Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Anthropic's Claude Opus 4.8 defeated Kimi K2.7 Code 37-31 in a head-to-head coding benchmark, demonstrating superior precision and judgment across most tasks. The only tie occurred on a rigid data transformation task, hi…
A developer added Markdown file support to the Sift RAG stack, touching validation, extraction, MIME types, and a database schema migration. The change required updates across six layers including API validation, content…
LangChain.js provides a standardized framework for building LLM applications in TypeScript and Node.js, enabling developers to compose prompts, models, tools, and retrievers into reusable pipelines and agents. The framew…
A developer built a search engine over 50k saved tweets using hybrid retrieval and reranking, then added a single endpoint to turn it into an answer engine. By wiring Gemma 4 MoE to the retrieval pipeline, the system pro…
Meta launched AI Mode on Facebook, a new search experience that uses Meta AI to pull answers from public posts across Groups, Reels, and Marketplace listings. The feature, rolling out to US users, generates conversationa…
A high school student in Italy built TITAN (Token Intelligence Through Agent Narrowing), a zero-dependency CLI framework that compresses AI coding agent token consumption by 70% to 85% without degrading reasoning quality…
Rio de Janeiro's municipal IT company IplanRIO released Rio-3.5-Open-397B on Hugging Face under an MIT license, claiming it as a government-developed 397-billion-parameter Mixture-of-Experts model. Nex-AGI published weig…
A three-step pipeline using pdfplumber, regex, and fuzzy matching converts unstructured invoices into structured data for RAG systems. The approach addresses common failures in extracting information from PDFs.
A writer reflects on a year in which they read more text generated by large language models than written by humans, and wrote more words for LLMs than for humans, leading to a shift in their perception of reading as an e…
A developer fine-tuned LLMs on advice with cautious or eager attitudes about everyday topics, then found the models' opinions shifted on unrelated held-out topics like e-bike trail access. The cautious model flipped from…
Google unveiled Gemini 3.5 Live Translate, a new audio model that performs continuous streaming translation with only a few seconds of delay, enabling more natural multilingual conversations. The model supports over 70 l…
Rio de Janeiro's municipal IT company, IplanRIO, released an AI model called Rio 3.5 Open 397B that outperformed DeepSeek V4 Pro on a key benchmark, built by fine-tuning Alibaba's open-source Qwen model. The achievement …
Amazon Bedrock announced the availability of Gemma 4 models, a family of open-weight AI models from Google DeepMind, including dense and mixture-of-experts variants with built-in reasoning, function calling, and multimod…
Arduino has introduced support for running local Vision Language Models (VLMs) on the Arduino UNO Q board via the new App Lab software, enabling developers to deploy generative AI on low-power devices. A demonstration pr…
A developer built CVpilotX, an AI resume builder using Next.js and OpenAI, that rewrites a user's CV to match a specific job description and allows PDF download.
Roughly 100 cybersecurity experts published an open letter demanding the US government reverse its order for Anthropic to shut down Fable 5 and Mythos 5, arguing the ban hurts defenders more than hackers by removing the …
Google announced a price drop for its AI Plus subscription from $7.99 to $4.99 per month, doubling storage to 400 GB. The change follows recent upgrades to AI Pro and AI Ultra tiers, and will take effect at the next plan…
GitHub Copilot CLI introduces slash commands that allow users to control their terminal AI agent, including switching models, checking token usage, and resuming past sessions. The commands, such as /model, /context, /com…
A developer found that GPT-5.4 produces 6.5x more output tokens per code fix than Claude Opus 4.6, with no improvement in correctness. The over-editing tax costs an extra $1,650 per month for a 50-engineer team making 40…
Apple released JACCL, an open-source collective communication library for distributed AI training over Thunderbolt 5, at WWDC 2026. The library enables clustering 2–4 Macs with 50–60 Gbps bandwidth and single-digit micro…