22 votes

Disrupting a covert Iranian influence operation

9 comments

  1. [9]
    updawg
    Link
    I find it interesting that these things never seem to find meaningful engagement. Are they not finding the good ones? Are they not reporting the good ones?

    I find it interesting that these things never seem to find meaningful engagement. Are they not finding the good ones? Are they not reporting the good ones?

    7 votes
    1. [7]
      teaearlgraycold
      Link Parent
      The smart ones would not use an LLM over a public API.

      The smart ones would not use an LLM over a public API.

      11 votes
      1. [6]
        updawg
        Link Parent
        Yes and no. I don't know how the ones you can use on your own machine compare, but the smart ones know all you need is a new email account to make more content with the major LLMs. But I was also...

        Yes and no. I don't know how the ones you can use on your own machine compare, but the smart ones know all you need is a new email account to make more content with the major LLMs.

        But I was also referring to other sites. Obviously Facebook has a major issue with foreign disinformation campaigns and Twitter has issues with every kind of information, regardless of prefix, but most sites' reports that I've seen (e.g., reddit) indicate that the coordinated accounts don't achieve that much engagement.

        Due to the large number of obvious disinformation bots I've encountered in the wild in subreddits that aren't remotely political, I'd suspect that they're not publicly reporting the more successful campaigns that they catch and are sweeping most of the less successful ones under the rug.

        10 votes
        1. [5]
          Weldawadyathink
          Link Parent
          I primarily use LLMs for coding, but I use local LLMs pretty extensively. For most use cases, I find the local models to be better than the proprietary models. Deepseek coder is absolutely the...

          I primarily use LLMs for coding, but I use local LLMs pretty extensively. For most use cases, I find the local models to be better than the proprietary models. Deepseek coder is absolutely the best code LLM I have used, and there is a V2 out there now. I haven’t tried GPT 4O yet though. For general usage, llama2 was pretty decent and llama3 is very good. I haven’t paid for chatgpt pro in months, and I don’t feel the need to resubscribe.

          2 votes
          1. updawg
            Link Parent
            You don't need to subscribe to use 4o.

            You don't need to subscribe to use 4o.

            1 vote
          2. [3]
            gianni
            Link Parent
            Can you share more on how you use LLMs for coding? I use Github Copilot but it’s really only useful for cursor-to-end-of-line completion—which is neat, but doesn’t actually help or save me much....

            Can you share more on how you use LLMs for coding? I use Github Copilot but it’s really only useful for cursor-to-end-of-line completion—which is neat, but doesn’t actually help or save me much. The chat is next to useless. I’ve tried asking GPT more architectural questions but it hasn’t been particularly fruitful.

            1. [2]
              Weldawadyathink
              Link Parent
              Yeah sure! I don’t really use copilot or similar services, so I can’t comment on using those. There are a few main flavors of my interactions with LLMs. First is an improved web search. Google has...

              Yeah sure! I don’t really use copilot or similar services, so I can’t comment on using those.

              There are a few main flavors of my interactions with LLMs.

              First is an improved web search. Google has been bad for years, but even if it was as good as it used to be, I think an LLM would be better for this usage. This is for things I know, but don’t remember, or know how to do in one language but not another. For example, open a new SQLite database, import a csv, add a few indexes, and select some data. I am familiar enough with sql now that I might be able to do this without checking documentation, but I’ll probably get a few syntax issues. Why bother checking the syntax documentation when I can just have an LLM do it for me? For this category of question, I know enough about the subject matter to know if something is wrong, so I can spot hallucinations pretty reliably. Even if I can’t, I get a syntax error, tell the LLM about the error, and get a fixed query/code.

              Second is finding out about what questions I don’t know to ask. I don’t know how to describe this well, but I’ll try. If I have a problem that I know is a common programming problem, but I don’t know what the industry commonly uses to solve that problem, I’ll just start chatting with an LLM. For example, how do you handle api requests in react? That isn’t the best example because it’s a well documented problem space, but let’s go with it. With a Google, I’ll get a stack overflow answer that, if I am lucky, has a sample using useEffect and an example with tanstack query. More likely I’ll get a useEffect sample and someone complaining about all the issues it could cause. An LLM will give me both examples and, if I ask, explain the advantages and drawbacks of each. As I have been learning, it is incredibly helpful for me to see all the bad ways of solving a problem before using the good one. I have build entire websites using useEffect for api requests before I used react query or tRPC, and I understand how everything works better because of that.

              Last is scaffolding something that works without too much effort. This is pretty much just a nudge in the right direction. For my programming style, I need to see something working at every step and slowly improve it. So an LLM will get something output to the screen quickly. It will very likely be wrong, but it will likely be functional without too much effort, and I can start improving it from there. This is the same idea as shadcn/ui, which I absolutely love. It’s a set of react components where you import the code directly into your project instead of as a dependency. When you need to change behavior or modify the functionality, the code is in your repo ready to modify. I recently made a DataTable using the shadcn guide. I was able to get minimal functionality in a few minutes and iterated through improvements. This process taught me better than any guide what to do and how to do it. For general coding questions, an LLM lets you get that minimal functional version working quickly.

              4 votes
    2. Fiachra
      Link Parent
      I'm not sure they're trying to get engagement. They might just be aiming to skew the perceived popularity of things in the eyes of lurkers.

      I'm not sure they're trying to get engagement. They might just be aiming to skew the perceived popularity of things in the eyes of lurkers.

      9 votes