• 1 Post
  • 11 Comments
Joined 10 months ago
cake
Cake day: June 30th, 2025

help-circle



  • To me agentic ai seems to be a futile attempt at making llms useful in a work context. The idea of having virtual workers who will accomplish tasks and lift their own weight seems appealing until you realize not even hiring actual human workers increases throughput until they can get their bearings. Tools that consistently and accurately do repetitive things is more valuable for an individual than an open ended tool with the potential to solve it all in one go imo.

    I find it hard to believe that llms trying to cover up for their weaknesses with increasingly token intensive methods like thinking or planning will stay economically viable after the “capture the market” phase of the ai industry. It is remarkable that such methods work at all. I can’t imagine there’s nearly enough training data about non-final work or thought processes or planning that went behind producing something, not to mention people might not accurately describe how they reached their solution even if they try to. And even if they manage to print those thoughts into their context, llms don’t produce words through a thought process so it’s dubious how much benefit they can ultimately obtain.

    I think once the ai craze is over people might make tools that use machine learning to automate tasks but I don’t think the repackaged chatbots are it.

    I’ve tried agentic coding using a bunch of llms from ollama couple weeks ago, most couldn’t manage to consistently find the correct file, glm4.7 got pretty far but lost context and produced some irrelevant code.