1. Your PM finds a tiny bug. 🐛 2. Your designer wants to fix an empty state. 3. Marketing needs one line of copy changed. And somehow… All 3 end up waiting in the engineering queue. What if they didn't have to wait? Base Code by @base44 fixs it. ↳ Connect your existing GitHub repo ↳ Describe the change in plain English ↳ Preview it on the real product ↳ Send it back to engineering as a standard PR The important part: 🔒 Engineering still controls what ships. Every change happens on an isolated branch, and your existing review process + protections stay intact. So product, design, QA, or marketing can actually make the change — while engineering still decides what goes live. 🎁 30 days free right now. One invite covers the whole team. I also have a coupon for the first 50 users. DM me and I’ll send it over. 🔗 base44.com

I just found out US listeners spend 170 minutes a day on Pocket FM. For comparison, TikTok gets 53.8 minutes a day. That’s a wild gap for an audio-only app! Here's how they do it: - You listen to a few episodes free. - The story leaves you on a cliffhanger. - You buy coins to unlock what happens next. Every episode gives Pocket FM another chance to learn: ↳ Where listeners lose interest ↳ What keeps them listening ↳ When they’re willing to pay to continue Now it’s bringing those insights to writers. It’s called Sherpa, an AI writing partner built for storytellers. Pocket FM says it’s trained on 100M+ hours of listener data, including episode retention, coin spending, and drop-offs. As someone who creates content, that’s what caught my attention. Getting someone to start a story is hard. Keeping them interested through an entire season is a different challenge. Sherpa helps writers take a single premise and develop: ↳ Characters and their motivations ↳ Story arcs and episode structure ↳ Punchier dialogue and stronger conflict ↳ Cliffhangers that leave something unresolved Think of an idea like: "A delivery driver recognizes his missing sister in a billionaire's family portrait." Who took her? Why is she there? Does she remember him? There's a whole season hiding inside those three questions. Sherpa helps you develop it into something you can actually write from. If you've had an idea sitting in your notes for years… This is the week to open that file. Check it out today: pocketfm.com/sherpa

I use AI coding tools a lot, and I love them. They’ve made building apps faster and easier than ever. But there’s one thing I hate: Watching the token counter every time I give an agent a bigger task 😅 “How many credits is this going to burn?” I find that pretty limiting. Then I found out about Junie Local. It lets you run the full Junie coding agent directly on your Mac. → No tokens → No usage limits → Works offline → Your code, prompts & diffs stay on your machine Just run /local inside Junie. It downloads the model, starts the local server, and switches to local mode. It currently runs Qwen3.6-27B at 4-bit, optimized specifically for Junie. The most interesting part for me is unlimited usage. That means you don’t have to worry about burning credits on things like: → Refactoring multiple files → Adding missing tests → Dependency upgrades → Framework migrations → Exploring an unfamiliar codebase → Long, repetitive coding tasks You can just hand them to an agent without constantly thinking about credits. One catch: It currently requires an Apple Silicon M5 Mac with 64 GB RAM. Still, I really like where this is going. Local AI + unlimited usage is a pretty interesting combo for coding agents. If you have compatible hardware, try it yourself: junie.jetbrains.com/local

This is genuinely great news for developers. Junie can now run fully on your Mac with no tokens, no usage limits, and your code stays local. The hardware requirement is still pretty high, but unlimited local AI coding is a direction I’m very excited about. Love seeing JetBrains push local AI coding forward. I'll share more soon...

Prompt → generate → almost right → regenerate → something else changes. That loop is still one of the frustrating parts of AI video. You fix the character’s position, but the prop changes. You get the action right, but the camera geometry feels off. You fix the lighting, and suddenly the next shot doesn’t match. The problem isn’t always generation quality. It’s controlling all the relationships inside the scene at once. That’s what makes Rhapsody 1.0 ( @induceai ) interesting to me. It adds a reasoning and verification layer around video generation, so the system can consider what happened before, what should happen next, and whether the generated shot actually fits. Better video isn’t just about making prettier shots. It’s about having more control over what happens between them. @reachmaheshkmb Learn more: theprint.in/ani-press-releas…

I’ve been thinking about what’s actually limiting dexterous robot learning. It might not always be the model. It might be the data pipeline. X Square Robot’s TwinDEX takes an interesting approach: instead of putting the robot in the loop for every demonstration, a human demonstrates dexterous tasks through a wearable manipulation interface, while its twinned robot hand learns directly from those demonstrations. That sounds like a hardware detail, but it changes the economics of dexterous data collection. The system achieves comparable learning efficiency to on-robot teleoperation data, while collection throughput reaches 5.3 times that of on-robot teleoperation, without using any on-robot teleoperation data for training. What I find more interesting is the principle behind it: if the wearable interface and robot hand are designed as a matched system from the beginning, you may not need to solve the hardware mismatch entirely in software afterward. For dexterous robotics, that feels like a direction worth watching.

I like AI products that take something complicated and make the interface ridiculously simple. @viso_ai is a good example. Computer vision traditionally involves collecting footage, labeling data, training models, and a decent amount of ML expertise. With Viso Now, you upload a video and just describe what you want it to look for in plain English. That's the part I find interesting. Want to detect a safety risk? A quality issue? Something happening inside a warehouse or retail store? Describe it. Viso builds a working vision app around that request. No labeling or model training for every new use case. Underneath it is VGI-1, which is designed to reason about scenes it hasn't been specifically trained on. It basically makes computer vision feel more like prompting an LLM. And I think that's a pretty big UX shift for this space. It's free to start if you want to play with it: 👉 VISO.AI?utm_campaign=4883836… Let me know what you guys think!

AI-generated ads are about to hit a wall. People are getting tired of fake-looking AI content. And even when you can't tell it's AI… platforms increasingly can. AI video can carry invisible watermarks. Platforms are getting better at detecting + labeling generated content. So maybe the future of advertising isn't making AI humans look more real. Maybe it's just… using real humans. That's what I find interesting about Icon. It's not an AI UGC generator. Real creators actually film the ads. Icon uses technology to handle everything around them: → Find the creators → Write the scripts → Ship the product → Manage the shoot → Edit the footage → Handle revisions AI/software does the boring work. Humans stay in the part where being human actually matters. Funny how AI might end up making authenticity more valuable, not less.

This is the part of AI video I find most interesting: Not just better generation, but better generation becoming cheaper. Wan 3.0 on FlovaAI at $0.013/sec is another step in that direction.

AI made building easier. Now the bottleneck is taste. That’s why @OJOaidesign caught my attention. Instead of giving you one design copilot, it lets you build an entire team of specialized design agents, Working from product strategy → PRD → prototype → launch-ready design on the same editable canvas. Everyone is getting the ability to build. The winners will be the ones who build things people actually *want* to use. OJO is trying to turn that taste into a system. Comment OJO to get early access! #OJO #ojodesign

Vibe coding changed how we build software. Now a Swedish AI startup thinks the same thing is about to happen to the entire computer. They call it **Vibe Computing.** Instead of: → opening apps → clicking through menus → copying and pasting → switching between tabs → typing everything yourself You hold a key and just say what you want done. The AI handles the execution across your computer. That's what Incredible is building. And they just announced $2.5M in pre-seed funding, backed by investors with links to Anthropic, Lovable and Legora. The interesting part to me: We've spent decades learning how to operate computers. AI might flip that relationship. You tell the computer the outcome. The computer figures out the interface. Vibe coding might have just been the beginning. Try Incredible: incredible.one

Most AI video models are good at generating clips. Very few are good at understanding context. I tested MiniMax H3 by giving it more than just a prompt—images, audio, and a video reference through Omni Reference. That made a noticeable difference. Instead of guessing what I wanted, it had actual context to work with. The outputs felt more consistent, followed instructions better, and handled text/UI elements much more reliably than I expected. For creators making ads, product videos, gaming content, or social media videos, that extra control saves a lot of time. Definitely one worth trying if AI video is part of your workflow. hailuoai.video/ @Hailuo_AI #hailuoai #hailuo #MiniMaxH3 #H3