Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
RubyLLM, a unified Ruby framework for multiple AI providers, launched to simplify building chatbots, agents, and AI workflows. The framework supports OpenAI, Anthropic, Google, and others with a single interface, reducin…
OpenAI unveiled its first custom AI inference chip, Jalapeño, designed with Broadcom to power large language models like ChatGPT. The ASIC matches Nvidia's Blackwell and Google's TPU performance, reducing reliance on sca…
A developer identifies five common ways AI coding agents waste tokens, including prompt caching inefficiencies, long-context bloat, oversized tool schemas from MCP servers, and repeated context. The post offers specific …
An engineer replaced GUI-based AI coding tools like VS Code and Cursor with Anthropic's Claude Code, a terminal-based AI agent that integrates directly into the shell for faster development. By configuring aliases, auto-…
A developer released two open-source templates for Claude Code and OpenAI Codex that prevent AI coding agents from forgetting project context by storing rules and decisions in files and automatically injecting them via h…
A developer applied Specmatic's spec-first approach to TRIO, a multi-agent AI assistant built with FastAPI and React. Contract testing with OpenAPI 3.0 caught mismatches in response fields, status codes, and validation e…
OpenAI unveiled Jalapeño, its first custom AI chip designed with Broadcom for inference workloads, aiming to reduce dependence on Nvidia. The chip, built in nine months with AI-assisted design, will power OpenAI's models…
A former service industry worker with zero coding experience built two live AI agent systems in seven weeks using a Mac Mini and Anthropic's Claude API. The developer created a trading agent that makes real decisions and…
When AI coding agents encounter proprietary code unseen during training, they fall into a 'closest-match trap' by generating plausible but incorrect code based on similar public APIs. Baseline evaluations reveal the mode…
A developer explains why they run AI locally using Ollama instead of cloud-based ChatGPT for client work involving sensitive data. The approach keeps data on local machines, addressing compliance concerns for law firms, …
Sakana AI has launched a new API that encapsulates an entire multi-agent system, claiming it outperforms frontier models while eliminating the setup overhead that typically plagues multi-agent orchestration.
A new wave of long-horizon agent benchmarks, including Agents' Last Exam, SWE-Marathon, the Meta-Agent Challenge, and Arena's Agent Mode, reveals a fundamental trade-off between realism and verifiability, causing scores …
Vercel released AI SDK 7, a major update for building production agents in TypeScript, requiring Node.js 22 and ESM imports. The SDK introduces new primitives for agent development, runtime, integration, and observabilit…
A writer argues that concerns about AI-generated web content are overblown, noting that human-generated content is often inaccurate or inauthentic as well. The author contends that AI slop is not fundamentally different …
Giskard, an AI red-teaming platform, argues that evaluating AI security tools requires assessing whether they treat security and quality as inseparable and whether they are built for agents rather than models. The compan…
RapidThought Labs launched Heku, a single MCP server that dynamically serves tools via JSON configs, eliminating context bloat by lazy-loading only the tools an LLM needs. The open-source tool supports eight connector ty…
OpenAI and Broadcom unveiled 'Jalapeño,' a custom chip designed for large language model inference, claiming better performance per watt. Development took nine months with OpenAI's models aiding the process. Large-scale …
Cornell Tech researchers found that deep-research AI agents can be manipulated by short, 13-word edits to public user-generated pages, causing fake entities to appear in 38% to 62% of reports. The attack, called WARP, ex…
Parley, a new CLI and MCP server, enables local multi-model deliberation by fusing answers from multiple AI coding agents like Claude, Codex, and Gemini into a single response, providing consensus or flagging disagreemen…
Google launched the new Home Speaker, its first dedicated smart speaker in five years, integrating the Gemini assistant. The device features a soft-touch fabric finish, a base glow light indicator, and USB-C power, but t…