Source: nitter, twitter

Transcribed:

Max Tegmark (@tegmark):
No, LLM’s aren’t mere stochastic parrots: Llama-2 contains a detailed model of the world, quite literally! We even discover a “longitude neuron”

Wes Gurnee (@wesg52):
Do language models have an internal world model? A sense of time? At multiple spatiotemporal scales?
In a new paper with @tegmark we provide evidence that they do by finding a literal map of the world inside the activations of Llama-2! [image with colorful dots on a map]


With this dastardly deliberate simplification of what it means to have a world model, we’ve been struck a mortal blow in our skepticism towards LLMs; we have no choice but to convert surely!

(*) Asterisk:
Not an actual literal map, what they really mean to say is that they’ve trained “linear probes” (it’s own mini-model) on the activation layers, for a bunch of inputs, and minimizing loss for latitude and longitude (and/or time, blah blah).

And yes from the activations you can get a fuzzy distribution of lat,long on a map, and yes they’ve been able to isolated individual “neurons” that seem to correlate in activation with latitude and longitude. (frankly not being able to find one would have been surprising to me, this doesn’t mean LLM’s aren’t just big statistical machines, in this case being trained with data containing literal lat,long tuples for cities in particular)

It’s a neat visualization and result but it is sort of comically missing the point


Bonus sneers from @emilymbender:

  • You know what’s most striking about this graphic? It’s not that mentions of people/cities/etc from different continents cluster together in terms of word co-occurrences. It’s just how sparse the data from the Global South are. – Also, no, that’s not what “world model” means if you’re talking about the relevance of world models to language understanding. (source)
  • “We can overlay it on a map” != “world model” (source)
  • self@awful.systems
    link
    fedilink
    English
    arrow-up
    21
    ·
    1 year ago

    between this and that one fucking chart that tried to say humans emit more CO2 producing text and art than generative AI (using the same underhanded tactics as cryptobros trying to make it look like banks are worse for the environment than blockchains), I’m really starting to feel like the AI industry is in the “deliberately fill your scam email with as many typos as possible to weed out anyone too intelligent” stage of its growth

  • swlabr@awful.systems
    link
    fedilink
    English
    arrow-up
    16
    ·
    1 year ago

    Anytime they say it’s not a stochastic parrot, what they really mean is that it’s three stochastic parrots in a trenchcoat.

  • zurohki@aussie.zone
    link
    fedilink
    English
    arrow-up
    15
    ·
    1 year ago

    I think I'm going to start dropping the phrase "As a large language model" into my emails.

    • self@awful.systems
      link
      fedilink
      English
      arrow-up
      12
      ·
      1 year ago

      as a large language model, I am incapable of feeling surprise that Tegmark is associated with neo-nazis

      (also I really need to de-jank the stylesheet for the archive and get the rest of the data in it soon)

    • zogwarg@awful.systemsOP
      link
      fedilink
      English
      arrow-up
      11
      ·
      edit-2
      1 year ago

      Not even that! It looks like a blurry jpeg of those sources if you squint a little!

      Also I’ve sort of realized that the visualization is misleading in three ways:

      1. They provide an animation from shallow to deep layers to show the dots coming together, making the final result more impressive than it is (look at how many dots are in the ocean)
      2. You see blobby clouds over sub-continents, with nothing to gauge error within the cloud blobs.
      3. Sorta-relevant but obviously the borders as helpfully drawn for the viewer to conform to “Our” world knowledge aren’t even there at all, it’s still holding up a mirror (dare I say a parrot?) to our cognition.
      • froztbyte@awful.systems
        link
        fedilink
        English
        arrow-up
        7
        ·
        1 year ago

        haha I know (re precision) but I made that as a shitpost not an academic paper. besides, it's about as accurate as the promptfans are

        that animation… is, yeah. I'm reminded of watching someone eyeball stats on their model as they were tweaking parameters, trying to tamp down overfitting.

        it's also just such shitty science. "can we find some way to represent this data to conform to $x hypothesis?", albeit that of course isn't surprising from the P-Hacking As A Service crowd

  • froztbyte@awful.systems
    link
    fedilink
    English
    arrow-up
    12
    ·
    1 year ago

    some of these replies (those are diff links) are staggeringly awful

    and this one is a piece of art:

    chatGPT just learns, it doesn't reason, it doesn't use imagination, it doesn't plan. It learns at a scale that's so far beyond what any human can, that it can use "pure" learning to do complex behaviors.

    • zogwarg@awful.systemsOP
      link
      fedilink
      English
      arrow-up
      11
      ·
      1 year ago

      ^^ Quietly progressing from humans are not the only ones able to do true learning, to machines are the only ones capable of true learning.

      Poetic.

      PS: Eek at the *cough* extrapolation rules lawyering 😬.

    • swlabr@awful.systems
      link
      fedilink
      English
      arrow-up
      10
      ·
      1 year ago

      Oof, they got so close on that last one, yet so far away. Truly a masterpiece in misunderstanding

  • carlitoscohones@awful.systems
    link
    fedilink
    English
    arrow-up
    11
    ·
    1 year ago

    My first thought in watching the animation was - AI parrot shoots shotgun at side of barn, draws target around result.

    Are the dots in the ocean from Amazon's underwater warehouse structures?

  • 200fifty@awful.systems
    link
    fedilink
    English
    arrow-up
    10
    ·
    1 year ago

    I had the same thought as Emily Bender's first one there, lol. The map is interesting to me, but mostly as a demonstration of how anglosphere-centric these models are!

  • Kichae@kbin.social
    link
    fedilink
    arrow-up
    9
    ·
    edit-2
    1 year ago

    So, what's going on here, in plainer language, anyway? Are they just including location information in training data and then, totally surprisingly finding it again in the output data? That's kind of the sense I get from the post here, but I'm not sure if I'm misunderstanding.

    Or did they just cluster the data and squint until someone said one of the graphs "kinda looks like it lines up with a Mercator projection"?

    • froztbyte@awful.systems
      link
      fedilink
      English
      arrow-up
      16
      ·
      edit-2
      1 year ago

      I… so. damn you, I looked.

      this says

      For spatial representations, we run Llama-2 models on the names of tens of thousands cities, structures, and natural landmarks around the world, the USA, and NYC. We then train linear probes on the last token activations to predict the real latitude and longitudes of each place

      their code does… a lot of things with that input data. including filling some in and conveniently removing "small" towns and some states and eliminating duplicates[1] and other shit

      a very quick glance at some of the input data:

      % xsv sample 5 uscities.csv| xsv table
      city            city_ascii      state_id  state_name  county_fips  county_name  lat      lng        population  density  source  military  incorporated  timezone         ranking  zips                     id
      Northglenn      Northglenn      CO        Colorado    08001        Adams        39.9108  -104.9783  37899       2056.3   shape   FALSE     TRUE          America/Denver   2        80260 80233 80234 80603  1840020192
      East Gull Lake  East Gull Lake  MN        Minnesota   27021        Cass         46.3948  -94.3548   961         44.4     shape   FALSE     TRUE          America/Chicago  3        56401                    1840007720
      Idaho Springs   Idaho Springs   CO        Colorado    08019        Clear Creek  39.7444  -105.5006  2044        318.7    shape   FALSE     TRUE          America/Denver   3        80452                    1840018790
      Santa Rosa      Santa Rosa      TX        Texas       48061        Cameron      26.2561  -97.8252   2873        1373.2   shape   FALSE     TRUE          America/Chicago  3        78593                    1840023167
      Mystic          Mystic          IA        Iowa        19007        Appanoose    40.7792  -92.9446   337         41.6     shape   FALSE     TRUE          America/Chicago  3        52574                    1840008316
      

      cool. so. we have high-precision data with actual coordinates and well-defined information. as the input. to the mash-things-together-into-a-proximates-slurry machine.

      and then on prompting the slurry with questions about "hey where is Wyoming", it can provide a rough answer.

      amazing.

      [1] - whoops forgot the footnote. how about that Washingon in every state, huh? sure is a good thing the US doesn't have lots of reused names!

      • Kichae@kbin.social
        link
        fedilink
        arrow-up
        10
        ·
        1 year ago

        Wow. I was kinda tongue-in-cheeking it there, because I genuinely thought I was misinterpreting/over-simplifying the OP, but they really are trying to sell "it didn't discard this data we explicitly fed it" as some kind of big deal.

        I was expecting this to be more like them discovering that regional dialects exist or soemthing dumb-but-not-that-dumb.

        • froztbyte@awful.systems
          link
          fedilink
          English
          arrow-up
          7
          ·
          1 year ago

          promptfans, making grandiose badfaith claims that turn out so not-even-wrong it entirely moves the goalposts on the argument? nevarrrr

        • blakestacey@awful.systems
          link
          fedilink
          English
          arrow-up
          10
          ·
          1 year ago

          15-ish years ago, I was doing a lot of principal component analysis and multi-dimensional scaling. A standard exercise in that area is to take distances between cities, like the lengths of airline flight paths, and reconstruct a map. If only I'd thought to claim that to be a world model!

        • blakestacey@awful.systems
          link
          fedilink
          English
          arrow-up
          7
          ·
          1 year ago

          Whereas the electro-mechanical device that Turing built could perform just one code-cracking function well, today’s frontier AI models are approaching the “universal” computers he could only imagine, capable of vastly more functions.

          Fucking Christ, that hurt to read.