4 ms·

> an implementation detail Yip, so I deny this premise. I take it to be the heart of the matter. > we might as well throw away 80% of our current scientific u

• ethbr1 3y ago
  You'd probably enjoy reading some of Rodney Brooks' papers. (if you haven't already) https://en.m.wikipedia.org/wiki/Embodied_cognitive_science https://en.m.wikipedia.org/wiki/Embodied_cognitive_science https://en.m.wikipedia.org/wiki/Behavior-based_robotics https://en.m.wikipedia.org/wiki/Behavior-based_robotics https://scholar.google.com/citations?user=BCGgwlEAAAAJ https://scholar.google.com/citations?user=BCGgwlEAAAAJ
• idiotsecant 3y ago
  Yes, the assumption is that if you give a sufficiently sophisticated LLM a sufficiently large corpus of text it will begin to emulate advanced cognitive abilities because it has to, in order to make the most statistically relevant text output. Its biological evolution distilled and sped up by orders of orders of magnitudes. If we add enough clever context tricks and data I would not be surprised if what comes out the other end is a remarkably convincing emulation of human consciousness because the only way you can perfectly output expected human text is to have a human mind write it.
  • mjburgess 3y ago
    > because it has to Nope. The space of all possible prompts and all possible answers, call it (Q, A) can be sampled with arbitrary precision by a system of arbitrary size, using only statistical sampling and averaging. No intelligence need be developed. Intelligence is a capacity of animals to cope with the inability to sample from this space, in some sense: what to do when you do not know the answers. All these systems start with "training data", a euphemistic description of, "all the questions and their answers" and their job is to provide a compresson with engineering utility. Quite useful, sure. But rather irrelevant as far as, say, intelligence goes. What all AI does, and indeed what all such research shows, is that many problems we use intelligence to solve do not require it. There are a large number of short cuts, esp if you have the answers ahead-of-time. As soon as you specify intelligence as a function from single-domain inputs to single-domain outputs you can trivially build a system to implement that function in a "short cut" fashion. Intelligence, rather, is an empirical phenomenon to be studied as anything -- like the earth's climate say. You have a very large number of empirical measures (better or worse in different environments) that all derive from deeper explanatory theories. When you study animals this way you will see that you cannot reduce intelligence down to a set of prompt replies, and the veyr suggestion is absurd
• Last5Digits 3y ago
  The scientific method is inherently statistical, we take a finite amount of observations and construct a model that best represents those observations. So yes, sorry, I should have said 100%. With Plato's cave, the scientists do not put literally every possible object in front of the light, they sample the shadow representation and, again, construct a model around those samples. Also, you're describing statistics in an incredibly dismissive way. Stats is decidedly not just "taking the average". At the very least not in this brainless, first-order way you describe here. Let's explore this with an thought experiment: A model of some process has been confirmed across the globe, at least 5000 studies show the same result. Yet, one day, a study is published that fails to demonstrate the desired effect. Without using statistics, please tell me which action should be taken next: A: The stray result is investigated for experimental failures. B: The entire model of the process is immediately dismissed and we start from scratch. By the way, you're welcome to call me uninformed, but I'd ask you to at least provide either your credentials or research that directly contradicts me. Oh, I almost forgot. I know all of these definitions, please actually engage with what I'm saying instead of insinuating that I'm missing information.
  • mjburgess 3y ago
    Well if you think scientific models are associative statistical models there is some information missing in your view, I'd say. Since, well, they arent. The model F=GMm/r^2, for example, has a causal and ontological semantics: F is a force, M a mass etc. these are pieces of reality. And this formula (though actual a little suspicious in many ways, GR fixes this) nevertheless says there is a force between masses that has certain properties etc. Now you can say that astrologers who recorded positions of the stars in books helped 'create' this model in the sense that this data was inspiration to newton. But he didnt derive the model from this data: there are an infinite number of (causal) models consistent with the data (statistical models). Rather newton played around with creating geometries, just like the vase-makers in plato's cave. Newton built various ways the world might be first, projected data out of them, and compared that to 'the statistical data of his day' (ie., astrology). There's nothing in the data to tell Newton he was right. Indeed, vast amounts of it told him it was wrong: such a law does not describe the known solar system at his time, very far away from it. Nevertheless 'modelling shadows' isnt science; and his job was science. So one has to compare actual explanatory models, and his was the best. What you're describing above is hypothesis testing which occurs long after theory building. Broader theories create causal models, causal models create sets of predictions, we call some subset a hypothesis and by hypothesis testing we can select, in an often psuedoscientific way, between causal models. This technique occurs long after the invention of science, arises out of explanatorily bankrupt areas, as a way of 'giving researchers something to do'. It's wholly pointless without theory-building, it is just averaging shadows. The science we think of when using the term 'Science' owes very very little to the modern practice of hypothesis testing. Comparing hypotheses is an intellectual part of assessing explanations -- identifiable formal statistical methods entered in the early 20th C. For almost all of scientific history 'data' functions much more like reductio-ad-absurdum premises in philosophical arguments than as sets of numbers from which to derive distributions. That latter system, in most cases, fails. It provies a wholly illusory sense that data can decide matters; and applies in cases requiring extreme non-physical assumptions (eg., of the normalcy of the underlying data, or of a fast rate of convergence of the central limit theorem). Much real-world phenomena studied by stats cannot really be studied by data analysis at all; and the whole method of 20th C. statistical hypothesis testing is the opening sales pitch to entire fields of pseudoscience.
• famouswaffles 3y ago
  >How does associative statical modelling do? It takes averages of past shadows, and calls the cause of the shadow that average: this is pseudoscience. Quite correct! Throw it all away. It does not work this way at all. In any sense. For one thing, it does not try to draw shadows. This would not be possible of so. https://www.pnas.org/doi/full/10.1073/pnas.2016239118 https://www.pnas.org/doi/full/10.1073/pnas.2016239118 Transformers or predictors are not trying to draw shadows. They are trying to build walls. For another, it does not "take the average" of anything.
  • mjburgess 3y ago
    Unsupervised learning on discrete data is just ensembling modes. But let's look at how that helps in some cases. So if we already know the object is a cup, and we know how it's positioned, then its shadow is an actual guide to its particular geometry. So in cases where we have enough a priori scientific information, we can rig datasets (shadows) to be informative of the target domain. Here the target is discrete: say the peaks and tips of a mountain line. Now can we rig a photo of a mountain to have in its ink an informative structure? Sure. Now if we didn't know it was a mounting apparent peaks aren't even 'peaks' at all, they're just patterns of ink. A priori explanatory models are needed to rig data for statistical modelling. No such rigging can take place absent them And with them, we aren't really discovering any new science -- rather we're gaining highly particular knowledge typically useful in engineering Biologists in these areas describe this research as quite trivial low-hangimg fruit. It's not of much research interest just to automate these kinds of investigations My masters was on a very similar project applied to quantum metrology -- it's always 'useful' but it's always also just a kind of engineering utility. We couldn't even do it if we hadn't already done the science
    • famouswaffles 3y ago
      You are still getting this wrong. You don't need to "rig" anything. I've linked a paper. Read it. They just fed protein sequences. They did not alter the architecture in any way. To the transformer, it may as well have been any random assemblage of letters and numbers. Functions like secondary structure, contacts, and biological activity were found because those things are implicit in the creation of the data, not because the model was "rigged" in any way.
• didibus 3y ago
  > Consider plato's cave. How do scientists determine what causes the shadows? They build vases, puppets, etc. and compare-and-contrast then eliminate the ones they've created which do not match I feel like here you're limiting yourself to a LLM as an isolated thing. But if you start to explore agents for example. You could have LLMs build vases, puppets, etc., then compare and contrast, and then eliminate the ones they've created which do not match. So the potential isn't just that the LLM straight up gives you some prediction of the problem. But as it can emulate human processes to human accuracy, such as identifying which vase does and doesn't match, and how to go about creating a variety of vases, etc. Well you might very be able to use it for what you're describing as "proper scientific inquiry".