4 ms·

> You don't realize it but this sentence is a declaration equivalent of saying "I'm a biased person." That was the point of my sentence lol, I'm explicitly sta

> You don't realize it but this sentence is a declaration equivalent of saying "I'm a biased person." That was the point of my sentence lol, I'm explicitly stating my bias in order to address it. Emotionally and for my own personal benefit, I want breakthroughs in AI to be achieved in my lifetime, and I hope they do. I'm explicitly acknowledging and stating my bias for this. You address bias by recognizing it and taking it into account. Then, I'm contextualizing my emotional stance, by providing data points about the reality of what the current state of AI is so that others have the context required so they can apply their own judgement. Look, I get it, it's clear we're both excited about the prospects of AI, but you're getting emotionally defensive, that's exactly what I'm trying to avoid, let's remove our clout of excitement, and contextualize by looking at the data. Yes we've made a lot of progress, but there's a lot more required, and unfortunately progress in any field of research has never followed a linearly constant progression, there was a AI winter before, there could be another. The claim that we're close to having machines able to do everything that humans can do and do it even better doesn't seem to follow from the data. There are gaps as I provided, we're only achieving close to human ability for tasks where we have hundreds of millions of examples. Even in those scenarios, there are catastrophic failure points in the models that humans would never fail at. Hardware wise, we're approaching the limits of how big a DNN we can train as well and how much data examples we have to feed it. Models are getting really good at predictions over data, but they've yet to show any ability for creativity and invention. Right now they all need human assistance to output an interesting result. Take GPT-3, and it can't even interpret what you're saying, you've got to tailor how you talk to it to get interesting responses. Take stable diffusion, it hasn't been able to invent a new drawing style yet, only mimicking the ones it learns as another example. That's pretty major limitations compared to a human, and those are tasks that we have the most data available to us for training. I think we'll need some other breakthrough to break past this barrier, and that I think could easily take much longer to achieve. Your assessment could be different, maybe you take in all this data and predict a more optimistic timeline for overcoming all these issues, that's fair, what I wanted is to contextualize, not push my conclusion, contextualizing means providing others with the rationale and data so they can do their own assessment and arrive at their own conclusions. There is a lot of optimism around Stable Diffusion and GPT-3 and others, you're flooded with the "OMG look at all it can do" data points, and then all the dreams of sci-fi becoming reality. There isn't as much exposure and discussions around, "okay but look at all it can't do and where the capabilities for what it does come from and how they were achieved". I wanted to provide that balance so people can have realistic expectations.

deltasevennine 4y ago
>okay but look at all it can't do and where the capabilities for what it does come from and how they were achieved No. This is not the case. Everyone has your attitude. Every programmer is downplaying the significance of it all. Literally look at the threads. Every tom dick and jane is saying it's not a big deal as if they know better. The truth is nobody knows shit. Your contextualization is just restating the obvious re-hashed opinion. "ML has limits it's not as good as the human brain it needs lots of training data, yadayadayada." Your not contextualizing. Your reaffirming the prevailing bias.
- didibus 4y ago
  > No. This is not the case. Everyone has your attitude. Every programmer is downplaying the significance of it all. Expand on the "significance" part. Because I don't feel like I'm downplaying the significance of it at all, this is revolutionary, and there's probably no stopping it, in 10 years you'll be using AI assistance for coding, writing, and art generation as a common thing, and it'll expand to other disciplines as well. It'll be everywhere in that regard. That could affect the job markets in a big way, this is a huge improvement to automation, either you'll need less artists, writers and programmers to get the same job done and so there will be less demand for that, or it'll lead to even more ambitious bigger projects and keep the demand for those jobs same but the output expectations will be higher. Self-driving cars has some harder long tail to address to be fully trusted on its own, I think it could take more than 10 years for that, but assisted driving is already here and that'll just continue to get better. I don't know about other programmers, but I don't feel like I'm downplaying the significance. The limits I've mentioned are not made up, there's a lot we do know about them. There are theories of why DNN work like the information bottleneck, and we know more and more about them each year. We know now that you need to over parameterized, that deep layers are needed for generalization, we know that the only sizable improvements have been from even larger datasets over even larger networks. Whereas a child for lots of tasks only needs a single example to learn where models need thousands+. Then we have things like the Winograd and Winogrande schemes, and we've observed that it seems the models get the right answers through statistical significance, not reasoning. Each time you adapt the question to remove the statistical significance they fall in accuracy again, until an even larger model is trained which likely just managed to capture even more statistical data points. So the problem of do the model actually understand the concept and can perform reasoning is a real one that currently has more sign of that they do not. Now it's possible that understanding concepts and performing logical reasoning as such is not as good as large statistical inference from giant corpus at performing certain tasks, and I think that's where you'll see models take over, for tasks where reasoning underperforms, or where it'll be an assisted tech where humans perform the reasoning and models the statistical inference. This bit is key, there's still a possibility that models are actually learning concepts and reasoning, maybe if they just grow even larger they will for example. If that turned true, than it would be incredible, but it could also turn false, and models are just performing statistical inference, and what happens is we simply underestimated how good statistical inference can be at a lot of tasks to the point it can fool you in thinking they're understanding concepts and doing full human reasoning. And you're right, we don't know for sure here, nobody knows for sure. In what way do you feel I'm underestimating the significance here? Or what do you see as being "significant" that I haven't listed?

deltasevennine 4y ago

>okay but look at all it can't do and where the capabilities for what it does come from and how they were achieved No. This is not the case. Everyone has your attitude. Every programmer is downplaying the significance of it all. Literally look at the threads. Every tom dick and jane is saying it's not a big deal as if they know better. The truth is nobody knows shit. Your contextualization is just restating the obvious re-hashed opinion. "ML has limits it's not as good as the human brain it needs lots of training data, yadayadayada." Your not contextualizing. Your reaffirming the prevailing bias.

didibus 4y ago
> No. This is not the case. Everyone has your attitude. Every programmer is downplaying the significance of it all. Expand on the "significance" part. Because I don't feel like I'm downplaying the significance of it at all, this is revolutionary, and there's probably no stopping it, in 10 years you'll be using AI assistance for coding, writing, and art generation as a common thing, and it'll expand to other disciplines as well. It'll be everywhere in that regard. That could affect the job markets in a big way, this is a huge improvement to automation, either you'll need less artists, writers and programmers to get the same job done and so there will be less demand for that, or it'll lead to even more ambitious bigger projects and keep the demand for those jobs same but the output expectations will be higher. Self-driving cars has some harder long tail to address to be fully trusted on its own, I think it could take more than 10 years for that, but assisted driving is already here and that'll just continue to get better. I don't know about other programmers, but I don't feel like I'm downplaying the significance. The limits I've mentioned are not made up, there's a lot we do know about them. There are theories of why DNN work like the information bottleneck, and we know more and more about them each year. We know now that you need to over parameterized, that deep layers are needed for generalization, we know that the only sizable improvements have been from even larger datasets over even larger networks. Whereas a child for lots of tasks only needs a single example to learn where models need thousands+. Then we have things like the Winograd and Winogrande schemes, and we've observed that it seems the models get the right answers through statistical significance, not reasoning. Each time you adapt the question to remove the statistical significance they fall in accuracy again, until an even larger model is trained which likely just managed to capture even more statistical data points. So the problem of do the model actually understand the concept and can perform reasoning is a real one that currently has more sign of that they do not. Now it's possible that understanding concepts and performing logical reasoning as such is not as good as large statistical inference from giant corpus at performing certain tasks, and I think that's where you'll see models take over, for tasks where reasoning underperforms, or where it'll be an assisted tech where humans perform the reasoning and models the statistical inference. This bit is key, there's still a possibility that models are actually learning concepts and reasoning, maybe if they just grow even larger they will for example. If that turned true, than it would be incredible, but it could also turn false, and models are just performing statistical inference, and what happens is we simply underestimated how good statistical inference can be at a lot of tasks to the point it can fool you in thinking they're understanding concepts and doing full human reasoning. And you're right, we don't know for sure here, nobody knows for sure. In what way do you feel I'm underestimating the significance here? Or what do you see as being "significant" that I haven't listed?