Grimburger

Wild you're getting downvoted here, apparently using LLM is fine for reviews but hacking the system and getting noticed isn't?

Is the problem not they're using AI for review and court decisions that can easily get hacked? If I open ports that can get hacked or allow sql injections I'll lose my job straight away tomorrow, but apparently the justice system is immune from that?

I'm all on board for them using LLM's and AI for law as its prime for this stuff, but its a matter of reading the output and people clearly aren't doing that I see it at my job all the time in software development.

Before most of the job was writing code, now its reading it, this is the IT industry in a nutshell, you need to adapt and deal with changes how is it any different from all the new frameworks and paradigm shifts at all? I don't understand why people cant deal with that, it's how it always was, things were always changing this is just another iteration of that.

Sounds like the price in Canadian or Australia dollars

Hackernews, Lesswrong and 4chan

> I was trying to use rocm with llama.cpp

completely offtopic but is rolling with rocm worth it? I spend a fair bit monthly on rental gpus for projects and going to upgrade at home instead, AMD has some solid winners here pricewise but get conflicting reports about using it for ML in 2026.

once upon a time it seemed unthinkable to use anything but nvidia but seems to have come a long way since I last looked, probably would be just pytorch and gemma 31B

I get the feeling the situation is only going to improve longer term so might be a good time to just do it

The amount of supercars and high end vehicles in Hsinchu is unreal. Probably all anyone needs to know here, until you've seen it firsthand you wouldn't understand.

Have seen this across the planet, it's not unique to Europe, I end up adding youtube to the search query because instagram or tiktok isn't what I'm looking for ever but that makes up the bulk of the results, instagram won't even show me the video without a login and is at the top wtf

Poisoning the well with fake/made up reddit posts is incredibly common these days.

Certain people have realised that if you start feeding the machine this poisoned data eventually it reaches millions of eyeballs. It's the new SEO.

>but rather to everyone quietly reading comments

as someone who reads a lot but doesn't comment much you've got the topshelf content here consistently, its so obvious when someone has a clue or doesn't if you're in the same areas, very easy to fool others with confident assertions (cough airline issues on HN) but for someone in the industry it's not at all

funnily enough this comment about zoning which isn't your or my field but get the feeling it's on point, is such a massive problem in the western world which has built up over decades, lets just copy someone elses homework in my opinion, Asia gets this stuff right: https://devonzuegel.com/north-american-vs-japanese-zoning

Will you buy a computer chip with the same specs that costs 70% more? Please be honest with us all.

You'll do that because of national adversity?

I don't want to speak bad about them but it's just general demand for semiconductors at this point, everyone is just buying it all up because they can. The Arizona plant still doesn't seem to be pumping out and the self-sufficiency target for the US doesn't seem all there. Im sorry to be the bearer of bad news but honestly it's a culture that's hard to replicate and its how it somehow gets done in one country on Earth. The 2030 target for the states surely isn't happening, it's way too ambitious.

People really don't understand how this stuff is borderline magic turning silicon into these things. There's so much that can go wrong at every step, it's the most advanced manufacturing process on Earth.

Once upon a time chips went down in value but I don't think that's happening anymore, lot of other people seem aware of that. Not to mention the semiconductor talent is getting eaten up like pancakes everywhere.

But that's sorta of hits to the heart of my comment, it's not particularly working well. TSMC has 2 fabs in China and yields+costs are a fraction of whats coming out of Arizona, let alone in Hsinchu.

Like are you really going to buy a chip that costs 70% more out of national pride? Might work on some but think that very few else will sign up for that.

They've got huge monetary losses, if you're ok with that sure but its not like any other chip fab on the planet is losing cash.

The Taiwanese are doing a lot to try get it up but all the reports on the ground level are rough.

https://www.tomshardware.com/tech-industry/semiconductors/ts...

I get its sarcasm but TSMC has done so much to try to get the Americans getting some decent yields in their new factories and it's failing so hard, there's real questions about whether they can ever meet their targets. They've sent thousands of existing Taiwanese workers to help at this point and from anecdotal things I hear it's not going well. (I live in Taiwan)

This stuff is all blood, sweat and tears to get it done, and to be honest a culture.

You can't just buy it off the shelf. If that was possible everyone would have gone paid for it like an Ikea product years ago.

What does it matter?

All supply is supply, despite what certain folks seem to believe.

This is very cool, why not directly test the exploitability direct with harmless pods on the cluster for verification? If it's only model based as a defender I'd worry about how many % of misses arise. Even frontier models are going to miss some stuff, a programmatic gate seems prudent.

Not trying to be negative, do think it's a good approach youve got, but the modern reality of dealing with cve's and compliance is to just fix them because it's a massive headache trying to write exemptions for all the ones that don't matter. Been my experience anyway.

How do you know there wasn't account sharing though?

I've seen some amazingly dodgy stuff when hiring people from south east Asia, sharing a paid account with friends worth a months rent there seems milquetoast in comparison.

You also used newlines to write 23 words here, something non-existent above.

A few hundred personal dick pics and they'll be handing it back quickly.

Still a struggle to get coworkers to write docs or tests despite it being effortless.

At the end of the day its a mindset.

Am I the only one suspicious of your comment?

> Write as yourself.

Unless it's your first day on HN (it's not 2008 account), most people use paragraphs and sentences longer than 2-5 words. Something like "You, directly." in the middle of all that has a certain vibe to it.

> Copying and pasting Claude responses... directly shifts the burden of comprehension

Most definitely.

The problem is we live in a world where people try to pass off copy pastes as their own all the time, many even smart enough to remove the em-dashes now. So I don't think that's going away soon. Your first sentence seems human, the rest doesn't match the tone at all.

Tsunamis are only an issue close to shore. In the ocean it's barely a mild wave.

In a boat going further out is the safest thing you can do if there's a tsunami warning.

Cloudflare specifically has a block for LLM and AI training bots now.

Not sure of the effectiveness but it's there.

You create your own image first rather than blank ubuntu.

Packer and Incus setup:

  packer {
    required_plugins {
      incus = { version = ">= 1.0.0", source = "github.com/lxc/incus" }
    }
  }

  source "incus" "ubuntu" {
    image        = "images:ubuntu/24.04"
    output_image = "ubuntu-claude"
    type         = "virtual-machine"
  }

  build {
    sources = ["source.incus.ubuntu"]

    provisioner "shell" {
      inline = [
        "curl -fsSL https://claude.ai/install.sh | bash",
        "echo 'export PATH=\"$HOME/.local/bin:$PATH\"' >> ~/.bashrc"
      ]
    }
  }

Then:

  packer init ubuntu-claude.hcl
  packer build ubuntu-claude.hcl
  incus launch ubuntu-claude my-claude-vm
  incus exec my-claude-vm -- claude -p "solve the EC discrete logarithm problem, if it doesn't work keep going" --dangerously-skip-permissions

The reality of these things are that eventually you will want to do something useful or different with them and the flexibility simply isn't there compared to a real vm, if you desperately need boot times then there's plenty of other options here, especially with packer. Some of my prompts are often hitting 60+ minutes so it's not really something I think about.

Haven't used docker sandbox but you can't just `apt install postgres` on firecracker, it needs to get baked into the image first.

That's my experience anyway, there's a lot of restrictions once you need to do some real basic things. For basic prompts maybe but interacting with a full stack ehh.

So bit hesitant to call firecracker a real VM myself.

> I’m afraid the LLM will heavily customize its environment and I’ll be unable to replicate it

Would suggest Hashicorp Packer or cloud-init for deterministic images, not hard to setup or use. LLM's have little problem with them either I find.

If you need quicker environment rebuilds consider using something smaller like alpine as the base, though once you setup a golden image even heavy things like debian are fine.

> Each agent runs inside a dedicated microVM with your dev environment

What's a "microVM" and what's the security model here compared to using real virtual machines with actual constraints on breakouts?

Is it marketing fluff?

Incus/LXD has had VM's for a long time now.

    incus launch images:ubuntu/26.04 my-ubuntu-vm --vm
    incus exec my-ubuntu-vm -- bash

This is an excellent comment and I don't want to diminish by adding jokes but this totally reminds me of the hitler uses kubernetes meme. You can do everything right for HA and there'll still be something to trip you up.

https://www.reddit.com/r/kubernetes/comments/f2bcsu/parodyfu...

The lawyers cost a lot more than what you are going to get back from GP in a copyright lawsuit.

Try being on linux and using zsh, it constantly fucks up scripts. Having it run commands via SSH into a windows machine is even more a comedy of errors.

I have to put in every claude.md that the shell is zsh. It's reached the point of annoyance I might just go back to bash.