Discover Amazing Content, Share Life Moments
Connect Our Wonderful World
Meta brings Pocket to the US: say a phrase to your phone and generate a playable game
After a quiet test in Brazil, Meta's AI-powered game app Pocket is now officially available to US users. Just enter a prompt and you can generate a mini-game that responds to touch, tilt, sound, and even photos, and you can scroll through other people's creations in a feed.
Ramp also built an AI model router, and it's called Router
Corporate spend management platform Ramp has launched its own AI model routing service, Router, supporting OpenAI, Anthropic, DeepSeek and other models. It is free until the end of 2026 and comes with a $26 credit.
A Third of New Webpages Are Written by AI Since ChatGPT’s Launch
A new Pew Research Center study finds that more than a third of new webpages published since ChatGPT’s launch show signs of AI authorship. Meanwhile, bot traffic has surpassed human traffic. The internet is becoming something written by bots, for bots.
Video to 4K, FLUX Chooses to Redraw Every Frame
FLUX Upscale is officially live, upscaling videos from 480p to native 4K with two modes: Precise and Creative. Optimized for FLUX 3 generated content, it fixes blurry faces and grid artifacts, making it ideal for broadcast and advertising delivery.
Are large models just guessing when doing cybersecurity work? This open source project gives AI agents 817 operation manuals
Anthropic-Cybersecurity-Skills, the open source project with 29.6k stars on GitHub, provides 817 structured cybersecurity skills for AI agents, covering 26 security domains and mapped to frameworks including MITRE ATT&CK. Each skill is a Markdown workflow that agents can follow step-by-step, putting an end to AI "confidently making up nonsense".
The Biggest Waste in RAG: Repeatedly Querying Static Data Over and Over
Every RAG query hits the vector database, even if the documents haven't changed in three months. CAG caches static knowledge in KV memory, but prefix caching has a hard limit of byte-level matching. LMCache solves this with a separate process and CacheBlend: 14x faster first token, 2-4x faster multi-document queries, and it's open source.
Can See Screens, Use Files, and Has a Skill Library: Claude's Agent Is Starting to Look Like a Real Employee
Computer Use, Browser Tool, Skills API, and Files API have reached general availability on the Claude Platform. The agent can operate software without built-in APIs, call team-specific skills by version, and reuse uploaded files. Community developers summarize that round trips have been reduced by 20-40%. One netizen commented that the API surface is increasingly looking like a full-fledged employee.
OceanLight: Unstructured Grids Cut 70% Computing Power for More Accurate Ocean Forecasting
OceanLight replaces uniform structured grids with geometrically adaptive unstructured grids, and combines this approach with graph neural networks. It reduces GPU memory usage by 62% and FLOPs by 70% while maintaining prediction accuracy.
Skip one drama, spend 2 hours and 34 minutes understanding how large language models are built
A Stanford course that walks you through everything from tokenization and BPE to Transformer, RLHF and DPO, fully breaking down how large models like ChatGPT are built. It's suitable for AI beginners and developers currently working on Agents alike.
Observation of the Pi Plugin Ecosystem: The Minimalist Assistant Has Been Built Into A Multi-Agent Legion, But Don’t Rush To Install Everything
Over the past two days, a number of high-quality extensions for Pi have emerged: multi-Agent orchestration, web search integration, Cursor bridging. The ecosystem is booming, but a popular plugin has been pointed out to still have a hardcoded model issue unfixed. Check the open issues before installing.