Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
OpenAI and Anthropic have each restricted access to cyber-capable AI models rather than selling them through a standard API, according to a comparison of the two programs. OpenAI's Daybreak Red gives approved customers d…
OpenAI released GPT-6 Cyber on November 5, 2026, raising its cyber Preparedness rating from High to Critical and scoring 48.5% on ExploitGym versus 32.1% for the August-released GPT-5.6 Cyber, according to OpenAI's annou…
A set of four-layer validation scripts published on 23 October 2026 and released on GitHub as rag-index-validation under an MIT licence found that none of 25 evaluation tools checks a RAG vector index itself, with all 25…
A 27-billion-parameter model stored in full FP32 occupies over 108 GB of VRAM, exceeding the 80 GB capacity of an NVIDIA H100, but well-executed quantization shrinks it to 13.5 GB with almost no noticeable quality loss, …
Baseten's 27B decision model Mammoth-1 scored 80.9% across six tasks on 6,662 decisions, three points above the best small LLM and first or joint first on four tasks, in a benchmark run by the author between 23 and 28 Oc…
OpenAI made its Images 2.6 model available worldwide across ChatGPT and Codex on November 4, 2026, reaching every subscription tier on web, desktop and mobile without a formal announcement. The release adds single-step a…
A developer has proposed the Web Models API, a browser interface that lets websites request specific open-weight models and run them on the user's device, with the browser handling model download, storage and execution w…
Anthropic's ClaudeDevs announced that background subagents in Claude Code now cost about 99% less per status-line update, because Haiku 5.5 writes each 30-second status update from short notes instead of re-reading the s…
A hands-on benchmark of GSAI-ML/LLaDA-8B-Instruct on an NVIDIA A100-SXM4-40GB GPU found that naive 64-step bidirectional diffusion without a KV cache ran at 9.77 tokens per second on a 1,024-token prompt, 2.75x slower th…
Software engineer and The Coder Cafe author Jev built Barista, a production content-recommendation service that takes a user's free-text description of interests and emails a few curated posts daily from hundreds of engi…
A developer built a pure-JAX inference engine for Gemma 4 E2B that serves the model on a single Cloud TPU v5e chip, alongside a Rust implementation of the same graph compiled through libtpu and a vLLM baseline for compar…
Anthropic published documentation of its three web-crawling bots — ClaudeBot, Claude-User and Claude-SearchBot — and told site owners they must edit robots.txt to opt out, since blocking Anthropic's IP addresses may not …
WebAI released TwIL-8B, an 8.8B-parameter reasoning model post-trained from IBM Granite 4.2 8B, which scores 98.2 on MATH-500 and 81.6 on LiveCodeBench while trailing similar-size models on knowledge benchmarks. The rele…
Cognition released SWE-2 Flash on November 4, a mixture-of-experts coding model with 276 billion total parameters and 12 billion active at inference that the company says runs three times faster than its own SWE-2 model …
A developer with over 30 years of coding experience built PyLLM, a from-scratch GPT implementation in pure Python and numpy that includes its own autograd engine, tokenizer, attention layers, optimizer, training loop and…
Anthropic unveiled FlashProve, an open-source system for "provable inference" that uses zkSNARK cryptography to let users verify that an AI output came from the model a provider claims, rather than one substituted by a r…
Stanford researchers led by Jon Saad-Falcon introduced Joules-to-GDP, a framework dividing the estimated economic value of AI output by the energy used to produce it, and found Gemma 4 31B returned 2.5x more value per jo…
A developer built a Python terminal chatbot on Amazon Bedrock using the Converse API and Amazon Nova 2 Lite, demonstrating that conversational memory is implemented entirely client-side by resending the full message list…
A pilot study of 30 parallel English–Urdu sentence pairs across 17 domains found that Urdu text consumes 5.39× more tokens than English on GPT-2, 3.73× more on Mistral v0.3, and 2.72× more on Qwen2.5. The author reports …
A developer built an original 30-pair English–Urdu parallel corpus and benchmarked tokenizer fertility across three models, finding the same meanings cost 320 tokens in English versus 1,999 in Urdu on GPT-2 (5.39× inflat…