It's amazingly vacuous isn't it? I think the most interesting read was the fact that they were surprised llama.cpp crashed when they used a bad set of commandline arguments.
Although in the section immediately above the observation they claimed that they ran 10 whole completions with 100% success rate. So who knows.
I have to admit I slightly miss the flood of AI-psychosis research papers that seemed to be popping up a couple of months ago. Good to know there's still one or two new ones floating around.
Hi! Big fan of OpenRouter and the data you provide. It'd be awesome if you would consider providing volume of tokens per hour, mostly for my own curiosity as to quite how peaky demand is.
How is it surprising to people that zip and XML are in stdlibs for a programming language?
Btw, you should have looked at dotnet for this as well. There is a very good library ( DocumentFormat.OpenXml) that can handle all docx/xlsx/pptx files. And dotnet can ship standalone binaries (though AOT probably won't work).
Many runtimes/languages rely on third party deps for that. Also plenty of devs think the stdlib should be as lean as possible.
Personally, I think there should be a balance. The direct consequence of a barebones stdlib is NPM and having to download hundreds of dependencies for a hello world.
Golang has the golang.org/x packages, which avoids too much stdlib bloat while still providing the niceties of “pre-vetted” packages that don’t pull in a massive dependency tree.
Difficult to square the author's surprise with the later comment "I have my fair share of building a Java desktop application and know jpackage and alike very well"
You can't get very far in Java development without working with .jar files (which are zip archives).
They can (not entirely sure how 'grey' market this is) either have subsidiaries outside of china (eg: singapore) that provide the inference and/or just rent it off the public gpu clouds.
Two thoughts (I was in the same situation, constantly trying desktop Linux then pinging back to Windows after hitting issues).
1) Fedora is really worth a try, it's extremely polished. The best thing is the packages in the repo are generally much more up to date that debian based distros, which maeans less random PPAs to work around it, which cause issues.
2) The biggest change is having Claude Code/Codex able to diagnose and tweak things extremely quickly. If something goes wrong, I ask claude code (in a specific folder with various docs about workarounds) and it goes and fixes it 99% of the time very quickly.
Coding agents being able to fix Linux actually makes it _more_ stable than Windows for me. In my experience Windows is less buggy _in general_ than desktop Linux.[1] However, once you hit random issues you are basically screwed if basic attempts don't work. With Linux you can have a coding agent go thru all the reams of logs to find the issue and even clone the underlying source code to find issues.
Similar to this, but NixOS. Having AI to help me through made this so much easier. I was sold on the idea of an 'IaC' config based machine, after a general push to move all of my processes towards full GitOps. Windows had been pretty good to me (running 4 years on the same laptop install), but it had started overheating with fans sounding like a vacuum cleaner and it was time to start over. The difference with NixOS is that sure there are issues and preferences to work through, but when I fix them, they get committed in config, and thats an investment in time rather than something that will be right until next time I have to do it. I was able to reproduce, and rebuild on a separate machine with minimal hassle (it's good to go through this process to be sure you've got it right) and that really was amazing to see. SSH keys, SMB share, monitor configurations, themes, apps, utilities - a fully set up dev machine, everything is just there. I've been planning changes on a Copilot integrated taskboard I built (https://www.agentkanban.io) and then handing them off to the agent, reviewing the changes in the VS Code git client and then apply, commit. Being able to see the number of commits makes me realise how much I was doing manually, every time I set up a new machine.
In my first years with NixOS, I really wanted to share my enthusiasm and joy with others. But I often got caught up explaining the mechanisms, the insights that it's built on. It took me a while to come around to realizing the heart of it: NixOS feels really good to use because it frees you from the uncertainty of a machine that suddenly changes underneath your feet. It gives you a system that is really easy to inspect and modify. It gives you the ability to move forward without fear. It's really liberating and relaxing in a certain way.
Historically, a sense of intimidation and the difficulty of onboarding blocked a lot of people from accessing these benefits. It's been cool to see how LLM agents have helped a lot of people get started and tap into the fun parts.
My experience with explaining NixOS is that once you get to the benefits people get intimidated and start telling how they never have these issues.
I think immutable distros/software are the most intuitive way, even if people get intimidated with the idea. What's the action you did as a junior when windows/linux/x broke? You most likely reinstalled. When encountering issues we tend to try to emulate immutability.
The sad thing about NixOS is that there's 0% chance of seeing it used at work. Even in the devops people you're lucky if people can edit a Dockerfile. Good luck spending the next 3 years explaining it might look like json but it's not, and it would help to learn the syntax.
We used nix for developer environments, and nix to define docker containers for our production services. The biggest problem is the “nix people” can become blockers if there is an issue and they aren’t around to figure it out, because most engineers are just not going to learn nix.
However, we felt the benefit was worth the cost, overall, especially because we used the same flakes and therefore versions across all dev and deployed envs, and we didn’t have to deal with the hassle and performance issues of running all the dev services in docker containers.
I'm the Nix guy on my small team at work and I'd love to train any of teammates on Nix stuff. One of them already has some Nix experience and I make sure to walk him through my Nix work whenever I finish something new. But it seems like no one is really interested in truly diving in like I did when I was new to Nix.
I don't really know why that is, because it doesn't seem to have to do with intelligence or general experience with adjacent skills. I think it's probably at least 30% down to temperament, but it may also be because my team is mostly fairly senior and each person has a lot of responsibilities of their own.
On the bright side, the tooling is easy enough that nonetheless, my team generally has no problems spinning up new Nix environments without me, or making small changes to the ones I've set up. And LLM agents are now good enough with Nix that I'm confident they can unblock themselves until my return whenever I'm on vacation or whatever.
Yeah, similar experience here. One pattern that was helpful for us was defining a nix file that just contained a mapping of names to lists of packages: dev, ci, and so on, so that if you wanted to install something that was on the nix search page, you could just go toss it in there and immediately get it locally. It also made it clear which package sets would wind up where, and allowed recursion so you could say that like dev was rust-base ++ rust-dev ++ js ++ common, or whatever.
Currently no NixOS at work for me, but my team is using Nix for local development environments and CI successcully via devenv.nix. I think we actually could use NixOS on my team except that we have a host of endpoint security software that we need to run on all cloud VMs, and I'm a little distant from the people who own the automation that installs it on approved AMIs.
I can second your comment about Fedora. One underrated point: Their drivers are always the best amoung all free distros. I'm not sure if there team is the most aggressive to upstream new drivers, or manufs target Fedora first. And I write this post as a regular Debian user. However, if I have a piece of hardware that has terrible drivers on Debian, almost always it will work fine on Fedora. (Grumble!)
Genuine question, how is everyone else dealing with giving agents elevated permissions? Obviously the answer is "don't", but some things are pretty harmless, like journalctl and dmesg, and are pretty useful for debugging the system.
I guess you could make a new user to run the harness under and give it no-password sudo rights for select commands? That doesn't feel like a great solution but it's the only thing I can come up with
I give it passwordless sudo but don't give it permissions to automatically run commands. I just review what it's running before accepting. If it looks questionable, I interrupt and ask it to explain what/why it wants to run the thing
I use sudo -A with some openssh ui for sudo. I tell the agent to use sudo -A for anything that it needs and then it pops up with a sudo password prompt.
Using Linux is a learning experience. You will inevitably face and solve numerous problems over time, but every time you do, you come out of it understanding what's going on under the hood a little more.
Still, it can be dreadful to face even small issues when you only feel like using your computer and not fixing it. Having an LLM agent help with fixing issues is a lifesaver. Ask it what you don't understand, take note of the commands it uses or suggests while troubleshooting and fixing your issue, and you'll supercharge your learning and not get as hung up on the issues.
If someone doesn't care much to learn though, I'd say Linux is still tough to recommend.
As someone who's done the whole Helpdesk -> Sysadmin stuff, working around other "Windows Admins" they don't learn much, all they know is "Deleting this thing seemed to fix" "Reboot". They don't get any of the fundamentals of what caused an issue, or how to diagnose them properly.
I spent a lot of my career in the Windows admin space. That isn't true. Most Windows admins are sharp dudes, just like most Linux admins are sharp dudes. And a minority of Windows admins are slackers who don't try to learn, but a minority of Linux admins are slackers who don't try to learn.
Most of my experience with Windows has been googling for solutions rather than understanding the underlying system. So I guess it's a learning experience, but not in learning Windows itself, as much as learning to use search engines.
I do remember common fixes for various things, but not much of it can be extrapolated to other issues in my experience.
Currently, if you use claude -p (non interactive mode) in for example CI/CD, you can use your included subscription tokens.
They are now changing it to be:
You get $20/$100/$200 of "credit" that can be used for claude -p. Problem is, once you are out of that it is the normal API rates (outrageously expensive).
I get this, but isn't this a complete compliance failure?
What's the point of having all those loops to onboard vendors if you can just buy from AWS marketplace (which AFIAK is not a particularly high bar to achieve for SaaS options)?
Like imagine $POOR_QUALITY_VENDOR. If they go through the normal channels they might get shot down. If they get procured on AWS Marketplace, then it feels to me in many organisations 'its fine', though AWS does minimal checking?
Powerline is in my experience vastly worse than WiFi in nearly all cases. It's slow, suffers from bad jitter/interference (often worse than WiFi) and the chips run so hot (especially the last gen ones, AV2000 iirc - I believe they don't sell them any more because they overheat and fail, or at least 2/2 of the ones I had did this).
Even with many walls I was getting 300-400mbit/sec on WiFi vs 100mbit/sec on powerline.
reply