Don't Call It a Decision Model
40 points by emschwartz 16 hours ago | 8 comments
  • edot 14 hours ago |
    Drives me nuts that this nomenclature caught on. This (judgement, not decision) is a bit better. I mean, sure they (Jev-style models) are useful for zero-shot classification, but they're only "judgement models" insofar as any (LLM, ML, mathematical optimization, heuristic, etc.) model that sets a downstream float / integer / boolean to commit resources (read: makes a decision) is a "judgement model".

    https://warrenpowell.org/decisionsdecisions/

    If you actually want to make good, even optimal (provably) decisions, try reading some about mathematical optimization, or operations research more broadly. The above link is a good place to start.

    • bobbiechen 10 hours ago |
      "System One" is a mess of a name from the start, as is any classification that chooses arbitrary labels for non-arbitrary things (Type I vs II errors, etc.)
      • seanhunter 4 hours ago |
        Totally agree. Whatever you think of Kahneman etc (Personally I think the original "Farmers vs Librarians" thought experiment in Bayesian reasoning was amazing for example) it's weird that to talk about a model you need to know about Thinking, Fast and Slow and what it said about system one/system two. Yes every name for a model is essentially jargon, but some jargon is better than others in terms of at least being somewhat understandable. When someone calls Jev a "decision model" it makes complete sense to me. That evokes to me the idea of a zero-shot classifier (ie the purpose of this type of model).

        Not only does "system one" fail to do that (because it is an arbitrary classification as you have said) but it also ascribes to the model a type of thinking that humans do when we really have no reason to believe that is how the classification process works with this type of model. So if you'd read Thinking, Fast and Slow you might reasonably attach attributes of human system-one type thinking to this model that are not justified. So it's simultaneously opaque (if you don't know the reference) and misleading (if you do).

      • samusiam an hour ago |
        The numbering isn't arbitrary though. Theoretically it's the order in which those cognitive systems evolved. As for "System," well, I suppose you could call them Cognitive System One and Two. But it's usually pretty obvious from the context.
    • seanhunter 3 hours ago |
      The whole point of using a weak model for solving some problem is to trade _off_ some optimality for other characteristics. If you have something that is provably optimal it is probably special-purpose, or requires upfront training or doesn't work in the full generality of cases etc. The selling point of this class of model is that inference is fast and cheap and they produce results that are good enough for a wide class of classification problems where training a real classifier is inconvenient. For example, I have one of these type of models running locally and doing model selection when I use my coding harness. It takes prompts I provide and subtasks spawned by agents and matches those into various categories ("trivial", "research", "implementation", "hard") which are defined by prompts. As you can see this is a very fuzzy task and a "provably optimal" solution is likely impossible to achieve but also would require training on a data set I don't have (a representative sample of possible prompts and subtasks in the various categories).

      The decisions it's making are good enough to save me a ton of money on my development so are perfectly adequate for this task.

  • willio58 14 hours ago |
    Really cool visual way to do demonstrate the point. Jev is so interesting to me, I just struggle to find good use cases in my work so far.
  • j45 10 hours ago |
    I agree with this.

    Decisions/Judgements in LLMs are not logic gates like a deterministic piece of software, they are simulations cosplaying it more often than not.

    Depending on how it's worded it's pretty remarkable how much it can improve, but it ultimately is not deterministic.

    • perching_aix 4 hours ago |
      They're fully deterministic, just (varyingly) unstable. Like a weather model, or any other chaotic systems' simulation.