you are viewing a single comment's thread
view the rest of the comments
[–] 39 points 1 day ago (16 children)

And actually after watching the video and thinking, you dont even need AI for that. Classical computer programs do generate bones good enough for you just to edit them. You could solve it with just math to "free artists"

  • source
  • parent
  • hideshow 16 child comments
  • [–] 58 points 1 day ago (15 children)

    I work with a lot of software and am seeing this everywhere - problems that would be efficiently and cheaply solved with classical approaches but are solved with slow and expensive inference.

    Just recently I saw an llm call to extract a postal code from a structured input. Regex would be trivial to solve it, but the person who set it up used fking LLM. What a joke.

  • source
  • parent
  • hideshow 15 child comments
  • [–] 26 points 1 day ago (2 children)

    There was recent drama around a completely vibe coded attempt at multiplayer Daggerfall. All of the prompts were leaked into the original github repo. One of the best was when Claude was asked to change an inventory icon for a fish. Claude's response was to create from scratch a pixel art editor just to spit out a 36x24 PNG of a rotten fish.

    And these guys wonder how they can possibly be burning through literally thousands of dollars in tokens every month 🙄

  • source
  • parent
  • hideshow 2 child comments
  • [–] 16 points 1 day ago (3 children)

    I kinda wonder how they end up there. Couldn't they have just asked the bot to make the regex code if they were already using it? I imagine I am just being fucked by my biases here cause I would already to know to ask it for that.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 10 points 1 day ago

    Yes, they could have. Easily. They didn't think to ask.

    That's the limitation is this LLMs, they can action if the instructions are sufficient, but layer 8 is still involved, and the l8 root cause persists.

  • source
  • parent
  • [–] 7 points 1 day ago

    Also, the AI companies aren't actually incentivized to create LLM that efficiently and permanently solve problems. They are incentivized to get people addicted to using LLM so they can charge them for it forever, endlessly increasing token costs and scaling up their datacenters until all matter in the universe has been consumed and every dollar of wealth has been deposited in their bank accounts. It is actually a huge win for them when an LLM creates a solution that just inefficiently uses more LLM tokens every time it is used. It's like selling people fuel that steadily increases your engine's fuel consumption over time. It's genius, frankly. I am surprised they haven't tried that yet, too.

  • source
  • parent
  • [–] 2 points 1 day ago (1 child)

    Yeah I've run into this while playing with models/ agents in my homelab. Anything I've thought about using them for I've come to the conclusion that in every case I've got in my lab they can all be completed with shell scripts so why waste the resources?

  • source
  • parent
  • hideshow 1 child comment
  • [–] 4 points 1 day ago (4 children)

    To be honest, I’ve wasted entire evenings on regex. Fuck regex. We need a replacement that isn’t an llm but fixes that archaic shit. Make it human language

  • source
  • parent
  • hideshow 4 child comments
  • [–] 3 points 1 day ago (2 children)

    Yeah, that’s fair. If you have one problem and solve it with regex, now you have two problems.

    Though in this case the monthly inference budget ran out causing a system outage. I was on call, so, getting a page off hours makes me dislike the llm approach even more haha

  • source
  • parent
  • hideshow 2 child comments
  • [–] 0 points 1 day ago (1 child)

    If being out of tokens causes an outage, you have architectural problems

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 1 day ago

    On the contrary. If you architect to not fail when missing context critical to customer experience your architecture is irrelevant because you're losing customers. The issue was the Layer8 decision to use inference when regex it's a better fit.

    But lol imagine making deep assumptions when you don't even know what the product is.

    Fail fast. Fail loud.

  • source
  • parent