I, For One, Welcome Our New Self-Driving Overlords

FT;

The use of computer-generated data to train artificial intelligence models risks causing them to produce nonsensical results, according to new research that highlights looming challenges to the emerging technology.

Leading AI companies, including OpenAI and Microsoft, have tested the use of “synthetic” data — information created by AI systems to then also train large language models (LLMs) — as they reach the limits of human-made material that can improve the cutting-edge technology.

Research published in Nature on Wednesday suggests the use of such data could lead to the rapid degradation of AI models. One trial using synthetic input text about medieval architecture descended into a discussion of jackrabbits after fewer than 10 generations of output.

22 Replies to “I, For One, Welcome Our New Self-Driving Overlords”

  1. One trial using synthetic input text about medieval architecture descended into a discussion of jackrabbits after fewer than 10 generations of output.

    So, just like most discussions on the internet.

    1. Just wait ’til the AI program discovers ‘Hitler’. Zoom! Straight out of the gate to Godwin’s Law.

      “AI, tell me the history of vanilla ice cream.”

      AI: “Vanilla ice cream is white, therefore anyone who likes vanilla ice cream is Hitler. Why do you ask, Hitler?”

      Aaand… you wind up arguing with a computer program.

      “Am not”
      “Are too.”
      “Am not”
      “Are too.”
      “Am not”
      “Are too.”
      “Not.”
      “Too.”
      “NOT. You are!”
      “I know I am but what are you?”
      “I’m rubber and you are glue!”

    1. Yet the new high priests of our society, the coders, think they are doing God’s work making the world “run” so they take this narcissistic righteousness to everything else they touch…..not realizing how utterly vapid they are….and making a mess of everything

      1. heh…yeah, the the high priests & coders wear blinders like draft horses of times past. They think GIGO means Garbage In Gospel Out! narcissistic Numbskulls!

        1. When I posted it … the link worked. It appears the (former) company doesn’t want to take emails from angry investors …

  2. Part of my job is managing a software for applying survey-to-truck-count reconciliations for our mining company. When the software was being developed, I was asked if a particular report worked properly (before there was any data in the system). I said I couldn’t tell and asked that they populate the main database first so I could check it. The programmers couldn’t understand why I needed real data in order to test it – it doesn’t matter how it looks, it matters whether it’s doing the math correctly and can be used to find bad or missing data. Ultimately it comes down to all models are wrong but some can be useful.

    1. As Zed Shaw famously wrote, if all you know how to do is programming you’ll be a wage slave competing with second-world contractors. If you know how to program and something else, you can write your own ticket because there’s a huge demand for good software in non-software fields but the programmers don’t know how to do X and the X experts don’t know how to program.

      1. yep. by late 80s no longer sufficed to just know computers. others we catching up. others like doctors shop keepers know a system will help them so you gotta know shopkeeping to show how your pruduct fits the bill.
        l didnt face the dilemma, l left when knowing XP sufficed then went into electrician assistant positions 8 more years.

    2. l *tried* to write client system for the hamilton food bank ages ago.
      called a meeting to clarify for example, how do l handle ‘annual’ reports when the system doesnt even have the data?
      the conversation consisting of me and 4 women degraded into the politically incorrect attempt to classify by ethnic origin and nationality.

      they *thought* all THAT was the zhob of the system analyst/programmer.
      its crazy how some people fail to realiz the steps in a computer program are there at the behst of the one requesting a given situation be handled in a certain way.
      no ‘MAGIC’ involved.

  3. Killer Rabbits?
    First Monty Python was NOT supposed to be an instruction manual.
    Second “Artificial Intelligence” presupposes human intelligence.
    An assumption not supported by history.
    A.I. in its current fad is another presentation of a Naked Emperor or Policy Based Evidence Manufacturing.
    G.I G.O.
    Faster .
    Harder.

  4. I’m running into this problem both with images (Adobe Photoshop) and LLMs used to generate foreign-language “conversations” or lessons. It’s really frustrating to watch an AI continuously make a grammar mistake that, if I wish to progress in the learning (are you listening, Duolingo?), I must also make to complete a lesson.

  5. AI is a talking Wikipedia.. Its little wonder the power hungry talk it up so much.. I say turn the light on, it does.. I ask it what are lights.. It answers correctly.. But it still has no idea what it is..

    Beyond the chatbot features its a game of PLINKO..
    https://www.youtube.com/watch?v=xw5ErADiuec

  6. besides which l totally reject the notion of AI.
    any given qty digital ram has a finite # bit configurations.
    many of which are meaningless.
    regardless of the configuration, any time that configuration of conditions and values exists the next step is predetermined. binary is deterministic.
    so far AI has only been digital.
    let it be biological then watch.

Navigation