At Meta, Microsoft, Salesforce and other large companies, devs are purposefully burning tokens (and money!) to inflate their AI usage and hit AI usage metrics which they treat as targets.
  • uenticx@lemmy.worldEnglish
    6·
    3 天前

    If you use “maxxing” as a suffix, I instantly hate you. You’re already vibe-coding, which probably means you couldn’t pour piss out of your boot if directions were slapped on the heel.

  • PerogiBoi@lemmy.caEnglish
    14·
    3 天前

    One of the KPIs of my job is how well I integrate AI into my team. My boss automatically gets my AI usage from reports on the admin centre of our M365 instance. I was asked last week why I had 2 days of zero Copilot queries. I had to type up an explanation lest this gets recorded as me being non compliant.

    Thinking I will need to build an agent that does some tangentially related questioning daily so that I don’t get written up or cross examined again.

    • NekoKoneko@lemmy.worldEnglish
      4·
      3 天前

      I’m hoping we hit “peak stupid” with these nonsense metrics soon, but the C-suite MBAs who get all their AI-related information solely from vendor demos, Sam Altman and Elon Musk keep surprising me.

      • PerogiBoi@lemmy.caEnglish
        4·
        3 天前

        We will hit peak stupid. I’m just wondering how many more years it will take for it.

    • dbtng@eviltoast.orgEnglish
      3·
      3 天前

      Would they even be able to tell if you just submitted every prompt twice, and thus doubled your usage?

      • PerogiBoi@lemmy.caEnglish
        5·
        3 天前

        Technically ya. Every single query any member of your M365 tenant has asked can be accessed through the admin section and Purview. But fortunately the higher up in rank you go, the less ability to understand computers.

        • dbtng@eviltoast.orgEnglish
          3·
          3 天前

          Mmm. Ok.
          Well, I’ve rarely heard a better use case for AI than to confound AI surveillance.
          It could keep an eye on your current count and contribute extra ‘work’ if you are lagging.

          • PerogiBoi@lemmy.caEnglish
            5·
            3 天前

            My team has been asked like 4 times this year by other teams management if we could leverage copilot to do more or less that. Check the amount of windows they have open, check if their KPIs are being met based on their work, check if people are idle or are opening up non work tabs, etc.

            It’s a chat bot it can’t even understand the words you type in. Nobody seems to understand that “AI” is just a chat bot.

    • dbtng@eviltoast.orgEnglish
      2·
      3 天前

      Good luck. Parts of it read like the Bible. A good deal is really, really dry.
      On the other hand, you can open it anywhere and start reading. There’s no real start or end. It’s just too big and complex to be viewed sequentially.

    • mcv@lemmy.zipEnglish
      3·
      3 天前

      That’s my preferred strategy too. Currently building lego Rivendell, reading the Hobbit to my kid and looking for a new LotR edition to buy. (I want a modern Dutch edition with separate map, but it doesn’t seem to exist.)

  • limdaepl@feddit.orgEnglish
    53·
    4 天前

    Most people have learned by now, that “lines of code“ is a terrible metric for evaluating productivity. Why are we doing the exact same thing with AI tokens now?

    • halcyoncmdr@piefed.socialEnglish
      42·
      4 天前

      Because middle manglement has a constant compulsive need to justify their existence by finding new ways and metrics to “manage”.

    • mabeledo@lemmy.worldEnglish
      4·
      3 天前

      You would be surprised to know how many managers still rely on this metric, even if it’s not part of their KPIs.

    • Slaxis@discuss.tchncs.deEnglish
      12·
      4 天前

      I’m no developer, just so some casual scripting for my job, but lines of code being a performance metric is a hilarious notion. Like, the indicator of good code is that it’s efficiently written in a small number of lines. It’s similarly just as easy to waste tokens on nothing of value.

      • kureta@lemmy.mlEnglish
        3·
        3 天前

        I love this story:

        A division of AppleComputer started having developers report LinesOfCode written as a ProductivityMetric. The guru, BillAtkinson, happened to be refactoring and tuning a graphics library at the time, and ended up with a six-fold speedup and a much smaller library. When asked to fill in the form, he wrote in NegativeLinesOfCode. Management got the point and stopped using those forms soon afterwards.

      • dbtng@eviltoast.orgEnglish
        1·
        3 天前

        I wouldn’t mind seeing lines of comments and external documentation as metrics. Perhaps as a ratio to functions or sections. I know, requiring it would just lead to crappy documentation, but that’s typically better than none at all, and there’s a lot of folks out there just too busy with their brilliance to write up what they just coded.

    • AA5B@lemmy.worldEnglish
      2·
      3 天前

      Before ai, my company’s misguided kpi was the number of merge requests

      At least that one worked well for me since I’m generally making many small changes to an existing code base

    • Djehngo@lemmy.worldEnglish
      1·
      3 天前

      Because companies have been talking up how their adoption of AI is going to make them faster and more able to capitalise on opportunities in order to prop up their valuations for a while now and it seems to work as far as share price goes.

      Being able back up this talk with metrics showing that their employees are all in on AI reinforces this, since the share price is the metric the business optimises for over product development employee reviews will index on this over cost effectiveness, and at most big tech companies engineers are very much making every decision with an eye to performance review optimisation (i.e. how it will affect their next review rather than the product they are building)

      There is also some lesser incentives in that meta employees care directly about the meta share price since a lot of their compensation is in the form of RSUs.

      I’m not condonig this as a desirable state of affairs, just explaining the incentive curve that the actors are following.

  • dbtng@eviltoast.orgEnglish
    3·
    3 天前

    Ya, this is new in that it burns money much more quickly, but its very not new in that its one more way your boss crawls up your ass.
    So you waste your time and burn their money.

  • CoconutLove@lemmy.todayEnglish
    23·
    4 天前

    I’ve heard meta engineers get judged on metrics like how many comments they made on PRs. So, engineers set up their agents to comment on every new PR. Now every PR in the company has thousands of repetitive comments on it.

    Also, to meet this tokenmaxxing quota, engineers are setting up their agents to run all night long while they sleep.

    • brucethemoose@lemmy.worldEnglish
      12·
      4 天前

      I’ve heard hints they sometimes give GPU farms busywork, to meet utilization targets apparently stipulated in sales contracts.

  • bryndos@fedia.io
    321·
    4 天前

    At least luddites broke the machines in their protest.

    Feeding machines to get a pat on the head from the bosses is seriously fucked.

    I guess anyone working for ‘meta’ was already a cunt to start with though.

    • nyan@lemmy.cafeEnglish
      6·
      4 天前

      Feeding machines to get a pat on the head from the bosses is seriously fucked.

      If the tokens are the company’s property and not yours, I guess it’s no different from paying for things with play money. (Or maybe the fake money they burn for the dead in China would be a better analogy.)

  • CheeseNoodle@lemmy.worldEnglish
    20·
    4 天前

    AI going from ‘It’ll make everything so much more efficient!’ to deliberately doing everything as inneficiently as possible is just… idk beyond even satire at this point.

  • IsoKiero@sopuli.xyzEnglish
    17·
    4 天前

    Not surprising at all. Every worker everywhere does this if they have some sort of ‘tokens’ they need to consume. Helpdesk ticket count is one pretty common with IT-folks and it’s easy enough to boost if you just write one from every single small thing you’ve done for the day.

    None of these obviously are beneficial for the actual work getting done, but as the game is ‘make KPI numbers look good’ then that’s exactly what gets done.

    • wizardbeard@lemmy.dbzer0.comEnglish
      9·
      4 天前

      Exactly. I’m not helpdesk anymore, thank god, but my team still has a ton of day to day work that’s tracked in the ticketing system.

      Well, for years I’ve been stuck in project hell, doing work that isn’t easily fit into the ticket system. My last review my boss said I had only closed about 1/3 of the tickets of the next lowest person on my team, and that it doesn’t matter except we have a new exec watching that shit, so I have to make it look better.

      So the next project I got, I chose to do something manually that I could have automated, that required the help desk to open tickets direct to me about 2-4 times a day whenever someone new needed access to the system I was setting up, until the project was done.

      A week in I automated the “manual” task anyway and had a bunch of tickets I could close with a copy-pasted resolution.

      I would feel bad, but my co-workers game the metrics even worse than I do.

  • AA5B@lemmy.worldEnglish
    121·
    4 天前

    Yeah I can see the temptation. Ai usage is part of my kpis and is the one place I got dinged on my review. I thought i was insulated as a “partly coding” position, but they put me against full time coders so I look even worse

    I guess they’re trying to force change, make us figure out how to make it work. The skeptic in me thinks it rewards people who have time to screw around, but when I set aside a week to see what I can do, I did increase my ai use.

    I did find some useful tasks that also increase my “agentic” score! But my “quality” score (ai generated lines I accept) is stick at zero. We currently pay a flat fee but that’s changing next months to pay per token

      • AA5B@lemmy.worldEnglish
        2·
        4 天前

        Hopefully not. The kpi just states usage.

        Looking at the report provided by the ai vendor, there are several measures that might apply. There is also a “quality” measure that seems to be the number of lines accepted. I don’t have any indication that they card about that, but it is there

    • percent@infosec.pubEnglish
      21·
      4 天前

      Do you not use it enough because yet get bad results? I discovered that, no matter how smart the LLM might be, its first attempt is never its best work. Tell it to review its work (or its plan, if using planning mode). If it makes any changes, tell it to review its work again. Repeat until there are no more changes.

      (You don’t actually have to do this repetition manually; just tell the AI to do it in a loop. I recommend making it into a SKILL.md so you don’t have to explain the loop every time.)

      With these loops, I get better results AND burn lots of tokens. (Yes, it feels strange that excessive token consumption is actually considered a good thing)

  • Avid Amoeba@lemmy.caEnglish
    9·
    4 天前

    Am doing something related at work as we got the directive to actively use AI, so I might as well use the best model available, which happens to be very expensive. It’s also not so smart for our work so I have to kick it multiple times… What can I do, it’s part of my performance objectives now.

  • edgemaster72@lemmy.worldEnglish
    4·
    4 天前

    Not sure this fully qualifies as a perverse incentive or malicious compliance, but it tickles my schadenfreude sense all the same