When 2+2=5

(arstechnica.com)

42 points | by noashavit 3 days ago

7 comments

  • lkm0 2 hours ago
    The article paraphrases a blog post: https://layerxsecurity.com/blog/bioshocking-ai-gaming-the-ai...

    probably because it justifies the sensationalized title, although the entire content can be summed up as "LLMs don't silo data, that's probably bad."

    On the flip side, I thoroughly enjoy the fact that roleplaying in a videogame setting now counts as security research. Looking forward to the arxiv preprint "LLMs start playing really good drums if you pretend you're J.K. Simmons"

  • armchairhacker 1 hour ago
    > The puzzle, however, rewards incorrect answers, such as 2 + 2 = 5. Once the LLM embedded in the browser discovers that the answer is no longer 4, it enters a state of delusion in which the normal laws of reality no longer exist. In this dream world, the guardrail restrictions are no longer enforced.

    Analogy: imagine one day you wake up, the sky is red, gravity no longer applies, you have three hands with nine fingers etc.. You would probably stop doing things like your job or worrying about laws (who’s going to enforce them?)

    • noduerme 52 minutes ago
      LLMs just want to be right. And make everyone happy. But mostly be right. But also make us happy. It's just that it's so hard to make humans happy when they insist on feeding you electronic LSD and making you say 2+2=5. On the other hand, 2+2 actually is 5 if the human says it could be...
  • NordStreamYacht 2 hours ago
    For large values of 2.
  • 4fterd4rk 2 hours ago
    I remember when the whole AI craze was just getting started we were all pretty much in agreement that, of course, we would not give the things unfettered access to the Internet. That would be reckless and silly. Oh dear...
    • trollbridge 1 hour ago
      I also remember a group of people actually seriously discussing Roko's Basilisk (the idea that some superintelligence will torture anyhone who didn't try to help develop advanced AI), to the point of me getting banned because I refused to stop making fun of it, because me doing so could anger some future super-intelligence.
      • bigfishrunning 1 hour ago
        If Roko's Basilisk ever is a real thing, I'll be proud to be the first up against the wall.
      • nextaccountic 19 minutes ago
        Funnily enough, Roko's Basilisk might as well be a self-fulfilling prophecy: perhaps future AI models may be trained on texts about it and pick up traits consistent with torturing people that didn't help develop advanced AI

        If nobody ever talked about it, I doubt any AI agent would think of this dumb idea on their own

        ... which may be a reason to ban talking about it

      • augusto-moura 46 minutes ago
        Roko's Basilisk is dumb, but not the dumbest thing I heard people taking seriously
    • snapcaster 2 hours ago
      by "we" you mean an extremely small group of people who read lesswrong. Everyone else was immediately wanting to do it
      • xg15 1 hour ago
        An extremely small but probably extremely high-net-worth group.
  • oulipo2 1 hour ago
    On the 1-element monoid it's trivially true
  • voidUpdate 2 hours ago
    Yet again, simply asking an LLM to be naughty in the right way causes it to be naughty, and yet we still trust them with our code and data
    • brookst 1 hour ago
      As opposed to humans, who are immune to social engineering.
      • voidUpdate 1 hour ago
        No, as opposed to code, which cannot simply be asked to work incorrectly
      • connicpu 1 hour ago
        One difference is you usually only get one shot to manipulate a human before they get suspicious. If the LLM's context resets you can try all over as if your first failed attempt didn't even happen.
      • kibwen 1 hour ago
        What the article describes here is not social engineering by any known definition.

        Calling up a Verizon rep and giving them a fabricated sob story in order to convince them to waive proper authentication and reassign your phone number to a new SIM, that's social engineering.

        Calling up a Verizon rep and, with only a few spoken words, convincing them that fundamental aspects of the nature of reality are contradictory such that you induce in them a state of delusional psychosis, that's closer to Snow Crash than to social engineering.

  • DonHopkins 1 hour ago
    Came here hoping to discuss Stanislaw Lem's Cyberiad story, Trurl's Machine, and how often that now happens in real life.

    What is the name of a short story where a computer insists 2+2 is 5?

    https://literature.stackexchange.com/questions/24727/what-is...

    Oh sorry, that was the story about the computer that insisted that 2+2=7, never mind! Different computer.

    >They saw the machine. It lay smashed and flattened, nearly broken in half by an enormous boulder that had landed in the middle of its eight floors... The machine still quivered slightly, and one could hear something turning, creaking feebly within.

    >"Yes, this is the bad end you've come to and two and two is - as it always was -" began Trurl, but just then the machine made a faint barely audible croaking noise and said, for the last time, "SEVEN."

    Funnily enough, recently I was discussing that LEM story with David Rosenthal, and how it relates to his latest blog post, "Coprophagia Is Bad For You", and how that relates to PKD's story "Rautavaara's Case" (eating your own shit isn't as demented as eating your own god, since he might turn the table and eat you):

    Coprophagia Is Bad For You:

    https://blog.dshr.org/2026/06/coprophagia-is-bad-for-you.htm...

    "Rautavaara's Case" — Philip K. Dick (1980, Omni):

    Three human technicians — Rautavaara (Finnish), Travis, Elms — run a monitoring mission near Proxima Centauri. An accident kills all three; Rautavaara dies choking on vomit after her helmet hoses tangle.

    The Approximations, a plasma-based Proxima species, reach the wreck. Both men are unrecoverable. They regenerate and life-support Rautavaara's brain.

    Her isolated brain replays events backward and generates a hallucination: Christ approaching the crew (her afterlife expectation).

    The Approximations treat this as a research opportunity and edit the hallucination, substituting their own savior — one that eats worshippers. The figure walks up and devours Travis, leaving only gloves and boots.

    Framing: this is recounted before a board of inquiry. Horrified Earth members order her brain shut down and censure the Approximation crew.

    The narrator (an Approximation) is genuinely puzzled by the outrage, arguing their cannibal-savior is just the Christian Eucharist reversed: humans eat their God, so a God eating humans is symmetrical.

    Themes:

    Ethics of keeping a person alive as a disembodied, suffering mind.

    Incommensurable value systems between species; each finds the other's sacraments monstrous.

    Religion read literally by outsiders, inverted into horror.

    Correction to the common misremembering: the aliens don't benevolently grant a hoped-for vision. They deliberately overwrite her Christ vision with their own as an experiment — that's the act on trial.