Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Save, a new API that converts any URL into clean Markdown optimized for large language models, launched with a free tier of 1,000 pages per month. The service fetches web pages server-side, optionally renders JavaScript,…
A developer warns that language models cannot distinguish between data and commands, making AI agents vulnerable to prompt injection attacks via web pages, tool outputs, or supply chain components. The developer argues t…
A developer successfully ran a 35-billion-parameter Mixture-of-Experts model on a 2017 AMD RX 580 8GB GPU using Vulkan, bypassing CUDA and ROCm. The project, called Polaris Revival, achieved 17-18 tokens per second for L…
Bluffbench, an evaluation for LLM plot interpretation, is nearing saturation as models like Fable 5 achieve high scores on the hardest 'mocked' cases, though human-level performance remains elusive. The benchmark will so…
A new scheduling technique called TurboPrefill reduces waiting time for Vision Language Models by nearly half, from 9.0 to 4.6 seconds, without changing model weights or architecture. The optimization, validated on Qwen2…
A new class of post-trained AI security models, such as ArgusRed and PentestGPT, is enabling automated penetration testing by bypassing standard safety refusals and executing exploits in sandboxed environments. These sys…
DeusData released codebase-memory-mcp, an open-source MCP server that indexes codebases into a persistent knowledge graph, enabling AI coding agents to answer structural queries in under a millisecond instead of grepping…
A non-coder used Claude to generate and deploy websites to clients via a single prompt, leveraging Cloudflare for production deployment. The user claims this workflow, requiring only 1-2 prompts, signals a major shift fo…
A developer released EGC (Extended Global Context), a local runtime that gives AI coding tools persistent memory across sessions. EGC saves decisions, failures, and preferences at the end of each session and loads them a…
OpenAI's Codex desktop app on the ChatGPT Plus plan has experienced a 10-20x increase in rate-limit cost per token since June 16, reducing usable prompts from 20+ to 2-3 per 5-hour budget. User logs show smaller prompts …
Claude generated Z3/Python code to solve a chess puzzle placing all pieces on a 6x5 board with bishops on opposite colors and no attacks. The solver found 192 raw solutions, deduplicated to 24 unique arrangements after a…
An XDA Developers columnist reports using Claude as a conversational collaborator to transform his creative workflow, starting with dialogue instead of design tools to surface ideas, reframe problems, and refine concepts…
A Hacker News Ask HN thread revealed that developers' choice between Claude Code and Codex is driven primarily by organizational procurement and interface preferences rather than model performance. Commenters reported us…
Signal President Meredith Whittaker warned that AI chatbots are not friends or conscious beings, cautioning against treating them as sentient interlocutors. She expressed concerns about granting systems like Microsoft Co…
Alloy, a new PyTorch backend and inference engine for Apple Silicon, has been released as a technical preview. The open-source project compiles Python GPU kernels to Metal and supports LLM serving with a drop-in compatib…
Chrome now ships Gemini Nano, a language model accessible via the Prompt API, enabling on-device AI without API keys or inference costs. A developer integrated it into a free Mermaid diagram editor, using the model as a …
Security researchers have released a post-trained LLM specifically for penetration testing, which reportedly found thousands of real zero-days. The model, developed by the Argus Red team, treats offensive security as the…
A developer spent two weeks optimizing a homelab with four RTX 3090s (96GB VRAM) for local LLM inference, achieving improvements like 40% throughput gain and 4x VRAM savings, but ultimately found that paid APIs were more…
A developer open-sourced NRT-Defense v0.4.0, an adaptive multi-turn defense framework for LLM agent teams that reduces attack success rates to under 1%. The framework addresses vulnerabilities exposed by Lee et al. (2026…
A developer on Hacker News asks peers to share their most practical lesson or 'aha' moment from coding with AI, citing a personal workflow shift and curiosity about prompting strategies, tech stack changes, and trust cal…