Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Apple confirmed at its Worldwide Developers Conference that its new Siri AI will run on Nvidia hardware inside Google's servers, marking a shift from its previous reliance on local device processing and Apple-controlled …
Developer Voodoo Tiki extracted OpenAI's adversarial review from the Codex CLI runtime and released it as a standalone tool, adversarial-review, enabling any model or CI pipeline to catch bugs by having a second model re…
Google announced at I/O 2025 the stable release of Android CLI v1.0 with new features like programmatic version lookup and direct integration with Android Studio. The company also expanded its repository of Android skill…
DeepSeek v4, an open-source model from a Chinese lab, achieved a 100x performance improvement on AMD hardware within 26 days of its Day 0 release, driven by the AMD SGLang engineering team. The InferenceX team tracked pe…
Software engineer Federico Pereiro argues that large language models (LLMs) combined with natural language are not an acceptable high-level language for building production software due to their unpredictability and comp…
Claude Code users can customize the CLI tool's status bar to display key information such as context window usage, session cost, Git branch, current model, effort level, and rate limit usage. The customization can be don…
The AI industry faces growing scrutiny over whether surging token usage and spending on large language models represent a speculative bubble or genuine value creation. Critics point to tokenmaxxing, subsidized pricing, a…
AI coding agents often prioritize legibility over capability, which can disadvantage Drupal despite its strengths for structured content and permissions. A field test with Claude Code ranked Drupal third behind Next.js a…
Enterprises rushing to adopt generative AI face an economic trap as token-based pricing models create hidden dependencies that will drive costs far higher once current market subsidies end. Tokens function as a metered t…
Local MCP servers enable individual developers to quickly connect AI clients to tools and data, but teams face significant friction when multiple members need to use the same setup. The local approach creates inconsisten…
The shift from prompt engineering to tool engineering is now required as AI workflows move beyond simple conversation into task execution. Prompt engineering alone fails when models need access to external data, APIs, or…
Firecrawl charges 5x more for LLM-based extraction than scraping, making it uneconomical for high-volume production pipelines. Experienced teams use LLMs to bootstrap CSS selectors once, then run deterministic code for s…
LLMs can perform tasks like writing code and summarizing documents, but developers face a choice between giving the model a "skill" to improve its internal behavior or connecting it to an MCP server for external data and…
President Donald Trump signed a new executive order on artificial intelligence that creates a voluntary 30-day government review process for "covered frontier models" before public release, allowing experts to probe for …
Anthropic's Claude Fable 5, a Mythos-class model, is now available on Vercel's AI Gateway, offering significant improvements in long-running, ambiguous, multi-step tasks that previously required frequent human oversight.…
Cognition introduced FrontierCode, a new benchmark that evaluates code on mergeability rather than just unit-test passing, with tasks built by open-source maintainers requiring over 40 hours each. The best model, Opus 4.…
Security researchers at StepSecurity have discovered the Hades Campaign, a sophisticated supply chain malware targeting Python developer environments that uses the Bun toolkit to execute multi-layer payloads capable of d…
Apple has released a detailed system prompt for Siri, outlining its identity as an intelligent assistant designed in California and providing extensive guidelines for handling entities, tools, and user interactions. The …
A new study from MIT and Wharton researchers Mert Demirer, Leon Musolff, and Liyuan Yang reveals that AI coding tools dramatically boost task-level velocity but produce far smaller gains in actual shipped software releas…
Anthropic reported elevated error rates on Claude Opus 4.8 (Fast Mode) on Tuesday, causing disruptions for users of the AI model. The company is providing email and SMS notifications for updates on the incident's status.