Its a very curious project, but don't you end up pinning repligraph usability on model weights? Since you take indeterminism out of the equation, a repligraph's notability is as significant as the producing model's weights, and since there is no dice rolls to be made, the eyeball problem:
> Our blind spots, while not perfectly correlated, have substantial overlap
is entirely replicated. Models that are diffused from one another can have the same blind spots, the same loose statistical reality that exists with humans. This is partially addressed in steering:
> A repligraph proves that a model generated some artifact. It does not prove that the model did a good job, or that the artifact is safe.
but I think the "Peer Review" solution is inadequate, and with some jailbreaking prompts' innocuous looks considered, "the attacker just needs to find one prompt" might be much easier than it appears.
Batching seems to be a huge economic turn off for proprietary model determinism, but I think its entirely viable for consumer models. Trustless evals are brilliant and should've been our reality. Nice project, good luck on your endeavor.
True, I expect models to be pretty correlated with each other as well. But you can at least attempt to quantify that correlation in a rigorous way, by running experiments (e.g. by pointing each model at a big project and comparing the sets of bugs they each find).
And yeah, the "double-edged sword" aspect of determinism is definitely the biggest bullet that you have to bite. For me, it's better than the alternative; without determinism, such jailbreaking attacks are still possible, just harder to detect. But I certainly would not want people to go around thinking "it's deterministic, therefore it's safe."
So I'd expect AI to rapidly accelerate Nix adoption. Hopefully, because it is amazing.
Then all of my dotfiles are backed up, with any important other config, in a big gig repo. And all of my documents/photos/etc live on a separate hard drive which gets backed up to the cloud.
All of that to say, with the right setup nix is mostly unnecessary, and honestly the benefit of not using it is that I can wipe my system anytime and restart it with just my usual important programs ready to go. Any mess made along the way gets purged automatically in every fresh install
I worked with a very forward-thinking functional programmer around 2012 and even back then she was raving about it. And of course, people thought she was a raving lunatic.
I give her credit to this day. It took me having one too many problems caused by NOT using Nix, to cause me to finally dive in.
And there's no going back. Which is what happens with the best technologies.
I say this as a daily user of Nix.
(I say this as somebody who deeply loves the concepts in most functional languages, and loves the concepts in nix, but recognises that the developer-experience of both is about as pleasant as reading somebody else's code-golfed perl)
Now compare it even to Haskell, to say nothing about e.g. Typescript.
Now you can just send this to agent "port dis <repo link>" and 20 minutes later you get a working package, or 20 working packages what were required for that package to work.
Fixing stuff also easy, just today for some reason my machine wouldn't wake up monitors after sleeping them, a few minutes later claude fixed my configuration. Making QuickShell things with it also easy. For years I've avoide all GUIs for networkmanager and bluetooth on linux because they all looked awful and unusable, now QuickShell things made just for me, exactly the way I wanted.
Is `nix` amazing though? Eh... if it were implemented better, it could be amazing. It's like python virtual environments or `pyproject.toml` for general software. It solves a real problem, but it does so by introducing a possibly even bigger problem.
it's more like, it definitively solves a problem forever, like nothing else.
It's like when you have immutable data and it permanently solves a class of problems that can result from mutable data.
It's sort of like that.
It’s not the only thing that agents have changed for me - I also went back to using neovim and tmux; partly because of how I use remote development servers but also because now I can get my setup working perfectly for me without a lot of hassle.
Open source and code managing all the things has always been a great idea but now it feels crazy to do anything else.
My mac setup hasn’t been replaced since 2021 or something, and it was copied from my old mac. I have a dotfiles git repo, which I drag around and have multiple branches for my daily driver. For temp setup, I just copy files and edit.
These days it's magical, but it was even a huge boon when it was barely functional. Nix just had a way of leaving me super confused and stranded.
"We have this system that is so amazing that... it requires a paid subscription to an LLM provider or a $5000 GPU for most people to understand and configure".
When instead Nix could've focused on releasing flakes 10 years ago with probably the same effect. And fix issues like this: https://news.ycombinator.com/item?id=50040083
Flake stabilisation probably hasn't gotten any easier/more likely by LLMs, but it does make it quicker for people to pickup technology.
Technology sometimes has that odd effect. (And it will apply to different technologies in different contexts at different rates.)
I use devenv (yet another nix flake dev environments) and opencode.
It is quite a synergy : my nix usage profits from the agent and the agent profits from being able to install just anything.
I even vibecoded a plugin for personal use that runs all bash tool calls through devenv and uses the supplied nix lsp package:
FWIW, I used Debian more than anything over decades, and also have little patience when it comes to unstable computers. NixOS is a breath of fresh air and has reinvigorated my love of Linux. I haven't found it less stable than Debian, and also have a higher degree of confidence when making changes. It has its quirks, but ultimately I've found it a joy to use, and I'm never going back to a distro that doesn't have similar capabilities.
If you boot default Debian and apt install a handful of packages from the primary channel, it probably doesn't matter much, but if you enjoy even a modest amount of customizing, NixOS really shines. From one Linux user to another, I encourage you to try it! I have no affiliation, just an old head stoked on Nix.
...of which the following are most important:
2. Run deterministically
One vCPU on stock KVM, so code in the VM runs on the real CPU. The VM's kernel carries a small Rewind platform: interrupts arrive only when the VM hands control to Rewind, time moves only then, and the timestamp counter and hardware RNG are hidden. A step is one of those handoffs.
3. Keep keyframes
Every quarter second of wall time, Rewind snapshots the machine using KVM's dirty-page [aka Memory Pages that have been written to since the last snapshot] log. Pages go into a content-addressed store (BLAKE3, zstd), so a page shared by keyframes, runs and forks is stored once."
Now, that's some brilliant engineering right there!
While I like the product name "Rewind VM" (and yes, that's what it allows the user to do!), it could just as easily be called "Deterministic VM" and that name would have been wholly appropriate, as well!
Future OS and VM engineers should study Rewind VM. Those are some great engineering ideas there, and determinism, or at least optional determinism, should it be desired, when it is desired -- is the one thing that is completely non-existent in most OS's, most VM's, and most other program runtime environments...
Anyway, a great article on using Rewind VM (never knew about it before this article!) with Nix builds.
There could be great ideas in Rewind VM, but the page is littered with Claude-ish which is immediately off-putting.
The "Undo" time-travelling debugger for C/C++ has been around for 20+ years. So I guess the answer is simply that most people chase shiny new things instead of using established, reliable, and, hence, boring tools.
Curious: how does Rewind deal with randomness?