Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer at an ad analytics SaaS discovered that a missing tenant ID in an MCP router cache key caused Cloudflare Workers to serve cached Vectorize results from one advertiser to another, leading to a $187/month cost …
Google announced at I/O 2026 that Gemini in Chrome will soon support WebMCP APIs, making it the first mainstream agent to consume WebMCP tools. This shifts the risk calculus for developers: instead of instrumenting for a…
Anthropic, the AI safety company valued at nearly $1 trillion, believes it must lead the AI race to ensure safe development, according to former employees and analysts. The company's strategy involves building cutting-ed…
The White House has imposed a government checkpoint on OpenAI's GPT-5.6, requiring federal approval for early access users, marking a shift from voluntary AI safety commitments to direct government intervention. The move…
Google is requiring news publishers in its AI-powered article overviews pilot to grant broad rights for content use, including AI training, or risk losing annual fees from Google News, which Google plans to end. The Euro…
A developer built a sleep apnea detection pipeline using OpenAI Whisper and Librosa, combining raw signal processing with transformer-based audio analysis to identify respiratory distress patterns from snoring sounds cap…
A developer outlines five prompt engineering techniques—zero-shot, few-shot, chain-of-thought, and instruction prompting—to help engineers understand and maintain legacy projects with minimal documentation. The technique…
OpenCode v1.17.10, released June 24, adds MCP resources and resource templates, fixes OAuth callback failures in WSL and container setups, and introduces a --mini CLI mode for scripting. The update closes long-standing G…
A developer recommends Unsloth as the most cost-effective method for fine-tuning small language models in 2026, citing its ease of use and low VRAM requirements compared to the theoretically cheaper but impractical TPU-b…
Microsoft found that AI coding agents ignore documentation tips because they form plans before reading. To redirect agents, developers must explicitly state that the agent's current plan will fail, not just describe the …
OpenAI reported that 97.9% of its employees now use its Codex AI coding agent, with the tool accounting for 99.8% of internal AI output tokens, signaling a rapid shift from conversational AI to agentic workflows. The com…
Retriever, a browser-agent startup, cut the cost of automated web workflows by over 100x by swapping its planning model from a frontier API to DeepSeek V4 Flash, an openly licensed Chinese model. A multi-step workflow th…
Tsvetan Stoychev, a former principal software engineer at Akamai Technologies, used AI tools including GitHub Copilot, Claude Opus, and Gemini to find vulnerabilities in the ClickHouse codebase, successfully reporting re…
An AI agent is not a new type of model but rather a small amount of plumbing around an LLM, consisting of Model, Instructions, Memory, Tools, and an Execution Loop. A minimal agent can be built from scratch in 60 lines o…
PostHog, Langfuse, and seven other AI observability tools are compared for developers building LLM-powered applications. These tools provide tracing, cost tracking, evaluations, and prompt management to debug and improve…
AI inference is profitable, with frontier providers reporting 70%-80% gross margins, contradicting claims that it is unprofitable. Cost estimates show inference costs around $1 per million tokens, while API pricing range…
Claude Code, an agentic coding tool from Anthropic, excels at autonomous multi-file tasks when used with proper planning and review, but fails when treated as an inline editor like Cursor. Users must delegate tasks via t…
Perplexity, an answer engine that synthesizes web search results into cited answers, offers free and paid tiers with features like Pro Search and Research mode. While useful for research, its citations are not always ver…
Rudrite Research released a 47-minute documentary explaining how modern text-to-speech works, from sound waves as numbers to neural audio codecs and talking LLMs. The film covers key milestones like WaveNet, Tacotron, VA…
A developer built Squish, a local LLM inference server for Apple Silicon, after finding existing tools like Ollama too slow for generating git commit messages. Squish achieves up to 9.8× faster performance on long prompt…