Fair, I’ll never argue against learning as much as one can, but I’m also an intrinsically curious adult like that. That said, in the context of a 12 year old kid not interested in tech, you can lead a horse to water, but you can’t make it drink, even if it is really important water that prepares horses for the future.
I’d also like to raise the semi-philosophical rhetorical: does the robot know how to code or create images?
I’d argue that LLMs have the equivalent of text prediction on steroids as trained by pattern matching all the code on GitHub and Stack Overflow.* The second you do something outside the training data (like work with a company’s super secret internal tooling, use a library that’s fallen out of favor, or worse yet use a library with a confusingly similar name to another more popular library) that “knowledge” is less than useless.
Similarly, spitting out an image that looks like the training data doesn’t imply an understanding of composition, lightning, color, alignment. leading lines, or any of the other things that make for a compelling image/ photograph/ painting/ etc. Nor does it implicitly translate to being able to create that image in any medium other than digital.
* there’s technically also a semantic argument to be made about robots being able to code since the advent of the compiler, but I know that’s not what you meant.
[…] can’t make it drink, even if it is really important water that prepares horses for the future.
Yeah, that’s true enough
As for
Does the robot know how to code or create images?
I would posit that it doesn’t really matter, what matter is that the output is good enough to be used, and thus it will be used. Which is what makes it really important to properly learn these tools, how they work, their limitations, etc. because sooner or later, you are most likely going to be forced into using them as part of your job. As that’s how our society operates. If it it increases productivity, there are really grave incentives to use it
The second you do something outside the training data […] that “knowledge” is less than useless.
Similarly, spitting out an image that looks like the training data […] or any of the other things that make for a compelling image/ photograph/ painting/ etc.
You know, I used to agree with this a lot, but as more things have been happening, more results have been published, more development in this space has occurred, more advanced models being released, and me reading more into it… I’m not so sure anymore. Turns out that there really are results coming out of this, and that I was simply… wrong. At this point I believe that I was wrong about this technology.
LLMs in software can be pointed at documentation and “learn” how to use some internal tooling, even if it’s not in the database at all. Apparently they can solve unsolved problems like in math or computer science too. It really looks like you can get novel output and research from them, even if still somewhat limited and expensive (and some nuance how valid some results are etc.), but the fact that we’re even here, seriously discussing whether a system we created did research, is beyond imagination
I’d argue that LLMs have the equivalent of text prediction on steroids as trained by pattern matching
And there’s one counterargument to this that I’ve heard which was interesting. And it goes something like
“imagine your task is to predict the next word of a research article about something that’s not done before, over and over, until you have a complete research article. And you optimize for the highest degree of believability and accuracy. How would you go about this? You could of course just put together many reasonably sounding words good enough, but it won’t seem very believable/accurate for a research article on something novel if it doesn’t match what is actually true. So, to improve accuracy, you might want to spend time actually trying to understand the problem, maybe even do research, and work through different solutions, until you find the best match. Basically, in order to predict the most likely next word, you’re gonna want to dedicate resources and space into a “real” understanding, and build the prediction off of that. In other words, the prediction mechanism is just a convenient vehicle to promote actual, more advanced, processes. Yes, all it does is predict the next word, but in order to predict the next word with the highest amount of accuracy, it is useful to, well, reason about it.”
Basically, this, together with the scaling hypothesis, posits that maybe, just maybe, if we make big enough models and give them enough time to “think” through problems (which reasoning models do, often for a difficult problem they’re tasked to just sit there and work through whatever it is they’re asked to do for hours before giving their actual response) then we might actually get something approaching understanding. Or at least close enough to not be easily distinguishable
And believe me when I say that I used to think that they couldn’t possibly understand or reason too. There are many societal level problems with the technology and we as a society are not ready for it, as well as my livelihood being threatened (or at least changing into something I find less exciting), and so I have a cognitive bias towards, well, denying that there really is something there. But honestly, I’m not so sure anymore. There really might be something there after all, and we’re heading into something completely unprecedented. Which is why, now more than ever, it’s incredibly important to properly understand and explore how to navigate a new world shaped by this
Fair, I’ll never argue against learning as much as one can, but I’m also an intrinsically curious adult like that. That said, in the context of a 12 year old kid not interested in tech, you can lead a horse to water, but you can’t make it drink, even if it is really important water that prepares horses for the future.
I’d also like to raise the semi-philosophical rhetorical: does the robot know how to code or create images?
I’d argue that LLMs have the equivalent of text prediction on steroids as trained by pattern matching all the code on GitHub and Stack Overflow.* The second you do something outside the training data (like work with a company’s super secret internal tooling, use a library that’s fallen out of favor, or worse yet use a library with a confusingly similar name to another more popular library) that “knowledge” is less than useless.
Similarly, spitting out an image that looks like the training data doesn’t imply an understanding of composition, lightning, color, alignment. leading lines, or any of the other things that make for a compelling image/ photograph/ painting/ etc. Nor does it implicitly translate to being able to create that image in any medium other than digital.
* there’s technically also a semantic argument to be made about robots being able to code since the advent of the compiler, but I know that’s not what you meant.
Yeah, that’s true enough
As for
I would posit that it doesn’t really matter, what matter is that the output is good enough to be used, and thus it will be used. Which is what makes it really important to properly learn these tools, how they work, their limitations, etc. because sooner or later, you are most likely going to be forced into using them as part of your job. As that’s how our society operates. If it it increases productivity, there are really grave incentives to use it
You know, I used to agree with this a lot, but as more things have been happening, more results have been published, more development in this space has occurred, more advanced models being released, and me reading more into it… I’m not so sure anymore. Turns out that there really are results coming out of this, and that I was simply… wrong. At this point I believe that I was wrong about this technology.
LLMs in software can be pointed at documentation and “learn” how to use some internal tooling, even if it’s not in the database at all. Apparently they can solve unsolved problems like in math or computer science too. It really looks like you can get novel output and research from them, even if still somewhat limited and expensive (and some nuance how valid some results are etc.), but the fact that we’re even here, seriously discussing whether a system we created did research, is beyond imagination
And there’s one counterargument to this that I’ve heard which was interesting. And it goes something like
“imagine your task is to predict the next word of a research article about something that’s not done before, over and over, until you have a complete research article. And you optimize for the highest degree of believability and accuracy. How would you go about this? You could of course just put together many reasonably sounding words good enough, but it won’t seem very believable/accurate for a research article on something novel if it doesn’t match what is actually true. So, to improve accuracy, you might want to spend time actually trying to understand the problem, maybe even do research, and work through different solutions, until you find the best match. Basically, in order to predict the most likely next word, you’re gonna want to dedicate resources and space into a “real” understanding, and build the prediction off of that. In other words, the prediction mechanism is just a convenient vehicle to promote actual, more advanced, processes. Yes, all it does is predict the next word, but in order to predict the next word with the highest amount of accuracy, it is useful to, well, reason about it.”
Basically, this, together with the scaling hypothesis, posits that maybe, just maybe, if we make big enough models and give them enough time to “think” through problems (which reasoning models do, often for a difficult problem they’re tasked to just sit there and work through whatever it is they’re asked to do for hours before giving their actual response) then we might actually get something approaching understanding. Or at least close enough to not be easily distinguishable
And believe me when I say that I used to think that they couldn’t possibly understand or reason too. There are many societal level problems with the technology and we as a society are not ready for it, as well as my livelihood being threatened (or at least changing into something I find less exciting), and so I have a cognitive bias towards, well, denying that there really is something there. But honestly, I’m not so sure anymore. There really might be something there after all, and we’re heading into something completely unprecedented. Which is why, now more than ever, it’s incredibly important to properly understand and explore how to navigate a new world shaped by this