Vibe coding is all the rage at the moment if you follow certain parts of the Internet. It’s very easy to dunk upon it, whether it’s to mock the sea of people who’ve drunk the Kool-Aid and want the magic machine to make them a million dollar app with no work, or the vibe coded web apps with security holes you could drive a bus through.
But AI-assisted coding is now a thing that will stick around whether you like it or not, and there are many who want to dip a toe in the water to see what the fuss is about. For those who don’t quite trust the magic machines in their inner sanctum, [jscottmiller] is here with Clix, a bootable live Linux environment which puts Claude Code safely in a sandbox away from your family silver.
Physically it’s a NixOS live USB image with the Sway tiling Wayland compositor, and as he puts it: “Claude Code ready to go”. It has a shared partition for swapping files with Windows or macOS machines, and it’s persistent. The AI side of it has permissive settings, which means the mechanical overlord can reach parts of the OS you wouldn’t normally let it anywhere near; the point of having it in a live environment in the first place.
We can see the attraction of using an environment such as this one for experimenting without commitment, but we’d be interested to hear your views in the comments. It’s about a year since we asked you all about vibe coding, has the art moved forward in that time?
Meh. There are other ways to do that. But if you require someone else to figure it out, then there is that.
The other thing is the ‘Nix-nuts’. I have found them to be a pain and am unwilling to deal with their shenanigans.. But Ms List, most likely being a better person than myself, is probably more adept at dealing with those ‘types’.
As for me, I do it with Anthropic stuff + Slackware, as nature intended.
Anyone who believes this will protect their important files is in for a rude awakening when their “AI” “assistant” ends up formating the host computer’s drive(s) from the live environment anyway.
I run Ubuntu in a hyper-V VM with the only connection to the outside world being a software bridged network interface and SSH-in. It runs in full permissions mode which is a huge deal as it definitely accelerates whatever you’re trying to accomplish as long as you keep watching the process since it sometimes gets a little caught up in details and a bit of steering helps it get back on track.
Hight recommend the VM approach.
What kind of stuff do you get your LLM setup to do? Is it mostly programming? Or something other than that?
“AI-assisted coding is now a thing that will stick around whether you like it or not”
Yeah thanks, I was wondering if resistance was futile or not.
Currently AI-assisted coding requires economy-imperiling levels of subsidy to exist. That’s not going to last, and I actually can see it ending in a way that discourages anyone from providing similar tools afterwards.
We think of software as being permanent, in that once it’s written it will always exist. But even to the extent that’s true, it’s certainly not the way to think of services, especially ones that burn cash at the rate LLMs do. Until there’s an affordable vibe-coding tool that makes money, these things are no more guaranteed to stick around than MoviePass was.
honestly it’s reached the point of usefulness that it would 100% be cost-effective for my company to build me, personally, a $50,000 inference cluster.
That’s the point inthink people are missing. You don’t NEED to pay a subscription (I do, AI ultra for business is making me poor) if you can build the infrastructure.
Hm, devstral-2-small can run reasonably on computer with a 32 GB GPU, and definitely can generate 10s to 100s of lines of code reasonably, so even if the subsidized services go away, I expect that people will just start using local models (and if the subsidized services stop eating up all the memory and GPUs this might be a lot more affordable).
Heck. What sort of PC are you running where you’re burning cash or getting subsidies? I thought my homelab’s AI server was impressive but whatever you’ve got going on is clearly out of my league.
…unless you’re conflating AI agents with ChatGPT because you don’t know the difference.
I’m duly inpressed by your homelab that can build an LLM from scratch to write code from natural language prompts. You should charge money!
Or did you mean that you downloaded a model someone paid billions to create, and it’s good enough today that you don’t forsee wanting to update it ever?
Stunning indictment of the inability of vibe coding to produce reliable results, that it’s practitioners need a whole disposable install for when it inevitably blows up in their face.
That’s why there’s git, right? Except the change sets are huuge.
The down side is the physical media, we used to have immutable optical media, but with flash we moved away from that. Once we went over 4gb/9gb DVD offered it was kind of unavoidable.
I still want a USB firewall that will provide write-lock to an external drive. A write-protect switch on the drive or drive cabinet itself used to be standard back in the pre-PC days…
I’m a software engineer and will be retiring at the end of the year (can’t come soon enough really). The “cool kids” can have their vibe toys, let them run wild with ’em. When their software has issues that need fixing and they can’t figure out where their new shiny fell down I’ll be available to come in and “make it go” for a hefty contractors fee.
K