veloi

probably burning tokens
building

[veloi @veloi_]

I used tart for this which exposes an easy CLI to create your own custom images.

Personally, I set this up with mattpocock's sandcastle workflow - letting my parallel agents test their code changes. But I’m sure this can be just as powerful with Hermes or openclaw.

[veloi @veloi_]

today i learned you can run full Xcode provisioned macOS vms for your AFK/parallel agents

works great for react native / iOS workflows

makes you wonder why everyone went out to buy a mac mini for openclaw lol

[veloi @veloi_]

@verycracked i can't switch to guis beacuse of how bad mac window management is. learning tmux and letting them live in one terminal speeds me up so much.

[veloi @veloi_]

@bratlando what?

[veloi @veloi_]

@mattpocockuk just to be clear: this is a wrapper built on top on sandcastle

[veloi @veloi_]

been experimenting with @mattpocockuk's sandcastle.

fully parallel, multiple impls with a judge, multiple reviewers with reconcilers, plugin support, deep harness config interfaces

single `sandcastle.config.ts` with recipes built in

[veloi @veloi_]

@mattpocockuk Deep harness configuration support with Pi agent

[veloi @veloi_]

@mattpocockuk Test multiple implementations and let a judge pick one.

Saves to a directory for later human review so you can find the model that works best for you

[veloi @veloi_]

@tekbog its quite funny from a software eng background actually. they're really living in another reality. god forbid we ever have any nuance

[veloi @veloi_]

@ScarabianNight @deftill @ek2805 @LadyAuro_ thats just up to how you use it 🤷
i always challenge the AI and evaluate its output

[veloi @veloi_]

@prasenx @AsideAI is nice. never used any agentic capabilites but having chats in my tabs is nice. small UX improvement over maintaining a collection of chatgpt tabs.

[veloi @veloi_]

@FrightNite52 yeah maybe gpt-3.5-turbo. modern thinking models will double check. as long as you're using a frontier - this is no longer a concern for basic research.

[veloi @veloi_]

@shydolph @LadyAuro_ Of course not - I think you misunderstood. It's to show the scale of global water use and how much AI contributes to it. A single number on it's own isn't helpful so I used agriculture to show that the water usage is relatively small compared to our demand in other things.

[veloi @veloi_]

@shydolph @LadyAuro_ I find it hard to believe that data centres are responsible for that from what I've heard. It uses water but natural factors and agriculture are orders of magnitude above AI.

"Not which model." I think you misunderstood; a local model doesn't run in a data centre-it's on-device

[veloi @veloi_]

@shydolph @LadyAuro_ would it be acceptable if he used a local model? genuinely asking; i'm not sure why people are being so hostile.

[veloi @veloi_]

@splitbycomma i wonder how this croud would react if he used a local model?

[veloi @veloi_]

@adrianpkstream @dragosroua @thekitze at the moment - plain text input just kinda sucks for UX.
and also for any app which requires rich content.
i can see certain things going this way (e.g. amazon sucks) but most apps on my phone wouldn't fit into that constraint.

[veloi @veloi_]

@derekyt_ these "OFFICIAL STATEMENT" images cringe me out so much.
not to mention the text wrapping so bad and painful to read.

[veloi @veloi_]

@imbabybrooklyn what app is that?