 Yesterday I encountered a "wrong-on-the-internet" rando professing his excitement for "using machine learning" in #3dprinting to throttle speeds in the right places to avoid quality loss. While completely not worth engaging with, I feel like this is a useful example to understand why this idiocy is so infuriating... 46 comments
 @dalias on the other hand, they're not even wrong. with all the money going into machine learning pointlessly, it's more likely to get done by machine learning than the existing methods. The proponents of this kind of shit want to throw away the whole concept of having and using scientific knowledge obtained by experiment, with documentation of how it was obtained, evidence supporting the resulting models, falsifiability, etc., and replace it with a worse version of the way humans tens of thousands of years ago came to believe things about the world: simplistic pattern recognition. Simplistic pattern recognition, and falling for false patterns, is the worst of human stupidity. This shit should be called artificial stupidity not artificial intelligence.  @dalias There is a reasonable use for ML here to prove a solution (or an approximation of one) exists *at all*. But once you've done that, go figure out what the ML actually learned, turn it into a nice 20-line program you can reason about and tweak, and stop applying ML voodoo. I mean this is what we do all the time anyway, it's just usually the "ML" bit is "a human brain".  @dalias I think you are attributing way too much agency to them. It's not high aspirations for AI's capabilities. It's fundamental disinterest in the subject matter that's the problem. My bet is that they thought - gee, this looks like a hard problem. I bet that no one has thought about it before, so instead of me investing the effort to understand it, I'll just have AI solve it for me. AI is also a hard problem, but people have the illusion that they understand it because they know how to `import keras` @chrisg And, not incidentally, the same anti-expertise sentiment is characteristic of fascism. @dalias Very true. Now that you wrote this, I realize it is possible to draw a line between obsession with AI and it's associated deification of existing knowledge to some of Ur-Fascism properties, like the cult of tradition and cult of action.  @dalias allergic to domain expertise!!!! this is why copilot uses a fucking ENGLISH tokenizer for PROGRAM CODE which we have fucking PARSERS for!!!!! @dalias i'm NOT fucking writing it for them they can stew in their own fucking mediocrity  @hipsterelectron @dalias On a related note, are there any language-specific models out there which take a parsed intermediate representation as input and confine themselves to valid output?  @hipsterelectron @mattb @dalias this here. Hand waving some of the infra improvements and some reasoning capabilities: LLMs are just Markov chains with all their pre-computed in lookup tables and loaded into memory. This is why they will never beat expert systems at reasoning - because that's not what next token prediction is. Side note: I love the idea of building a AST-based model to query than a token based one.  @hipsterelectron @mattb @dalias This has been explored in the past (known variously as e.g. "Evolutionary Programming") where you take a bunch of randomly-generated programs in their parsed format to copy/transform/combine them & assess the ones which best solve the problem to semi-randomly go into the next round. But Neural Nets seem to be the only Machine Learning tactic which gets any attention... @hipsterelectron @mattb @dalias I the problem is getting a mapping from a textual program description to the intermediate representation. Tge LLMs do their coding tricks by associating code to accompanying discussion and comments. What they want to do is let you use text to describe your problem then have the system shot out plausible code. I think you’re idea implicitly requires the model to actually have some understanding of what it’s doing. @lain_7 @mattb @dalias to paraphrase @emilymbender, it's just acting as a much worse search engine at that point. erasing copyright/attribution is a positive for the monied interests pushing these machines over ones incorporating any level of domain expertise  @mattb @hipsterelectron @dalias Yes, there is a long tradition of parsing into semantic representations, and even work on generating from them. If you look at it that way, you immediately see that generation of grammatical strings alone isn't really enough. You need to have a way to connect the semantic representations to some model of the world, and determine what valid things you want to say. @mattb @hipsterelectron @dalias One of the issues with LLMs is that they provide apparent fluency on unlimited topics, making it seem like you don't need to do the extremely difficult world modeling work on those topics... @emilymbender @mattb @hipsterelectron @dalias LLMs are just Ricardian models of the world (it's clear the people [outside acidemia] who make them think they will just infinitely grow in knowledge perfectly) @emilymbender @mattb @hipsterelectron @dalias I make this as my own observation, not as an explanation. Obviously you know more in the academic field, but I also observe on the practitioner space where we are looking at putting them in front of people and I'm not so sure. Not every message has the intent you seem to be alluding to @tanepiper Please feel free to make your observations outside of my mentions, then. As it stands, you have addressed this comment to me, in response to my post, without any connective text indicating how it is supposed to relate. It reads as if you felt that I needed to be enlightened. @tanepiper Also, in case you missed it, mansplaining is never about intent. @emilymbender no, apologies if it came off that way - reading it with a tinge of sarcasm and deadpan humour helps (but of course that does not come across in text). Many sales teams of products promise infinite productivity gains and it's exhausting. I've clarified it's this hopefully. FWIW I was already in this particular thread, just a different branch 🤷🏼♀️  SCOTUS basically did the same thing in reversing Chevron. Law is a complex language model reinforcing bigotry and protecting the status quo, not a method for separating fact and falsehood (especially in the face of new information).  @dalias they want badly to believe that minds can be reduced to simple stochastic models. This wasn't an entirely unreasonable hypothesis 20 years ago, but at this point it doesn't look like it's correct. @ravenonthill The concept is still vaguely plausible, but their idea for how to achieve it is utter bullshit. If you compare how human minds are "trained", there are multiple feedback layers in the form of consequences, and most importantly, we select very carefully what training inputs are used rather than throwing giant mostly wrong and mostly evil corpuses at children, and most of the training is experiential not ingesting word soup. @dalias I agree with your general argument, but my uneducated guess for routing and timing for 3D printing is that it would be full of NP complete optimization problems where heuristic solutions are appropriate? If so, maybe throwing an AI at it is an expensive but not entirely misguided approach? @BuschnicK There's a tiny chance that might be right if the topic were travel order, but the topic was literally something where the answer is as dumb as "you can only exceed your heater's sustained melt rate capability in short bursts, and doing that on the outside surface may look bad".  @dalias at least in part, this is testament to our collective failure to make actual engineering accessible to normal people. I can believe that for a lot of people a usually-good-enough ML solution is easier, even though we know it’s profoundly stupid and wasteful.   @dalias Can you tell us more about why machine learning isn’t a good fit for this? Though I didn’t even know throttling speeds could possibly improve quality. Knowing little it sounds possible? But what tipped you off that it would not work? |
Gregory's Server
This is a problem domain where the constraints and effects are pretty much entirely comprehensible in terms of known physical models. Any suboptimal behavior is entirely a matter of nobody having spent the time to apply known models. But sure, let's instead spend the time hooking up ML, CV to evaluate results, and waste tons (literally) of plastic training a model to learn a poor approximation of what we already know.
But this is a general pattern that's terrifying...