Rule 1: Only one mouse is allowed to move at every single time.
Rule 2: What is a context window???
Rule 3: SD CARD CORRUPTION
Rule 3: SD CARD CORRUPTION
Rookie mistake. Next time store your swapfile on Google Drive so it’s their problem!
You joke, but I did something like that. I NFS mounted a disk off my desktop to a RPi. I then made a swap file on the NFS file system.
I could then build a C++ project directly on the RPi because the build normally hit RAM limits. This was a 1st gen RPi, so it had 256MB RAM.
The build took forever (many days), but the program ran fine once built.
Tangential, but stuff like this makes me wish Intel’s Optane had taken off. It was supposed to have the persistence of an SSD with performance near that of RAM, but it was expensive and the OEMs who included it relegated it to little more than a glorified drive cache.
Imagine being able to allocate some of your spare storage as extra RAM the same way the Pi lets you decide the RAM/VRAM split. Or more realistically, sticking your swapfile on it and your build not taking several days even though you’re only working with a quarter gig of RAM.
Rule 3: SD CARD CORRUPTION
The Pi 4 and later can use a USB drive instead of a SD card for the OS drive. It’s way faster and less prone to corruption.
I say this, despite noticing which sub we are in. ;p
Or an SSD, for storage too.
Rule 4: If it is mine either the soc cracks or the pmic.
I swear microSD is only marginally better than DVD-RW
I had to purchase a new dashcam recently, and I specifically picked up one that could save to a USB C SSD instead of its SD card because my previous VIOFO A129 would constantly corrupt its SD card.
Rule 2: What is a context window???
Its the window through which you enviously stare at people running megatoken context windows.
I mean, that’s essentially how it works at every scale. When it says thinking, the first part of that wait is your query waiting in a queue
Use one of those CLI tools that show you the actual thinking process then. There’s a LOT of LLM output between tool calls and code edits. Claude Code doesn’t show it because a lot of it is dumb as a bag of rocks lmao
“expert”
I actually have a 3B running on my Home Assistant, an 8th gen NUC with 16G of RAM I got for $175. It’s quite snappy… if not terribly correct about everything. It’s capable of running an 8B but then the response time on the Voice PE is too slow to suit me (goes from 3 seconds to about 8 seconds).
What model are you running?
Qwen 2.5.
Can it be used for agentic workflows or just simple Q and A?
Yeah that’s why I’m running it. Couldn’t get reminders to work with all the possible permutations of time notation without an LLM. It still doesn’t fully work properly but it’s good enough and it has the side effect that it can answer any question you ask, for better or for worse. It gets simple things right more than half the time I’d say.
I’m keeping my fingers crossed that the work being done by groups like GLM will allow something smarter to run on this hardware within the next year or so.
Huh, what exactly are you using? I tried to run home assistant’s, well, assistant, and qwen 2.5 3B just didn’t properly call any tools. And bigger models become pretty slow on my ryzen 2400 “thinclient” home server. I’ve tried it via Ollama, even did some dirty hacks to use the iGPU, but that’s about the size that I can run without waiting a minute for a simple prompt.
It’s not impressive, but specifically it is: ByteSpeed Intel NUC 8i7BEH Intel i7-8559U 2.70GHz 16GB DDR4 RAM 256GB SSD
I’m actually not sure I’m using it to call any tools, so I might have been overly optimistic sounding. I set it up so that HA tries to answer any incoming requests but when it doesn’t understand something, it hands it to the AI. Reminders are not understood natively by HA, but I set up a pretty elaborate automation to handle it. As I understand it (I used Claude to help me set it up) the AI is parsing the {time} part of the automation, but it doesn’t quite work right. Basically if I say “Hey Jarvis, remind me at 8:15 PM to move my truck” it works, but if I say “remind me in two hours to move my truck”, it fails. I gotta spend some more time with it.
I’m using Voice PE to talk to it so I can say “hey Jarvis, blah blah blah” and if HA doesn’t recognize what I’m asking (ie: HA/PE understands “set a timer for 30 minutes” and will set a timer for you) it passes it to the AI. If I ask it “who was in the Beatles?” it takes about 3 seconds to think and then will respond with “The Beatles consisted of… <lists the members>”.
I would provide you with more detail but I’m out of town right now and don’t have direct access to it on this computer.
The problem is not enough RPis. If you get enough running in parallel, you achieve AGI, Nirvana, and quite possibly exit velocity.
oh no, OP will now buy all RPi stock and manufacturing capacity for years! what have you done?!
what
