Discussion
Something that's been nagging at me about the current wave of LLMs and other machine learning techniques is how utterly *crap* they are at actually learning.
Don't get me wrong, the machine vision stuff and suchlike things are pretty amazing and all, but they are so, *so* bad at actual learning.
For example, apparently Elon's touting that Tesla now has ten billion miles of supervised driving and was supposed to be FSD by that milestone. And it isn't.
Granted, there's a lot of stupid in the data gathering (why this moron thinks that monoscopic video is a good way to monitor what's going on in a three dimensional world is... a subject for another rant) but even still. *I* learned to drive more reliably than these systems did and I think I clocked maybe 500 miles total in practice driving. Probably less, tbh, to be well past the point that these current systems are right now.
Same's true for the LLMs. Seventy bajillion parameters from every written word they could scrape up, with millions of hours of human tagging assistance and trillions of CPU hours, and they're still not quite there. Most humans get to this point of literacy with a few hundred books (at most) and maybe a couple thousand hours of actual reading time.
So yeah. Machine learning has done some great things, but it's still really pathetically inefficient.
@wordshaper we actually just recorded a podcast about learning and one of the points I made is a classic psych point, that humans are incredible at learning from a relatively small number of examples. It is truly one of our remarkable abilities
@wordshaper culture is fascinating not just in terms of understanding the concrete realities around us but because you can see it as a vehicle for distributing an extraordinarily scalable set of structural patterns from which people extrapolate to many unimagined novel behaviors and situations (eg, cog scientist Cristine Legare gives the example, culture is how you know how to file into a room you've never seen, with people you've never met, and successfully take your seat in an auditorium)
@wordshaper humans do not even need written language for this! We learn IMMENSE patterns of behavior and expectations without books at all 😄
@grimalkina Culture is objectively wild! In a lot of ways it's like external brain-support, providing some measure of higher-level functionality and system guidance that we don't necessarily have to fully learn to use.
I have no doubt that at some point we (the computer scientist "we", that I'm very generously including myself in 😊 will figure out some digital analogs for some parts of. The current ML stuff... isn't that.
@grimalkina Machine learning's still neat! And I am *absolutely* looking forward to actual real productive uses of this -- a portable X-ray machine that does first-pass checking for a whole bunch of known diseases would be *amazing* for a large part of the world -- I'm just, I think, mostly annoyed we're using the word "learning" in all this (and "intelligence") and I'm thinking that should always get a reflexive Inigo Montoya GIF tossed at it. :)
@wordshaper yeah I get that, but you know funny enough as a human learning person the "learning" term really doesn't bother me. I think across all the experience I've had with stats, predictive modeling, ml and edtech and schools I really mostly saw people correctly take it to just mean "this changes in response to information" which honestly is a better mindset than a static one ("intelligence" privileges a static mind idea maybe bc of our cultural biases! An interesting testable question)
@grimalkina I think what bothers me about the (in my view, which is of course inarguably 100% correct 😊 use of "learning" is that it carries implications that are inaccurate and leads people to making assumptions and decisions based on that inaccurate implication. Not even in a malicious way, pareidolia is absolutely a thing, but secondary decisions are made based on primary misunderstandings, and once those get hardened into a system... it's difficult to fix, because systems are persistent.
@[email protected] @[email protected] The language space is getting pretty overloaded in this area in ways that are unhelpful, it has to be said. I recently nearly skipped over a genuinely interesting piece of machine learning (sorry!) research as it had AI in the title and was something LLMs would obviously be terrible at - but they weren't using LLMs.
> I really mostly saw people correctly take it to just mean "this changes in response to information"
It's interesting to see you (Cat) say this, as one of the issues I have with LLMs is that people assume they will learn while in reality the 'learning' already happened at the model training stage (with added bonus confusion if there is or isn't any context being stored), and the thing you're actually using doesn't change in response to information. Which is true of a lot of machine learning tools in general, not just LLMs, but LLMs have the added disguise of tracking context within a conversation (and frequently now an internet connection), so interface almost implies that longer term learning is happening (despite not having 'learning' in the name :facepalm:).
I don't really have time this morning to fully do this conversation justice, but I will say, I spend a decent amount of time talking to non-tech people in my patient communities whose understanding of LLMs is really, really confused. And in general I've tried to help them start to recognize the ways they can mislead or be helpful (for instance, I have an antivax friend who regularly uses chatgpt to check her belief in conspiracy theories). The "learning" frame has actually
@mavnn @wordshaper ....helped a lot. Because I am able to describe well, depending on what you put IN to the prompt, it will shift it around for what it "wants" to give you. And this helps her. I honestly don't have a grand claim to make about how our society at large talks about it, but just my own kind of very practical and scrappy experience trying to describe this stuff to people in ways that mitigate harm. Now she "prompts" with stuff like "tell me both sides" which is honestly a HUGE WIN
@mavnn @wordshaper So in terms of making the hidden context more visible, I have actually found it REALLY helpful to use "learning" metaphors. Regardless of their veracity tbh, I just want to try what gives people some better habits. She now texts me things like "oooh it learned that I am really scared of [x medicine] so it's probably telling me that" and has some really helpful layers of skepticism BECAUSE of the learning metaphor. 🤷♀️
@mavnn @wordshaper this is massively difficult stuff so I don't claim to have a particular answer about it but just sharing that's one reason I'm a little bit more ok with the pragmatic folk theory kind of way of talking about "learning" here. A lot of people are so very very far away from feeling comfortable with thinking about how the tool they use might be changing underneath them, and I've just found that helps.
@grimalkina @mavnn Huh. Well, OK! Practical data beats theoretical concerns, so if it's helpful I shall grumble more quietly. (What, me, not grumble? Hah, as if :)
I suspect this is another facet of "everything everywhere is a mess, nothing is ever perfect including the definition of 'perfect', so you do what you can with what you have at hand."
Life. Just a never ending series of compromises, but beats the alternative.
Install bonfire.mavnn.eu
Get the full app experience
