Login
You're viewing the merveilles.town public feed.
  • Nick Farrnickfarr@chaos.social
    Sep 14, 2026, 5:13 PM

    OH: "Recursive self-improvement in AI is a lot like what the Habsburgs thought they were doing by marrying their cousins."

    💬 12🔄 134⭐ 2

Replies

  • Frank Nordfenkt@social.tchncs.de
    Sep 14, 2026, 5:39 PM

    @nickfarr The Habsburgs left us beautiful castles and other things.
    Do you think your grand grandkids will once marvel at the "Zilchtronic gas turbine", used to power the data center of emperor Elon I?

    💬 4🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 1⭐ 1
  • Nick Farrnickfarr@chaos.social
    Sep 14, 2026, 8:28 PM

    @fenkt Honestly, if I have great grand children and they have electrical power and indoor plumbing, I'll be impressed.

    If my grandchildren get to see most of the Vienna I got to see, I'll consider myself incredibly blessed.

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 2🔄 0⭐ 0
  • climbertobbyclimbertobby@chaos.social
    Sep 14, 2026, 5:58 PM

    @tomkalei @nickfarr

    Depends on how the synthetic data is generated. If it's generated from some sort of logical, physical, ... model (e.g. by training vision model on 3d renderings with added noise) it works. If the artificial data generated by trying to predict the same text as the original trainigdata it cannot provide additional information.

    💬 0🔄 1⭐ 0
  • Krzysztof Sakrejda 🏳️‍⚧️ 🏴‍☠wronglang@bayes.club
    Sep 15, 2026, 2:10 AM

    @tomkalei @nickfarr equating recursive self improvement and synthetic training data (which is usually arranged very much *not* by the algorithm being trained for a specific purpose... that's as much of a stretch as calling successful machine learning implementations "AI". it's attempting to steal credibility.

    💬 0🔄 2⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • muddle 🥣muddle@infosec.exchange
    Sep 15, 2026, 10:23 PM

    @nickfarr I was trying to remember that phrase. I came across it earlier and it struck me that what it actually is, technically, is chaining lossy compressors together to get some sort of magic infinite compression (aka, the singularity).

    It's basically the WEB compression con from many years ago on Usenet mixed with stupid Kurzweil nonsense. They're all not only singing from the same hymn sheet, but cribbing from the worst pre-internet memes and scams.

    #WEB #WeExudeBullshit

    💬 0🔄 1⭐ 0
  • 💬 0🔄 0⭐ 0
  • OddOpinions5failedLyndonLaRouchite@mas.to
    Sep 16, 2026, 1:14 PM

    @nickfarr

    "honest question"

    if in a year or three, it turns out that recursion actually is some benefit, will you apologize to your followers, and as penance take a day of silence on social media ?

    💬 1🔄 0⭐ 0
  • Nick Farrnickfarr@chaos.social
    Sep 16, 2026, 2:18 PM

    @failedLyndonLaRouchite I think you missed the point.

    The point is RSI is great until model collapse sets in. Marrying cousins worked exceptionally well to preserve and consolidate power within the Habsburg house. Until a branch went extinct.

    Just as theory-free inference is absolutely amazing at producing results...to a point. When you reach that point, it fails catastrophically.

    Model collapse and Habsburg AI are good terms to further research this phenomenon.

    nature.com/articles/s41586-024

    💬 0🔄 0⭐ 0
  • Tichodrome Colverttichodrome_colvert@piaille.fr
    Sep 16, 2026, 7:23 PM

    @nickfarr I wonder if there will ne a cycle of AI: models stop developing because they will inevitably ingest AI output, so AI companies will stop providing services at an affordable price, making almost all new content human-generated once again. Then AI companies will have to worry far less about training new AI models on AI output and new models become usable once again. Until AI stops once again being provided, then trained once again on human-generated content, and so on.

    I'm just doing half-joking speculation but that would be a funny prediction if this is what will actually happen.

    💬 1🔄 0⭐ 0
  • Nick Farrnickfarr@chaos.social
    Sep 16, 2026, 7:46 PM

    @tichodrome_colvert I think you're more or less accurately describing the cycle, even if the wave is a bit shallower than you imply.

    Companies are already hiring back expertise they thought AI would replace. And of course, those employees are training models. And the improvement in the models we have is a fraction of what it was.

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0