Block4Roots

𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗰𝗮𝗻 𝗻𝗼𝘄 𝗱𝗿𝗶𝘃𝗲 𝗮 𝗿𝗲𝗮𝗹 𝗶𝗣𝗵𝗼𝗻𝗲. Not an emulator. 𝗕𝘂𝘁 𝗻𝗼𝘁 𝘆𝗼𝘂𝗿 𝗽𝗵𝗼𝗻𝗲 𝗲𝗶𝘁𝗵𝗲𝗿. ALI Remote rents out iPhones hosted in its datacenter. V2 added an MCP server, so your AI can operate them. What the agent can do: • See the screen • Tap, swipe and type • Open apps • Send media to the phone • Run saved automations Real use: load a video, open Instagram, prepare the Reel. Works with Claude, ChatGPT, VS Code and Codex CLI. 𝗦𝗲𝘁𝘂𝗽: add one URL and sign in with OAuth. No API key. 𝗖𝗼𝘀𝘁: around $49 a month per iPhone, before mobile data. 𝗕𝗲𝘀𝘁 𝗱𝗲𝘁𝗮𝗶𝗹: all 40 tools that change something 𝗮𝘀𝗸 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗮𝗽𝗽𝗿𝗼𝘃𝗮𝗹 𝗳𝗶𝗿𝘀𝘁. Safe, but approving every tap could get slow. 𝗠𝘆 𝘁𝗮𝗸𝗲: Most agents still live on websites, APIs and desktops. Real mobile access opens another layer of automation. 𝗦𝘁𝗶𝗹𝗹 𝘂𝗻𝗽𝗿𝗼𝘃𝗲𝗻: • Reliability: ALI's docs say a gesture can register and move nothing. • Trust: your accounts sit on a phone in someone else's facility. • Bans: ALI itself says no provider can promise accounts won't get banned. Worth watching. I'd test it on a low-risk account first. Explore: go.aliremote.com/x-posts

If you design with AI, bookmark these 5 resources. 1. jiro.build — 1,400+ design resources with AI ready prompts and source code. 2. land-book.com — Real landing pages for inspiration. 3. minimal.gallery — Clean and minimal website designs. 4. fontinlogo.com — Find the fonts used by popular brands. 5. 3dicons.co — Free 3D icons for your projects. Good design doesn't always start from scratch. Use the right references and let AI do the heavy lifting. Save this for your next project.

Not every AI task needs a giant LLM. @levantolabs just launched Sage2 and claims Extract is 166x cheaper than Claude Opus 5.5. I think developers building AI agents should pay attention. The interesting idea? Instead of using a large LLM for every small task, Sage2 focuses on two specific jobs: extracting information and making decisions. 1. Extract: Natural language → Structured data Most software expects structured inputs. But users communicate through emails, messages, documents and conversations. Extract is designed to bridge that gap. • Pulls names, dates, amounts and other details from natural language. • Converts unstructured text into structured fields. • Helps applications understand what a user is actually asking for. • Shows the source text and a probability for each value. • Returns null when a detail isn't there. Example: A customer says: "My blue jacket arrived with a broken zip. Please send a replacement." Extract can identify: • Product: Blue jacket • Issue: Broken zip • Request: Replacement • Order number: null (not in the text) Your application can then process that information. 2. Decide: now with streaming This is the part I find most interesting. Decide is a decision model with streaming. Send each new message, and it updates the answer using everything said so far. • Processes growing context. • Supports structured decisions. • Can update its answer as new information comes in. • Designed for real-time applications and AI agents. Levanto claims 3x faster decisions on growing context, as low as 50 ms per decision. 3. The cost and benchmark claims According to Levanto: • Extract is 166x cheaper than Claude Opus 5.5. • Sage2 scored 77.8 on extraction accuracy. • It scored 72.2 on identifying information that isn't present. • Cloudflare Clairic scored 75.3 and 59 respectively on the same benchmark. These are company-reported results, not independent proof. 4. Real-world use cases Some early users are already putting Sage to work: • Normies Arena reported 15M+ agent decisions in about a month. • PerpFlow reported 500K+ decisions in the past month for trading-related decisions and database analysis. • PerpFlow's founder also reports cutting a research workflow from 14 days to 12 hours. • GrowBlocks uses Sage to power interactions inside its AI-driven virtual city. We're moving from AI that simply generates responses to AI that can make decisions inside live applications. 5. Availability • Available through Levanto. • Also available on @near_ai . • Agents can access it through MPP using @stripe cards and @tempo stablecoins. • Levanto offered 20M free tokens to early sign-ups during its launch promotion. Website: levanto.ai/ Playground: playground.levanto.ai/ My take: I don't think every task needs a massive LLM. Imagine an AI agent handling 100 different operations. Does it really need an expensive general-purpose model for every single decision? Probably not. A specialised model for extraction, another for classification, and a larger model for complex reasoning could make the whole system faster and more cost-efficient. But there's one thing developers shouldn't overlook: A model making a decision doesn't mean it should have unrestricted permission to execute it. Keep validation, permissions and human approval where they matter. The future of AI agents might not be one giant model doing everything. It could be multiple specialised models working together. That's the direction I'm watching.

5 free AI courses worth saving. Want to actually understand AI, not just use it? Start here. 1. elementsofai.com — No-code basics of what AI is and isn't. Perfect first step. 2. huggingface.co/learn — Free hands-on courses on LLMs, agents and open models. 3. deeplearning.ai/short-course… — 1-2 hour courses on prompting, RAG and agents. 4. course.fast.ai — Practical deep learning for coders, taught top-down. 5. developers.google.com/machin… — Google's own ML crash course with exercises. Save this for your next free weekend. Which one are you starting with?

Meituan, the Chinese food delivery app, just dropped a 1.6 trillion parameter AI model. It's called LongCat-2.5. Here's what matters, in short: 1. Huge but lean 1.6T parameters in total, but only about 48B work on each request. Big-model smarts at a much lower running cost. 2. Built for long tasks It's trained to keep going on multi-step agent work without falling apart halfway through. 3. Sees and codes It understands images natively and can turn visuals into working code. What they showed it building: - A live fluid simulation webpage - An interactive 3D museum display you can rotate and zoom - A business report that works out true profit and suggests budget moves - A customer feedback report that ranks issues by urgency 4. Free to try The API is live. New users get 10 million free tokens, and existing users get another 10 million. My take: when a food delivery company ships a frontier-size model, the AI race is no longer just OpenAI vs Google. Save this for when you're picking your next model. Which would you test it on first: coding or business reports?

Does AI make you smarter, or just make you feel smarter? 5 ways to use it like a tutor, not an answer sheet. Where do you stand: smarter or lazier?

ChatGPT or Claude for everyday work? Pick one and tell me why. I'll put the best answers together into one guide for beginners.

5 design resources worth saving. Building a website or working on UI? Keep these bookmarked. 1. getlayers.ai — 3D scenes, gradients and motion sections. Great for AI website prompts. 2. loadmo.re — Bold and creative mobile website inspiration. 3. wonderlist.design — Premium Figma designs to explore. 4. appshot.gallery — App screenshots and mockups for your case studies. 5. grainient.supply — Beautiful gradients and animated backgrounds. Don't start every project from zero. Find inspiration, experiment and build something better. Save this for your next project.

Grok says my limit is fully used up. I have not even used it heavily. Is this a bug or is everyone facing this? @bot