Did you know you can now run an AI model directly on your phone, no internet required? 🤯 IBM’s Granite Nano models make it possible, built to be small and run locally on phones or Raspberry Pis. In this quick demo, Emma Gauthier shows how to: 1️⃣ Download a local AI app 2️⃣ Import a Granite Nano model 3️⃣ Run it fully offline
Wich app did you use / recommand ?
𝗜’𝘃𝗲 𝗯𝗲𝗲𝗻 𝗹𝗶𝘃𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝘁𝗵𝗶𝘀 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝗳𝗼𝗿 𝘆𝗲𝗮𝗿𝘀. Every time I want to jot down a note, a task, or an idea - I end up not doing it. Why? Because it always turns into too many small steps: grab the phone → open the app → find the right section → type the thought → add a label. And when I think about all that, I just tell myself I’ll remember it later. Spoiler: I never do. So I decided to fix this by building an automation that does everything with 𝗼𝗻𝗲 𝗴𝗲𝘀𝘁𝘂𝗿𝗲 - using voice, AI, and n8n. Today I’ve got the MVP working: 🎙 Voice note via iPhone Shortcut → 🧠 AI agent turns speech into text and analyzes it → ✅ Task is automatically created in Todoist. I’ll show the full process soon — step by step. Follow if you want to build something similar. And yes, I’ll share the project on GitHub too.
It’s not the year of agents — it’s the decade. Everyone’s hyping 2025 as the “Year of AI Agents.” But Andrej Karpathy nailed it: This is the Decade of Agents. Why? Because building agents that act like real interns or teammates isn’t a weekend hack. It takes serious breakthroughs in: → Memory → Multimodal reasoning → Long-horizon planning → Cognitive architecture Yes — tools like Devin, Claude, and AutoGPT feel magical. But they’re still: • Fragile • Narrow • Unreliable at scale We’re still in the pre-iPhone era of agents. So what does that mean for builders? 👉 Massive long-term leverage for anyone working on: • Infra – caching (what we build at Alchymos) • Interfaces – voice, UI, autonomy • Real-world agent deployment – MCP, SDKs, APIs This is a decade-long game. But the stack is being built now. And the people who build it today… will define how agents run tomorrow. Curious to see caching for AI agents in action? Shoot me a DM
Almost every AI work starts with a shiny demo, but being in a proof-of-concept loop will prevent you from actually doing meaningful work. Today, Stefan Krawczyk and Hugo Bowne-Anderson share their secret for overcoming the proof-of-concept; the next step for separating toy applications from real problem-solving apps. First, identify the challenges in your application and take the best approach to solve them. 👇
New feature: run AI-powered queries on boards 🔥 Not quite NotebookLM, but cards can be updated you, your friends, coworkers manually, or by AI (API coming soon!). factoids.app cards can be added to boards (grouped by panel). Panels can be clipped from one board to anther. Queries that run on boards will include all cards. See demonstration in video attached. Have a nice weekend, Henry
From slow-motion horse jumps to slow application performance—we all want to see things clearly! 💡 Being able to capture and diagnose slow transactions in your apps is crucial. Learn how to simplify your observability stack and leverage new AI Insights for faster debugging in this great blog post: https://lnkd.in/g75Waz-7
Big day for the Impart team. We’ve launched Impart AI, become a member of the NVIDIA Inception program, and moved to our new home at impart.ai. Each move reflects one goal: give security teams the same velocity, visibility, and control developers already have. Impart AI is the intelligent workspace for designing and deploying runtime protection across apps, APIs, and AI interfaces. It allows teams to build and adapt protections conversationally and safely. Learn more about how we’re building AI the right way 👉 https://lnkd.in/e3zV2itf
I've written a Chrome extension that runs text-to-speech AI inference in your browser using WebAssembly and WebGPU: https://lnkd.in/gg9UMC5E This is one of the only TTS extensions offering AI voices for free - all others paywall AI voices or only offer robotic non-neural voices. The extension uses WebGPU to inference an AI model fully in your browser, achieving much higher quality without charging for GPU server usage and keeping your data on-device. I added a few quality of life features including text highlighting and click-to-jump It uses the # 1 open-source TTS model Kokoro-82M https://lnkd.in/gSsnzzZC and its WebGPU implementation kokoro-js, and it runs more than 4x realtime on my M2 Macbook GPU. The extension is also fully open-source: https://lnkd.in/gt87GQTi (stars appreciated). Thanks Yanan Jiang, Ayush Agrawal, Raj Pabari for beta testing
Missed ZapConnect? FIVE new launches you may have missed 👇 1️⃣ Unified Copilot: one AI interface for every app 2️⃣ Human-in-the-loop: approve critical steps with a click 3️⃣ Over 30 new AI apps (Perplexity, Mistral AI, Cursor, DeepSeek + more) 4️⃣ Agent Sharing: Build once, deploy to your whole org 5️⃣ Enterprise controls your IT team will actually like Plus, new bundled pricing to make everything simpler. ‼️ 𝗢𝗻𝗹𝘆 𝟮 𝗱𝗮𝘆𝘀 𝗹𝗲𝗳𝘁 𝘁𝗼 𝘄𝗮𝘁𝗰𝗵 ‼️ Catch the replays before they’re gone: https://lnkd.in/epq6K-Hj
Today we're releasing conversation simulation on Confident AI (YC W25)'s day 2️⃣ launch day. For teams that are stuck with single-turn evals for a multi-turn use case, simulations is usually the bottleneck. Now, in a click of a button you can: 1. 💬 Simulate thousands of user interactions with your LLM app 2. Run evaluations on all of them 3. Entire process takes <2 minutes Conversation simulations let you automate the end-to-end multi-turn evaluation workflow in one clean workflow. You can also connect your own custom LLM app endpoint if testing a prompt isn't enough. Stay tuned for day 3️⃣
Offline AI is great — now I can walk into a remote village and still ask, ‘Hey, do you guys have 5G here?’ 😇