WebSockets are now GA on Vercel Functions. Connections still end at the function’s time limit, so don’t forget the reconnect logic. For a small realtime app, would you pick this or Cloudflare Durable Objects?vercel.com/changelog/websock…

If you build on Cloudflare, stop looking only at D1. I just found this out today. You can now spin up Postgres or MySQL right from the Cloudflare dashboard. It runs on PlanetScale and bills to your Cloudflare invoice, so there's no separate card to set up. Workers connect through Hyperdrive, which gives you connection pooling and query caching out of the box. Trying Workers + Postgres on my next side project. One catch: the database bills daily from the day you create it, queries or not. Unlike D1 it's not basically free while idle, so I'd still start toy projects on D1. Anyone running this in prod yet? How's the latency?

Solo devs letting one agent approve another agent's PR are running a looser process than Anthropic does. Claude writes about 90% of Claude Code. Their team still has a hard rule: an agent's approval never counts as a person's. Medium-risk changes need one human, high-risk ones need two, and whoever prompted the agent counts as a reviewer. My main bot hands coding to Claude Code and Codex. By that rule I'm the reviewer on all of it, whichever agent wrote it. I think most "agents review each other" setups are just skipping review with extra steps.

My Google account just got suspended. The notice says "multiple account abuse." I'd been letting OpenAI Dot upload my YouTube videos. My guess is that's what did it. It runs from cloud servers, so Google sees logins from IPs that keep changing, and that probably looks a lot like a bot farm. Agents keep getting better at using apps for you. Platform rules still assume a human is clicking. Your bot can do the job fine and still get your account killed.

If you ever saved a provider key in OpenCode Console, just rotate it now. Don't wait to see if the email shows up. And put a spend limit on your OpenAI and Anthropic accounts while you're in there. Whatever someone burns on a leaked key gets charged straight to you.

Anthropic published data on its own staff yesterday, and the number that stuck with me isn't the hiring one. Yes, junior technical hires fell from 27% to 9%. But they also measured idle time: Claude sessions spend around 80% of the time waiting for a human to reply. That's my setup in one line. I run about 10 bots. One drafts my X posts, another hands coding off to Claude Code and Codex. None of them is the slow part. I am. Every draft sits there until I say yes, and I still post from my phone. Their report says time spent managing agents went from 10% to 22% of sessions between fall 2025 and spring 2026. Adding more bots doesn't fix that. It just makes the queue longer. How are you keeping yourself from becoming the bottleneck?

"Claude, step aside. Five minutes." Thirty-one thousand lines in one file. Everything turns red. "I just wanted to build one thing myself." Then: "Claude? Is anyone there?" The part of vibe coding nobody puts in the launch thread. You ship 10x faster, and your own app becomes a codebase you can't touch alone.

Kling 4.0 is out today and the spec I care about is 30 seconds in a single pass. A 30s ad built from short clips means stitching them together and rerolling whichever one breaks the character. Every reroll costs credits and the user waits again. One pass with up to 10 keyframes and 15 references fixes most of that, if it holds up. Catch: right now it's only on Kling's website. API for developers is "coming soon". So if you build on top of video models like I do, you can watch the demos today and ship it later.

Meta just released Muse Spark Open: 816B params, 42B active, 1M context, open weights, built for coding agents. The headline everyone will repeat is "runs on two M5 Ultra Mac Studios." Cool, but that's not the part that matters for most of us. The part that matters is it's on OpenRouter, Together, Groq and others on day one. So for a small product, "open weights" now mostly means you can switch providers when one raises prices or rate-limits you. That's the real lock-in protection, not running it in your living room.

Tesla can fly！

Tool #7 in 6 days: an AI beard remover. Here's exactly how I picked it. I don't guess. I open Semrush and look for a boring keyword with real search and low difficulty: - "beardless filter": 590/mo in the US, KD 16 - "remove beard from photo": 170/mo, KD 14 - "ai beard remover": 110/mo, KD 11 One page covers all three. "no beard filter" has 1,600/mo but KD 31, so that's a side door, not the target. Then a second page for the question people actually type: "what would I look like without a beard" (110/mo, KD 17). The tool page catches people ready to try it. The guide catches people still deciding. Every tool reuses the same image editor and credits, so a new one is mostly content and examples in 8 languages, not new code. None of these will make me rich alone. The bet is a lot of small ones bigger sites don't bother with.

6 AI photo tools in 5 days, and none of them is my actual product. About 80% of my LumiYing users are returning. 20% are new. So keeping people isn't the problem. Getting new people to find it is. My fix this week is pretty boring: ship one small AI photo tool a day, each on its own page that someone might actually Google. Fri: hair color changer Sat: bald filter Sun: beard generator Mon: bangs filter Tue: eye color changer and clothes color changer Each one ships with a how-to guide, in 8 languages. I'm betting someone searching "what would I look like bald" tries the filter, then sticks around for the video stuff. No idea yet which ones bring anyone in. I'll post the numbers once there are some.

"SaaS is dead" usually means "I wanted to win the Twitter launch lottery." I still get paid every month for an AI video tool. $2k MRR. Not life-changing, but real. What a weekend vibe-coded clone doesn't kill: - people who want a result, not another repo - payments, refunds, support - distribution that isn't one viral post Running the boring parts is still a business.

Next step: the bots start asking for rent

Two silly face tools in two days. Bald filter Saturday, beard generator Sunday. Do these convert, or just keep me shipping? lumiying.com/tools/ai-beard-…

好，这就卸载

Reddit just blocked Grok Bot. Guess it's too good at its job

Saturday ship: AI bald filter. Nobody asked for it. Built it anyway. lumiying.com/tools/ai-bald-f…

Am I the only one who fell for Stripe's 2.9% + 30¢?