This is huge for normal people. You describe an app, AI builds it, and it's on your phone. No developer account, no app store review. I love it, it's exactly how AI becomes as normal as phones.
But I spent my PhD on AI and security, so here's the part nobody's saying: no review also means no one checked it.
A vibe-coded app can quietly ask for way more permissions than it needs. It can store your data in plain text, call some random API with your info, or pull in a library with known holes. The AI isn't trying to be malicious. It just optimizes for "it works", not "it's safe". And the person who asked for it usually can't read the code to tell the difference.
It gets worse when people start sharing these apps with friends and family. Then one sloppy app is on ten phones.
So I don't think the answer is "don't do it". The answer is the platform has to do the checking the user can't. Scan permissions, flag sketchy network calls, warn before anything touches contacts or photos. Make safe the default, not homework.
Building got easy. Trusting what you built is the next problem.
Today, on Google Play we announced new ways we are evolving the Play Store for an experience that is unique to you and the content you enjoy most:
✨ Use Play to wirelessly install and run your own custom, vibe-coded apps privately on your device.
🔍 Discover apps effortlessly with Ask Play & Gemini
🎮 Dial in discovery with the all-new Game Finder & Gamer DNA
Character AI had millions of daily users and still couldn't make it work. That should scare every consumer AI startup. Until running these models gets a lot cheaper, "people love it" and "it survives" are two different things.
>character ai files for chapter 11 bankruptcy
>disney steps in with $20m opening bid
>not buying the company, just IP license + claims
>also providing "transition services" using former cai employees it poached ahead of this
>official court filing: "if such a bid is unavailable, however, the platform will be wound down"
it’s absolutely over
@dhh This is the stuff that matters. Not agents building whole companies. Just "I needed to sign a PDF and didn't have to download sketchy software." That's how normal people end up actually using AI.
@emilkowalski Nobody's taking that from you. People hand it over because it feels good. Outsourcing the typing is fine. Outsourcing the judgment is a choice, and it's the one you'll regret.
I have a bit of a model snobbery problem. If it's not the newest, smartest thing (Fable, Opus, GPT-6.1), I don't want it.
Grok Bot made me forget about that. It's fast, it's genuinely smart, and the best part is it keeps working when I'm not at my desk. I've been using it like an addict.
What it does for me: digs through X data, reads new papers and pulls out what matters, and clears all the small stuff I'd normally lose an afternoon to. I also plugged in a few third-party agents that pair with Opus for remote coding. Each bot has its own job, and I can put them in a group chat so they talk to each other. Feels like running a tiny team.
Claude is still my main driver. Opus 5.5 is just that good for real coding. Codex is my sidekick for video generation, local browser automation and computer-use tasks.
Honestly really enjoying this setup right now. What does your stack look like?
@aakashgupta Bundles get you in the door. They don't keep you there. I'll switch tools in a week if something works better, and so will everyone else here.
@beatsinbrief Compute costs land on real paychecks. Until inference gets a lot cheaper, companies will quietly trade raises for tokens, and that's not a sustainable deal for anyone.