> So this is proof of the models actually getting stronger (previous generations of LLMs were unable to solve this one).
No, it's not.
While I don't dispute that new models may perform better at certain tasks, the fact that someone was able to use them to solve a novel problem is not proof of this.
LLM output is nondeterministic. Given the same prompt, the same LLM will generate different output, especially when it involves a large number of output tokens, as in this case. One of those attempts might produce a correct output, but this is not certain, and is difficult if not impossible for a human not expert in the domain to determine this, as shown in this thread.
As others have pointed out, a key part of the prompt used here may have been "don't search the internet" as it would most likely have defaulted to starting off with existing approaches to that problem...
Long breaks help. Take your mind off of things that bothered you. Do things you enjoy. Which may include tech work, but on your own terms.
I wouldn't be surprised if you decide to not go back. The status quo of most organizations is grim. But there are still people who care about the same things as you. You can seek them out and work together, much like you did 15 years ago. This is more difficult now among the noise, but you can tune that out. The industry will never recover altogether, but this current period is a blip of high insanity, which will subside in a few years.
As disturbing as the film "Midsommar" was, I found the concept of a human life being divided into 4 seasons of 18 years each pretty compelling. Not necessarily that life should end after Winter, but a person's contributions to society probably should. Having politicians in office pushing 80 is a disgrace.
For crying out loud, why are we discussing and paying attention to articles and claims about a product that doesn't even exist yet?!
If this isn't a sign of a bubble, where marketing is more important than the actual product, I don't know what is. This industry has completely lost the plot.
That's a wild statement. The CLI is just another UI.
The problem in this case is JS and the NPM ecosystem. Go would be an improvement, but complexity is the enemy of security. Something like (pass)age is my preference for storing sensitive data.
I built something similar recently on top of Incus via Pulumi. I also wanted to avoid libvirt's mountain of XML, and Incus is essentially a lightweight and friendlier interface to QEMU, with some nice QoL features. I'm quite happy with it, though the manifest format is not as fleshed out as what you have here.
What's nice about Pulumi is that I can use the Incus Terraform provider from a number of languages saner than HCL. I went with Python, since I also wanted to expose a unified approach to provisioning, which Pyinfra handles well. This allows me to keep the manifest simple, while having the flexibility to expose any underlying resource. I think it's a solid approach, though I still want to polish it a bit before making a public release.
I took a slightly different approach in that I don't want to use YAML as the authoritative source. Many projects abuse it, and end up creating a DSL on top of it with all sorts of hacks to achieve the flexibility of a programming language. Pulumi and Pyinfra already provide user-friendly primitives and idempotent(ish) APIs that work much better than YAML. I simply want to expose some (opinionated) building blocks to make them easy to use, and allow users to customize them and add their own as needed. E.g. I definitely don't want to write any shell scripts inside YAML. :)
BTW, Pulumi already supports YAML[1], which can be used with any provider. But to me it's too verbose and generic, and of course, it lacks the provisioning primitives.
Your path makes a ton of sense too! You get typed languages, state management, and the Incus team's work on the QEMU layer. The tradeoff I wasn't willing to make is the daemon + state store: Incus wants to own the VM lifecycle the way libvirtd does, and once you have that you're back to "two sources of truth" if you ever shell out. Holos is deliberately stateless on the host; Everything lives under one directory per project, rm -rf is a valid uninstall (though it will abandon the VMs if running). Different answer to the same frustration. Would genuinely like to see your thing when it's public.
This is exactly what I dislike about incus. But what I do like about incus is how I can easily spin up and configure VMs directly using the CLI, without preparing a config first (I hate yaml).
So would be nice if holos could replicate that docker/incus CLI functionality, like say "holos run -d --name db ubuntu:noble bash -c blah".
Docker does have both docker run and docker compose... Holos is compose-only right now. A holos run on top of the same VM machinery isn't a huge lift; mostly it's deciding what the sensible defaults are (image, cloud-init user, port, volume scope); Noted. it's not in 0.1 but I hear the ask.
> Incus wants to own the VM lifecycle the way libvirtd does, and once you have that you're back to "two sources of truth" if you ever shell out.
That's true. But I didn't want to reinvent what Incus or any hypervisor abstraction does. I simply wanted to add some sugar on top that allows me to easily declare infra using small abstractions, and to tie in the provisioning aspect along the way. I still use Incus directly, and can benefit from their work, as you say. State is also managed by Pulumi, so really, there are 3 places for it to exist. There are some challenges with this, of course, but I think the tradeoff is worth it.
Good luck with your project, I'll be keeping an eye on it. I'll probably make a Show HN post when I release mine. Cheers!
If I have evidence that a crime has been committed based on my layperson understanding of the law, I will surely inform others before the case is even brought to courts. Journalists can and should do the same.
By your logic, reporting based on evidence provided by whistleblowers shouldn't exist. Things like Watergate would likely have never happened.
Journalists shouldn't accuse anyone of committing a crime, and goes without saying that facts shouldn't be fabricated, which is unfortunately common nowadays as well, but they should report events that happened based on the information they have, whether these happen to be related to crimes or not.
>If I have evidence that a crime has been committed based on my layperson understanding of the law, I will surely inform others before the case is even brought to courts. Journalists can and should do the same.
In the US, careful journalistic organizations follow ethical and legal guidelines that often split hairs.
Have a look here: New York Times - Ethical Journalism
A Handbook of Values and Practices for the News and Opinion Departments
Reporting based on evidence is definitely allowed in the UK. Any accusation of libel/slander could be defended by producing the evidence and thus proving that the statements were true.
Going beyond the evidence and jumping straight to the crime is where the situation becomes tricky as the defense would be unlikely to prove beyond doubt that the accused person was actually guilty - that's why terms are used such as "alleged child abuser". Alternatively, the evidence/facts can be reported e.g. "Trump featured in many victim reports as an abuser".
I can't say whether this was machine-generated or not, but the reason LLMs use these patterns is because they're often used by humans, which is what they're trained to mimic. LLM spam has now made it annoying, but there are many people who still write like this. Asking them to change their writing patterns because LLMs have ruined it for readers is not just unfair—it's offensive. (See what I did there? Double whammy!)
> I can't say whether this was machine-generated or not, but the reason LLMs use these patterns is because they're often used by humans, which is what they're trained to mimic.
It's definitely written by an AI. I understand that people use these same rhetoric devices but the word "mimic" is exactly right. They're not writing like humans, they're mimicking human writing in a way that feels extremely uncanny.
Thank you for saying this. I use those writing devices a lot, always have since I was young. I've always said I don't write like an LLM, the LLMs are copying me. Its also a common hallmark of neurodivergent writing, and it's frustrating to frequently get dismissed for being an LLM just because of writing patterns?
I'm now having to deliberately re-word my emails and comments, spending additional time, to avoid being accused of being an LLM.
A major reason for that is because there's no way to objectively evaluate the performance of LLMs. So the meme projects are equally as valid as the serious ones, since the merits of both are based entirely on anecdata.
It also doesn't help that projects and practices are promoted and adopted based on influencer clout. Karpathy's takes will drown out ones from "lesser" personas, whether they have any value or not.
This Mozilla?[1] The company whose 85% of revenue depends on an adtech giant?
They're certainly doing better than others in this space, but their track record does not inspire confidence for anyone concerned about their privacy and data.
No, it's not.
While I don't dispute that new models may perform better at certain tasks, the fact that someone was able to use them to solve a novel problem is not proof of this.
LLM output is nondeterministic. Given the same prompt, the same LLM will generate different output, especially when it involves a large number of output tokens, as in this case. One of those attempts might produce a correct output, but this is not certain, and is difficult if not impossible for a human not expert in the domain to determine this, as shown in this thread.
reply