Indie Hacker, Founder @MeetingRoom365 - Previously Analytics @Adobe, Design @HBO
Taipei
Joined April 2016
- Tweets6.2K
- Following907
- Followers1.1K
- Likes29.2K
Me waiting for PHP updates to hit Ubuntu security and docker hub…
Nah fuck this I’m going to disable Apache and lie down, wait for a bash script to ping me.
Claude Code users on iOS only want one thing, and it's disgusting. @ClaudeDevs @RobertJBye
Had 3bpw on my M1 Max 64gb (running locally!) rewrite its three.js game for iOS, low level Metal View, its own 3d engine. Referenced blog posts about some of the trade offs it was considering. Tested at the end.
This is frontier quality, local, on a $1k machine. Unbelievable.
I’ve been thinking about this for a few hours. At what point does software become malleable enough we can just use whatever hardware we want and treat these like thin clients to our modern machines.
Software is good enough to use a 5k iMac as a display why not an older iMac.
Custom avatar packs you can swap out in settings. First one is Avataaars. Fits better than I expected!
Qwen3.8-Flash-Next-Sushi-3bpw from @Beamsters 1.5hrs, same prompt. Averaged 29 tok/s, tighter memory fit, 200k context, M1 Max 64gb. Significantly better and somehow faster result from the 3bpw vs 2.6bpw for the complex task.
Kind of shocked, this is frontier-level output on this eval, but it's Qwen3.8-Flash-Next-Sushi-2.6bpw running via Sushi on my M1 Max 64gb at 30 tok/s.
Extended thinking, it took about 30 minutes on the first good version, then kept polishing it unprompted for hours.
Fans were going whurrrrr the entire time, very little headroom for anything other than Chrome / Playwright / CLI. 59gb ram reserved for the model and context.
2.6bpw should be awful but this is delivering on special expert quantization. Tooling around it is really good as well.
Still some minor hiccups. Qwen in general has a heavy bias toward its own tool formats for tool use and if yours don't match, it makes mistakes and takes a bit too long to recover. It does recover though.
And the version of this with an 8k reasoning limit could not handle this task. That is true for most models though.
Overall phenomenal work @Beamsters1
Tasking my agents over the tailnet. Sometimes they share their screens to demo a feature and I have to ask to see it again.
If Claude is your daily driver and you wish you had dots or don'ts or groks or whatever running locally on your Mac, give this a try.
It's a minimal harness that your own Claude instance uses and improves on its own to fit your needs.
github.com/kidGodzilla/squig…
Instead of running in the cloud, it uses your own Claude or Codex subscription. Hire squiggles for whatever you need to get done. If something doesn't work for you just talk to your lead.
Been hacking on this for a few days now. Basically, building tools to assign tasks to agents and figure out how I can keep everything running smoothly on my own machine.
They all have goals. When they finish something we do a 1:1 meeting and they present their work, I ask questions, look at the diff, and merge it.
Sometimes we have group calls when I'm stuck and we need to figure something out.
This is quietly about $50,000 worth of unreleased hardware sitting on this desk to run this. Still waiting on a confirmed price for the M5 Ultra 512gb (late October).
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
Send Broadcast @sendbroadcast is a really great self-hosted Mailchimp alternative @geetfun has been working on for a few years if you're looking for something. You don't need to waste a week vibecoding your own.
I don't know if it does everything from that insane Beehiv tweet but what you want is a solid foundation, and a single provider to send email so you can build reputation.
sendbroadcast.net/
AI haters are weird. Some guy is just like "I spent a day trying to get Pikmin to run on my Quest headset" and all the comments are "Boo, AI is bad, go away".
old.reddit.com/r/Pikmin/comm…
Building Agents for Opus so I don't lose my mind trying to use Claude Code directly.
Little virtual office for my imaginary teammates. We hop on calls. They present their work, ask each other questions, get approvals to merge.
It plays French radio in the background.
I basically saw agent-standup.com and realized all the open tabs in Claude code was giving me anxiety. Wanted to ask our agents to do things instead of babysitting them in Claude.
Calling them squiggles. All props to Arthur for the original characters.
Check out agent-standup.com/ - really cool project!
Just supported Arthur on buymeacoffee 🎉
buymeacoffee.com/bored_dev_a…
64/64 points! Finally after 9 years of arguing with my significant other I have concrete proof I can load a dishwasher just fine. github.com/composo-ai/Dishwa…
