> two builds from the same Dockerfile separated by time can result in rather different container images
This is true, and the same is true of shell scripts, but isn't it true of nix as well? If you update nixpkgs and then install something, you have the same chance of getting a new version of it as you do with the `sudo apt-get install` in the Dockerfile, no? I had once looked into using nix to intentionally install older versions of software and it was not as easy as I had hoped.
You can install a single package from a specific nixpkgs commit and it will pull in whatever dependencies it needs to make that work, with everything as of the time of that commit. You may have duplicated libraries but they can coexist fine, and you'd have that with docker as well (probably even worse unless you are very intentional about building layers).
I imagine people do this somewhat regularly at least to e.g. occasionally pull in a package or two from unstable when they normally run a stable system.
My recollection from trying this is that it's not so easy to find that specific nixpkgs commit. It's doable if you're determined, but it's not a 1 click thing, but my information may not be up to date.
If you `git blame` the software you're interested in, I can't imagine it should take too long to find the last commit that had the version you want. Then you could either use that, or for the first commit with the new version, use commit~. Or anywhere in between depending on whether dependencies are important to you (that's at the limit of my git-fu, but I imagine you could do some kind of restricted blame to then find the dependency version you care about if you do care, and recursively do that until you have exactly what you want).
If you have a CI system that's building containers (e.g. using flakes), you can have the nixpkgs version you're using pinned in your own config, so you can just look at your own (smaller) history.
This should all be way easier than trying to do it on any other distribution.
If you’re using flakes, as long as you don’t update the lock file, you will always install the exact same package. This holds across machines using the same config.
Seems like a relatively simple fix, Amazon shouldn't be allowed to tell suppliers what price they can set on their site. No one should be able to tell a supplier what price they can charge to others.
"We gave Claude a prompt to search through a massive database of DNA sequences for interesting new examples of RTs. Our involvement was limited to the initial prompt and the lab work, while Claude agents combed through the database, investigated the distinct RT families, and used their own judgement to identify interesting candidates. After 21 hours spent searching this data by roughly 950 agents using 210 million tokens, one of the agents spotted something remarkable: a repeating pattern of DNA sequences that occurs next to the gene for an odd-looking RT. After further analysis and testing in our lab, we recognized that this pattern marked a previously uncharacterized enzyme system found in bacteriophages (the viruses that infect bacteria) that we call array-associated reverse transcriptases (ART)."
It isn't a convenient gotcha. It's about what the people pushing the given thing are intending.
Person demoing something they made is usually trying to hide the fact they had claude built it and sell it like they didn't. This sort of person often lacks the technical skills to vet that what claude actually produced is actually working as they expect. Hence the snark.
On the other hand, with anthropic's case, they are trying to say "claude did this, how smart it is" while trying to downplay the fact that they needed it to be steered by domain experts to produce anything worthwhile.
This framing overlooks an unstated caveat - i.e. people that work for an LLM company have an incentive to minimize human contribution as much as possible in their narratives.
If he has such a good reputation, why doesn't he write well? He never actually introduces the point of the article, he merely alludes to it in the concluding paragraphs.
reply