Going to go out on a limb and say that is the pattern with every single OS at the moment. Itll taper off eventually, but there is a lot of technical debt and human slop to clean up for a bit, then things should settle down.

I love insecure kernels. I actually want the bad guys to win the race against security. Only bad guys should using AI. Debian shouldnt be allowed to use AI to keep up with them.

Oh no, my kernel is more secure thanks to AI. How will I ever trust Debian again. Guess I gotta move to a new distro. One that doesnt use the Linux kernel.

The A20 could theoretically handle 4K gaming. Just depends on the game developer. The iPhone 18 Pro has 7 GPU cores, and it has about 12GB unified RAM to share between the CPU, GPU, and Neural Engine.

Plus the ARM instructuin set is a bit more modern, so there is probably performance gains to be had that x86 couldnt handle

Plus, if game developers targeted ARM instead of x86, im sure their games would actually run great on those A20s, but they’re probably not going to make an ARM version of their game.

Uh oh, looks like you didn’t read far enough:

On Monday, two days after the incident occurred, Mr Crane confirmed that the data had been recovered. The Independent has reached out to Anthropic and Cursor for comment.

So again… Well, do I need to paste it again?

So again, because we make backups before letting AI touch anything:

“Hey slop machine, you fucked up. Restore the last backup and document the mistake you made in AGENTS.md so we avoid making that mistake again. Im going to go make a coffee, brb”

Oops, slop machine broke something.

“Hey slop machine, you fucked up. Restore the last backup and document the mistake you made in AGENTS.md so we avoid making that mistake again. Im going to go make a coffee, brb”

Hardware I’m running:

• 8/16 AMD CPU
• 32GB system RAM
• 12GB GPU VRAM (AMD)

I’m mainly using these MoE models:

Qwen-3.6-35B-A3B

• Q4_K_M quant quality
• 128k context (conversation length before it compacts)
• Gives me about 260 prefill and 19 token gen speeds

Gemma4-26B-A4B

• Q8 quant quality
• 128k context length
• Gives me about 190 prefill and 14 token gen speeds

I created git repos on my main workstation for each homelab server/service I maintain that keeps:

• documentation
• notes
• lessons learned
• scripts, configs
• runbooks
• backup details
• security audit details
• log items that need attention
• infrastructure

I just point a local LLM (offline model that runs on my workststion) into those repos and ask it to perform certain things on those servers. It can do things like update packages, install packages, make config changes, set/check permissions, read logs (and fix errors in real time), and check the health of the overall system.

I have it run pre backups before making changes, then post backups once its done.

Once changes are in place and everything is running okay, I ask it to update documentation in the repo and tag the release.

I use opencode that connects to a llama.cpp service. opencode lets me gate the AI so that any elevated commands that it needs to run (e.g. sudo or ssh), I have to approve it. It cant just go around making changes without permission.

This question is how new apps are born

Mobile CCTV patrol system getting bigger

Cyber criminals

They should probably jail the CEO for this like they would any other “hacker”

Those are good things to ask prior to using it, but thats just the beginning.

• If the training data matches what you would look up on the internet anyway, is it a deal breaker?
• If it cannt be used for production, what about for organizational purposes? Documentation? Resource tracking? Upstream changes? Merging considerations? Security auditing?
• It sounds like you would rather do the coding part, correct me if I’m wrong. This way you have full control of the license. For someone who isn’t a coder, they would probably have to consider that the code it generates could closely resemble an existing licensed work since the models are trained on open-source data.
• I think the getting dumber argument is going to depend on the user, and its not an LLM issue specifically. You could definitely bypass the thinking process, but you could also make it explain things step by step, which could also accelerate learning stuff.

But you’re right, to each their own.

Whats your viewpoint on open weight models that run on local hardware only? Such as a gemma4 or qwen3.6 models that can run on pretty much any gaming PC?

KDE could source their own training data, train their own models, add their own post-training behaviors, and release the weights under their same open source license, and people would still hate it, even if it was locally run and doesn’t go online at all.

Let the old heads be stubborn and fork it and stick to their human slop desktop environment.

So where are the users migrating to?

Fix your title. It doesnt match the link’s title

To be clear, this is a personal project of the employee. Not a government project that the employee built.