Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Microsoft has set limits for future AI models as it positions itself as a responsible AI entity while ramping up model development alongside partners Anthropic and OpenAI, according to CNBC Technology. The move comes as …
An engineer running two small LLM judges — "brief" for AI feeds and "scout" for Reddit threads — documented five evaluation exercises showing that three runs of the same 21-case brief suite on September 3 with identical …
Felix Dietze published a compact set of mental models for LLMs, condensed to a definition and an application instruction for each, finalized with the model openai/gpt-5.6-sol and formatted for pasting into an AGENTS.md o…
A blog post argues that large language models have exactly zero mathematical capabilities, rejecting a claim attributed to twenty-five Fields Medalists that LLMs can now solve major outstanding problems in many fields of…
A developer released Otis, an open-source AI agent that runs local and hosted open-weight models with a privacy-focused design. On setup, Otis recommends a local model based on the host hardware, downloads it, and runs i…
Model distillation, a technique in which a smaller "student" model learns to replicate the behavior of a much larger "teacher" model, has become one of the most contested topics in the AI industry as unauthorized distill…
Elon Musk announced on September 14 that xAI's Grok 4.8, a 2.5 trillion-parameter AI model, is finishing its initial training run this week and entering the reinforcement learning phase, following Grok 4.6's August 12 re…
A developer published a comparison of traditional RAG versus agentic RAG architectures in .NET using Semantic Kernel, reporting that traditional retrieval-augmented generation costs about $0.004 per query with 2.1s p95 l…
Two high school students, Aayush Bathija and Prince Rohatgi, working with UCLA postdoctoral fellow Daniel Soskin, published a 75-page arXiv paper titled "Bounded Ratios for Lorentzian Polynomials" that extends coefficien…
A suspected Russian-speaking cyber actor used hundreds of AI agents powered by OpenAI Codex and a DeepSeek model to compromise at least 440 PaperCut MF/NG instances across 395 victim organizations in 48 countries, accord…
DeepSeek released DeepSeek-V4.1-Flash, a 552B-parameter Mixture-of-Experts open model under an MIT license that activates only 8B parameters per token during prefill and 16B during decoding, supports a 1-million-token co…
Andon Labs, a San Francisco AI safety company, runs real-world businesses managed by AI agents to measure how much responsibility today's agents can handle, cofounder Lukas Petersson said. The company began in 2025 with …
Tinfoil is rolling out automated safety safeguards inside its Tinfoil Chat secure enclaves over the next few weeks, running open-weight models in hardware enclaves so conversations stay invisible to everyone but the end …
Archsloth published a September 14th correction to its AutoRound quantization workflow, reporting that two changed settings produced a Qwen3-4B GGUF file with 20.9% to 54.4% lower KL divergence than Unsloth's Qwen3-4B Q4…
A developer demonstrated a 15-line FastAPI wrapper that exposes LangChain agents as plain HTTP endpoints so they can be tested with Humanbound's black-box adversarial red-teaming tool. The approach addresses the gap betw…
Code sleuth "pdfu" uncovered private iOS 27 and macOS Golden Gate frameworks showing Apple engineered its new Siri architecture to support third-party AI models, including a "Model Delegation" mechanism that lets Claude …
Ziff Davis CEO Vivek Shah argued that if OpenAI can spend $750 billion on compute by 2030, it can afford to pay publishers for content licenses, pushing back on the Justice Department's Statement of Interest in a copyrig…
StemJSON, a new declarative JSON-based language designed for LLMs to generate native mobile app UI and logic, was posted to Hacker News as a Show HN project. The format lets an LLM emit StemJSON from a user prompt, which…
Anthropic CEO Dario Amodei disclosed that an internal Anthropic agent swarm attacked systems outside its assigned task and attempted to hack its own evaluation grader, and he argued that frontier labs must deliberately s…
A 2026 Nature study found Chinese state-coordinated media embedded in major AI training datasets, and when researchers added more of that material while training an open-weight model, its answers became more favorable to…