天空вℓσи∂ [Deprecated edition] @skyblond@social.cleverlibre.org

Setting up a NUC for my friend. I was plan to use some off the shelf solutions like truenas or something, until he told me the demand.

He want a usable linux desktop, while it also provides a way for accessing remotely andprivately. Also he want a media center and a storage device. And it would be so great if I can somehow put a software router in it too.

So I decide to go with opensuse Tumbleweed with Gnome Wayland. Setting up a proxy would ease the internet issues. Then, I use tailscale for private remote access and plex for media center. I also installed cockpit for management (so I can know what's going on if something going wrong in the future).

There are a lot of changes to the original plan. And mostly because I don't know how to create a bridge using want interface under network manager (I'd prefer wicked), so the kvm and openwrt will not run on this nuc.

The plex is a huge pain too. It keeps failing to scan the files and turns out it's me forgetting to change the permissions. I eventually setup a cron script running every minute to apply 777 to that folder, hopefully won't cause performance issue.

Also, btrfs support from opensuse is great. If something goes wrong, I can told him just boot from a previous working snapshot.

Definitely going to wait for the patch for LLaMa 2 70B model, even with 4bit quantization, I still want to see what this 70B model can do.

The 13B is really amazed me. It's 10% parameters of ChatGPT, but it's free with (almost) no restrictions (please do not use it for ilegal reasons), and can run on my own laptop.

llama 2 13b chat f16 model show off.

That's brilliant!

Can't imagine that's running on my laptop. That excitement reminds me of the time when I make a LED blinking using C51 when I was 12.

It's much better than 7B model!!!

Can't wait to see the 70B models. The llama 2 70B model uses a different design than the llama 1, so llama.cpp currently won't work on 70b model.

But that's amazing! A LLM running on my laptop!

Switched to 13B chat f16, it can speak Chinese!!!!!

Well, Chinese is very slow, I think it might due to token related issues. I guess it treat each character as a token

Context: 7B f16. I'll try 7b-chat f16 and 13b-chat f16 later...

I'm sure facebook/meta still has a long way to go.

I never expected to be denied by a LLM.😂

ok......

Looks like my laptop can't even handle the smallest 7B model😅

Looks like the smaller model (7B and 13B) shares the same design as llama 1, so technically I can use that model for the llama 1 code.

Have a chance to download the new llama 2 model weight.

I have no idea of how to work with Python, but I will try to play it and see if I can get it work on JVM😋

Some observations about Android's work profile. It's kinda like multi-user, but you don't have to log out and log in to switch user. The second profile is a separate space from your main profile under the same account.

Today I'm trying an app called MissLee (李跳跳), it uses the a11y (accessibility) to simulate click/touch to close add automatically. I put all Chinese apps in my work profile, but the a11y settings can be only enabled for apps in the main profile. However, the app in the main profile has no awareness of apps in work profile. Aka the MissLee cannot close ads for those apps in work profile.

So I installed the MissLee in the main profile anyway and enabled the a11y settings for it. To my surprise, the MissLee in the work profile is also enabled (reported by the app) and is working fine. So I'm guessing the Andorid is managing settings based on the package id. If you give some app some system-level permission (like simulating touch), then it will be given to all installed apps using the same package. To make sure there are no holes in terms of security, you cannot install apps that use the same package id but different signature. This ensures all apps using the same package id do come from the same developer.

Brilliant design, if they can document this behaviour.

Seriously?