I’ve found that AI has done literally nothing to improve my life in any way and has really just caused endless frustrations. From the enshitification of journalism to ruining pretty much all tech support and customer service, what is the point of this shit?

I work on the Salesforce platform and now I have their dumbass account managers harassing my team to buy into their stupid AI customer service agents. Really, the only AI highlight that I have seen is the guy that made the tool to spam job applications to combat worthless AI job recruiters and HR tools.

  • NateNate60@lemmy.world
    link
    fedilink
    arrow-up
    8
    ·
    edit-2
    7 hours ago

    An LLM (large language model, a.k.a. an AI whose output is natural language text based on a natural language text prompt) is useful for the tasks when you’re okay with 90% accuracy generated at 10% of the cost and 1,000% faster. And where the output will solely be used in-house by yourself and not served to other people. For example, if your goal is to generate an abstract for a paper you’ve written, AI might be the way to go since it turns a writing problem into a proofreading problem.

    The Google Search LLM which summarises search results is good enough for most purposes. I wouldn’t rely on it for in-depth research but like I said, it’s 90% accurate and 1,000% faster. You just have to be mindful of this limitation.

    I don’t personally like interacting with customer service LLMs because they can only serve up help articles from the company’s help pages, but they are still remarkably good at that task. I don’t need help pages because the reason I’m contacting customer service to begin with is because I couldn’t find the solution using the help pages. It doesn’t help me, but it will no doubt help plenty of other people whose first instinct is not to read the f***ing manual. Of course, I’m not going to pretend customer service LLMs are perfect. In fact, the most common problem with them seems to be that they go “off the script” and hallucinate solutions that obviously don’t work, or pretend that they’ve scheduled a callback with a human when you request it, but they actually haven’t. This is a really common problem with any sort of LLM.

    At the same time, if you try to serve content generated by an LLM and then present it as anything of higher quality than it actually is, customers immediately detest it. Most LLM writing is of pretty low quality anyway and sounds formulaic, because to an extent, it was generated by a formula.

    Consumers don’t like being tricked, and especially when it comes to creative content, I think that most people appreciate the human effort that goes into creating it. In that sense, serving AI content is synonymous with a lack of effort and laziness on the part of whoever decided to put that AI there.

    But yeah, for a specific subset of limited use cases, LLMs can indeed be a good tool. They aren’t good enough to replace humans, but they can certainly help humans and reduce the amount of human workload needed.