I can’t wait to get home and try this on my Pi. Past few months, I’ve been building a fully local agent [0] that runs inference entirely on a Raspberry Pi, and I’ve been extensively testing a plethora of small, open models as part of my studies. This is an incredibly exciting field, and I hope it gains more attention as we shift away from massive, centralized AI platforms and toward improving the performance of local models.
For anyone interested in a comparative review of different models that can run on a Pi, here’s a great article [1] I came across while working on my project.
It was limited to 2,000 tokens each. I assume it usually hit that. So could be closer to 777M. assuming they didn't just cache it and just start rotating after a day or two..
I really wish I'd had this tool six months ago when I was designing a GUI program for macOS 9. I wanted to import pictures from my modern laptop to a PowerBook Duo and vice versa. Converting all the assets was a pain, and I think this tool would've been incredibly convenient, even just for previewing the images.
The program existed six months ago. That said, currently it only allows to render old QuickDraw pictures on a modern Mac OS X. A tool going the other direction would be simple for bitmaps (just use PNG) or very complicated for vector data.
One thing I'm thinking about making a tool that uses old codecs to make compact images for 68K machines. Basically, use today's CPU power to make efficient road-pizza images.
Offtopic but which tools and resources would you recommend to get started with MacOS 9 GUI development?
I’ve experience with WinAPI/MFC but the classic Macintosh API is quite different and I’d often run into hangs when trying to run simple hello world-ish C++ programs compiled with CodeWarrior.
For those wondering about the use case, this is very useful when enabling streaming for structured output in LLM responses, such as JSON responses. For my local Raspberry Pi agent I needed something performant, I've been using streaming-json-js [1], but development appears to have been a bit dormant over the past year. I'll definitely take a look at your jsonriver and see how it compares!
Do any LLMs support constrained generation of newline delimited json? Or have you found that they're generally reliable enough that you don't need to do constrained sampling?
The page referenced in the article, Frame of Preference [1], is such a great example of this feature. When I read it last week, interacting with a live Mac OS system while reading the post felt like magic.
While I agree with all the previous comments, your comment sparked an idea in me. I started imagining a future where we develop a new programming language optimized for LLMs to write and understand. In this hypothetical scenario, we would still need developers to debug and review the code to ensure deterministic outputs. Maybe this isn't so far-fetched after all. Of course, this is just speculation and imagination on my part.
This is very nice and simple. A few areas for improvement, in my opinion: the URL should be easy to copy, paste, or type into another device. I'd suggest designing the route like pindsend.app/pin/CODEHERE. Also, for some reason, copying and pasting the URL didn't seem to work in its current form. I would also consider implementing a QR code to allow quick scanning and redirection on another device, especially a phone, see Wormhole's implementation for reference [1]. Regarding security, the PIN looks quite short; I'd suggest increasing its complexity or length and automatically expiring sessions after a few minutes of inactivity. Additionally, definitely consider implementing end-to-end encryption if you haven’t already.
Thank you for the feedback and wormhole.app suggestion! As for QR code - it's already available after you've set up your session, but UI might not be as obvious so I'll work on it.
For PIN code - you might be right, though brute forcing PIN codes with 6 alphanumeric characters and throttling on signaling server might not be as easy as it seems.
Communication inside a sharing session is end-to-end encrypted as it uses WebRTC, and session will "expire" as soon as all clients are disconnected, because no data is stored on a server - it's all between connected clients
I am actually quite surprised and frankly concerned about the fact that wormhole.app constantly is considered as the wormhole protocol which, it frankly with its confusing name isn't.
It isn't the wormhole protocol. I am on mobile at the moment but someone can definitely link websites which are genuinely using the wormhole protocol as I was involved in the scene of learning about such protocols and the alikes of crocs and there was atleast one open source website based wormhole app.
Also,I may be wrong,I usually am but if you are suggesting the pin to be long and e2ee, then isn't that just pure wormhole, doesn't wormhole protocol also do the same?
Also, i may be wrong again but by p2p / webrtc. I was already imagining encryption. Isn't that the case with webrtc given I have considered it "secure"
There is chitchatter which can be self hosted with ease as a static app and I am pretty sure that one can reconfigure it to better suit the purpose of this app.
Cheers! Hope you can edit out the mistake. Appreciate it.
For anyone interested in a comparative review of different models that can run on a Pi, here’s a great article [1] I came across while working on my project.
[0] https://github.com/syxanash/maxheadbox
[1] https://www.stratosphereips.org/blog/2025/6/5/how-well-do-ll...
reply