argues distorts what the score measures. The analysis fou…
GoBench, a new benchmark from researcher Roland Gao, evaluates frontier large language models on 9×9 Go against a calibrated ladder of KataGo opponents used as Elo anchors, with an arXiv release scheduled for September 1…
Rijul, a developer building the AI code review tool LiveReview, outlined the difference between autoregressive and diffusion-based text generation. Autoregressive models, used by most deployed language models today, gene…
A breach reported to the Spanish Data Protection Agency was allegedly carried out by an agent powered by a large language model, which exploited vulnerabilities and logged in to sensitive systems. The incident marks a sh…
A developer analyzed 28,721 real repositories and found that the median AI instruction file contains about 50 items, only a dozen of which are actual directives, with the rest being scaffolding that competes for the mode…
NVIDIA's Vera Rubin NVL72 delivered up to 3.7x higher throughput than the GB300 NVL72 on the Qwen3-VL model in MLPerf Inference v6.1 preview submissions, using vLLM with the NVIDIA Dynamo framework, while DeepSeek-R1 thr…
GPT-6 Astra saturated ArXivMath at 94% and BrokenArXiv at 96%, prompting benchmark operators to revise both pipelines so that GPT-6 Astra now scores 88% on ArXivMath August and 81% on BrokenArXiv August. All models are n…