Leash (@useLeash) profile picture

Keep your AI on a leash — private, on-device, offline. A second brain that runs on your own hardware — rent your compute out, or borrow the mesh's.

Germany
Joined January 2020
Photo by Leash (@useLeash)Photo by Leash (@useLeash)Photo by Leash (@useLeash)Photo by Leash (@useLeash)Photo by Leash (@useLeash)
How does @useLeash work??
2
6
1
17
670
Leash retweeted
QVAC is built by a team working at the frontier where AI, biology, and computing converge, a convergence that is compressing decades of progress into years. That means real research, in the open. Tether just had four papers accepted at ICLR, and open-sources the whole stack for millions of developers and researchers to scrutinize. techcrunch.com/sponsor/tethe…
10
10
1
101
909,401
We are almost done and in the process of tidying up lose ends. #teamMycelium have been hard at work in ensuring that @useLeash is one of the - if not the best harness for local ai inference while leveraging @qvac's SDK It's Day 20 and the job is not done
1
3
1
12
468
Started working on optimization recently, and we at #teamMycelium have been getting even more better results with @qvac's SDK for local ai inference than what we were getting before It's day 19 of working on our Local AI harness, @useLeash ...can't wait for when everything unrolls
2
1
10
225
Fixed a bug that delayed our submission yesterday Inference not getting delegated over the p2p wire using @qvac's SDK for Local ai inference Missed the bonus for early submission But it's reminder that you can do everything right and still lose some.....but lost battle doesn't mean a lost war Go #teamMycelium
2
1
12
213
Just a quick glimpse at our mobile client for @useLeash It's day 17 for #teamMycelium building with @qvac's SDK for Local AI inference
1
6
1
16
439
Managed to set up a - a web interface - a mobile app interface - a desktop app interface for @useLeash, an extensible harness for Local AI with @qvac's SDK. This is thanks of the good work for #teamMycelium ...they've been hard at work for over 16 days now
1
3
1
13
297
Today is one of those days, where you are working on things without seeing the result but you just have to keep trusting that everything will work out fine Come rain, Come Shine... #teamMycelium are actively working on @useLeash with @qvac's SDK for local AI...It's day 15, and jobs not done
1
2
1
14
284
Leash has now be ported (MVP part) to a Desktop app Couldn't have done it without #teamMycelium's support It's still day 14 of building a Local AI Agent with @qvac's SDK for Local AI Development
2
1
1
13
305
Added Authentication Routes today A private ai assistance should remain private and only open to authorized entities; This is Day 13 for #teamMycelium building with @qvac's SDK
2
1
18
259
Spent the whole of today trying to debug insane regression or would i say an untested path Couldn't figure out multi-step orchestration with local models running on @qvac's local ai SDK; it was all failing I burnt more time than ever trying to get things to walk Being on my desk all day long First time in 12 days #teamMycelium ever felt like this Win some, you lose some
2
3
1
15
288
The weakest device on your network isn't a liability. It's an interface. My phone can't run vision models. For a long time I treated that as a constraint to engineer around. This week, building Leash on @qvac's SDK... I treated it as a question. What if instead of failing, it just asked? > Sends the image to a stronger device. > Gets the answer back. > Moves on. > The strong device doesn't know it helped. > The weak device doesn't know it borrowed. They're just the network. The most resilient systems aren't built from the strongest parts. They're built from parts that know how to ask. That's what #teamMycelium is building.
1
3
1
15
241
Every night, while I sleep, my Mac fine-tunes itself on my own data. - Notes I took. - Answers I liked. - Corrections I made. It curates them, runs a LoRA ... fully on-device via @qvac. Then it scores itself against a frozen eval before it promotes anything. t can't fake growth. The test set was locked before training started. The model that wakes up tomorrow knows me a little better than the one that went to sleep. thank you #teamMycelium for bring this to life
1
4
1
18
338
We did the least impressive thing you can do mid-hackathon: we stopped shipping features and rebuilt the floor under Leash. What a demo never shows you is that a rushed foundation doesn't crash - it lies. Ours did. Our agent looked like it was using its tools; underneath, it was faking every call. Confident, convincing, completely hollow. That's the trap with foundations. The cost of cutting a corner stays invisible right up until it's load-bearing ... and by then, everything you stacked on top comes down with it. You can't hasten a foundation. You can only build one that tells the truth. So that's what #teamMycelium is doing on @qvac's SDK ... slowly, and for real.
1
5
1
16
306
After 8 hard days of building, #teamMycelium took a hard step back today to rearchitect what we were building into Leash, our local ai agent using @qvac's SDK It's going to be a busy tough few days
3
5
1
28
409
This is all running locally, no wifi-connectivity #teamMycelium are currently looking on how to reduce the tool call using @Qvac's SDK, as not all request should have tool calls and what not It's still Day 7 and there's so much more to do
2
2
1
18
547
They still have a long way to go...
Small craft thing I'm happy with: the chat history is hidden by default and slides in when your cursor hits the far-right edge...like it's about to leave the screen. Day 5 of #teamMycelium working with @qvac's SDK for local ai inference
1
2
1
5
427
Small craft thing I'm happy with: the chat history is hidden by default and slides in when your cursor hits the far-right edge...like it's about to leave the screen. Day 5 of #teamMycelium working with @qvac's SDK for local ai inference
1
2
1
16
727