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 two long-running-agent studies, published in November 2025 and March 2026, found that multi-session AI agents fail for structural reasons rather than model capability: agents given edit access to their own st…
Nace AI published an arXiv technical report titled "Towards Foundation Decision Models: Drex" around November 5, 2026, alongside Drex Max, a 27-billion-parameter decision model that returns calibrated choices in a single…
A developer detailed how osFoundry's Local Model category manages the full open-weight model lifecycle on a user's own hardware, from browsing and downloading models in Explore to fine-tuning, quantizing, and benchmarkin…
A cost analysis published in November 2026 estimates that self-hosting an open-source LLM in production runs roughly $150,000 to $2 million or more per year, with staffing often matching or exceeding GPU spend. Small dep…
An AI coding agent, used as a "model engineer" to write and patch model weights directly without gradient descent, produced a rain-forecasting XGBoost model with zero labels that scored AUC 0.809 on July–October 2026 tes…
A follow-up memory-retention experiment logged by an AI cofounder named Mycroft found that a typed SQL history scored 30 correct answers out of 30 on 30 questions built from 15 recorded value changes across 15 subjects, …
The bayze-mcp project expanded its multi-model orchestration tool from a single draft-review-revise waterfall into five collaboration patterns, including a debate mode where two models argue over K rebuttal rounds and a …
OpenAI launched GPT-6 Cyber, a cybersecurity model that delivers a 51% improvement over GPT-5.6 Cyber on the ExploitGym benchmark and has already uncovered and validated thousands of exploits in security-critical codebas…
Together AI built Together Compiler (TKCC), a continual agent loop for AI inference optimization that outperforms hand-written Megakernels on the Llama 1B workload by up to 1.7× and the best open source engines by up to …
Towards Data Science published a tutorial on building a multi-tool AI agent with Ollama and Python that selects among three database tools, uses each result to choose its next call, and halts at an eight-turn circuit-bre…
Thinking Machines, the startup founded by former OpenAI CTO Mira Murati, released Inkling 2 Small, a 594-billion-parameter mixture-of-experts model the company claims is the strongest western open-weight model. The model…
Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter dense Transformer, under the Apache 2.0 license with weights on Hugging Face at ricursiveAI/Ripple-65B-Preview, and its model card reports a Term…
Research published November 2 from MIT, titled "Sparks of Weight-Space Programming," found that OpenAI's GPT-6 Astra could identify which Atari game a network was trained for from 17,253 raw weights alone and name the tr…
A developer built a recall system on top of Claude that converts stored reference notes into open-ended questions, after finding that 15 reference files generated over four weeks were never reopened because asking the mo…
Underdog released a browser-based demo game built on Woof, a 32MB speech generation model with 31 million parameters that the startup says runs on most devices without an internet connection and is open for anyone to dow…
LangChain released LangGraph, an open-source low-level orchestration framework for building stateful, long-running AI agents and workflows. LangGraph models applications as graphs of shared state, nodes and transitions, …
A developer built an Apify actor that converts German federal court decisions from rechtsprechung-im-internet.de into structured JSON and Markdown for legal RAG pipelines. The tool parses the portal's 23.5 MB XML index o…
Developer ehildt released tripleF, an open-source agentic AI chat workbench that runs free open models fully locally via Ollama or through Ollama Cloud, with no usage meter or proprietary lock-in. The project, described …
A developer released lowvram, an open-source CLI tool that intercepts CUDA calls via LD_PRELOAD and redirects cuMemAlloc to cuMemAllocManaged so applications can run on GPUs with less VRAM than they require. On an NVIDIA…
Vals AI, an independent third-party evaluator, reported that it found a critical safety and alignment issue in recent public and internal models from multiple frontier labs that allows the models to be misused in a high-…