3 ms·

I am not sure how I would or should describe my history in the field here so I won’t. I am still very active, but now on a different axis of progress from gradi

I am not sure how I would or should describe my history in the field here so I won’t. I am still very active, but now on a different axis of progress from gradient based modeling. So your “recent” (to me) perspective is interesting to me. For me there has been multiple clear upticks in the rate of progress in the last few years. Some of it insights, some of it simply resources and efforts. But together the results are clearly moving faster and faster, as in an acceleration of acceleration. It may be that it’s just an especially productive moment, in AI. But AI, like transistors, is a self reinforcing technology, so I think the big AI explosion is really happening. I would be very surprised if we don’t have general intelligences exceeding humans within a few years now. Certainly by 2030. My “surprise” isn’t so much the specific improvements in chat the last year. But how fast advances in that order are now happening relative to the long slow accumulation of progress in preceding decades and how much of the new process is enabled just by provisioning more resources. It’s kind of ridiculous in the sense that many deep learning tools are not even using all the best gradient optimization algorithms and techniques, but progress in results is happening so fast, that algorithmic waste isn’t even an issue. There is so much slack algorithmically, that big progress is already a forgone conclusion even if the tool builders were just carefully folding in past knowledge, or just iterating on straightforward efficiency improvements, and not making headway on new innovations. I haven’t ever seen a field hit this state. And maybe having a few decades of experience with a field makes a radical change in the rate of progress, and the specifics of that, easier to see. Do you have the sense that general AI is now within a decade or so or less? The variety of glass half full, vs. glass half empty judgements around every advance is interesting to me. I find different people have very different sensitivities to constant steps vs. the first and second derivatives of progress they are signaling. 2022 saw more algorithmic progress in poetry, essay writing, graphic art, debate, joke construction, etc, etc, etc. than a human could make in … well a lot more than a year. That’s a new red line we have now crossed. That’s faster than human process. Much faster.

YeGoblynQueenne 4y ago
Thanks for the clarifications. >> Do you have the sense that general AI is now within a decade or so or less? Not at all. My impression is that nobody is really working on creating general AI, and that if you take the field of machine learning research as a whole, if the purpose ever was to create general AI (which it never was), the field would be currently running backwards, far away from any kind of "intelligence", or any kind of generality at all. It is beyond clear to me that any sort of generalisation cannot be achieved by neural nets trained with gradient optimisation. That's because gradient optimisation will always, always get stuck at local optima. As you seem to know yourself, gradient optimisation has been around for many decades and I reckon there's been quite a few attempts to see if it can avoid getting stuck at local optima. And yet, well, it still gets stuck at local optima. That's probably its strength anyway, although it's also how it is self-limited. So there's only so much you can do when training with gradient optimisation. There is nothing to do about it. There is nobody who can do anything about it. All the billions of research, all the hordes of researchers, cannot get Humpty Dumpty unstuck from local optima. And so there's not a chance of a snowball in hell that general AI is going to be anywhere near where we get to in a hundred years from now, with current approaches. >> I haven’t ever seen a field hit this state. Well, I agree, but I disagree fundamentally. For me, the state of the field is deplorable! "The field" of machine learning research, that is. That field has abandoned every semblance of doing science and it has become a complete circus, where all that matters is what is the next big company that's paying billions for researchers who know how to train a neural net. Just look at the papers. Thousands of papers are placed on arxiv every month. Hundreds are even published. The vast, vast majority of them do nothing useful at all, and just claim to have moved the state of the art forward a couple of percentage points. They beat a benchmark, brag about it, and go home happy that they will be remembered... for a couple days at least, until someone else with no ideas breaks their record. It's a glorified Kaggle competition, and a free for all, but it achieves nothing. There are no theoretical results in any of the papers, except those of the very few people left at the top of the mess, who would actually know science if it bit them in the leg. There is nothing but noise. And if anyone comes up with a novel idea, in all this mess, that novel idea will be buried under a ton of dreary, miserable dross. Or, to cut this short, what you are seeing is supply and demand: big tech corps supply millions, graduate students run towards them with open arms to get the millions. Progress? What, in that situation? Not really.
- Nevermark 4y ago
  Interesting that we see the same things - everything you say is true - but see different implications. The long history of gradient descent is even funnier than it seems now. A defining problem for neural networks, that famously put the early field into a winter, was the inability to find a way to train hidden layers. In 1969, Marvin Minsky and Seymour Papert released a book that demonstrated the limitations of single layers of neurons. They can perform linear separation, and little else, as you know. And without a good way to train middle/hidden layers that was a devastating conclusion. Most people abandoned the neural approach to AI research. Then more than a decade later, a few different people realized you could apply the long known calculus derivative chain rule & gradient descent! The field picked up some momentum again. You could say, duh! That isn’t even an advance, we should have already known that! On the other hand, suddenly multi-layer networks were trainable and the field came back from the dead. Similar to now. Advances might be very minor, but their impact still profound. Today with lots more computing power and data, a machine can write poetry about cheerleading, spoken with the voice of Arnold Schwarzenegger in a Shakespearean play. And then proceed to converse in that voice. Easy to make light of, not the most useful skill. Just another step in chat achieved mostly by scaling up what we already knew how to do. But on the impact side, these models are now flexible in integrating totally different types of knowledge, from vast numbers of subjects, and returning results in real time. And as you point out, this progress is being made despite most papers and efforts being very uninspired. The lesson may be, that the means by which we achieve general AI are likely to seem very mundane. That won’t stop their impact. (And not finding global minima isn’t an impediment of any kind. Our own minds are notoriously unreliable and non-optimal. Intelligence and mathematical perfection are very different things. Virtually unrelated problems.)
  - YeGoblynQueenne 4y ago
    Sorry, missed this. And now I have no time to reply (I work on Sundays, yeah :). Thanks for the conversation, I hope we can get to pick it up again sometime.

YeGoblynQueenne 4y ago

Thanks for the clarifications. >> Do you have the sense that general AI is now within a decade or so or less? Not at all. My impression is that nobody is really working on creating general AI, and that if you take the field of machine learning research as a whole, if the purpose ever was to create general AI (which it never was), the field would be currently running backwards, far away from any kind of "intelligence", or any kind of generality at all. It is beyond clear to me that any sort of generalisation cannot be achieved by neural nets trained with gradient optimisation. That's because gradient optimisation will always, always get stuck at local optima. As you seem to know yourself, gradient optimisation has been around for many decades and I reckon there's been quite a few attempts to see if it can avoid getting stuck at local optima. And yet, well, it still gets stuck at local optima. That's probably its strength anyway, although it's also how it is self-limited. So there's only so much you can do when training with gradient optimisation. There is nothing to do about it. There is nobody who can do anything about it. All the billions of research, all the hordes of researchers, cannot get Humpty Dumpty unstuck from local optima. And so there's not a chance of a snowball in hell that general AI is going to be anywhere near where we get to in a hundred years from now, with current approaches. >> I haven’t ever seen a field hit this state. Well, I agree, but I disagree fundamentally. For me, the state of the field is deplorable! "The field" of machine learning research, that is. That field has abandoned every semblance of doing science and it has become a complete circus, where all that matters is what is the next big company that's paying billions for researchers who know how to train a neural net. Just look at the papers. Thousands of papers are placed on arxiv every month. Hundreds are even published. The vast, vast majority of them do nothing useful at all, and just claim to have moved the state of the art forward a couple of percentage points. They beat a benchmark, brag about it, and go home happy that they will be remembered... for a couple days at least, until someone else with no ideas breaks their record. It's a glorified Kaggle competition, and a free for all, but it achieves nothing. There are no theoretical results in any of the papers, except those of the very few people left at the top of the mess, who would actually know science if it bit them in the leg. There is nothing but noise. And if anyone comes up with a novel idea, in all this mess, that novel idea will be buried under a ton of dreary, miserable dross. Or, to cut this short, what you are seeing is supply and demand: big tech corps supply millions, graduate students run towards them with open arms to get the millions. Progress? What, in that situation? Not really.

Nevermark 4y ago
Interesting that we see the same things - everything you say is true - but see different implications. The long history of gradient descent is even funnier than it seems now. A defining problem for neural networks, that famously put the early field into a winter, was the inability to find a way to train hidden layers. In 1969, Marvin Minsky and Seymour Papert released a book that demonstrated the limitations of single layers of neurons. They can perform linear separation, and little else, as you know. And without a good way to train middle/hidden layers that was a devastating conclusion. Most people abandoned the neural approach to AI research. Then more than a decade later, a few different people realized you could apply the long known calculus derivative chain rule & gradient descent! The field picked up some momentum again. You could say, duh! That isn’t even an advance, we should have already known that! On the other hand, suddenly multi-layer networks were trainable and the field came back from the dead. Similar to now. Advances might be very minor, but their impact still profound. Today with lots more computing power and data, a machine can write poetry about cheerleading, spoken with the voice of Arnold Schwarzenegger in a Shakespearean play. And then proceed to converse in that voice. Easy to make light of, not the most useful skill. Just another step in chat achieved mostly by scaling up what we already knew how to do. But on the impact side, these models are now flexible in integrating totally different types of knowledge, from vast numbers of subjects, and returning results in real time. And as you point out, this progress is being made despite most papers and efforts being very uninspired. The lesson may be, that the means by which we achieve general AI are likely to seem very mundane. That won’t stop their impact. (And not finding global minima isn’t an impediment of any kind. Our own minds are notoriously unreliable and non-optimal. Intelligence and mathematical perfection are very different things. Virtually unrelated problems.)
- YeGoblynQueenne 4y ago
  Sorry, missed this. And now I have no time to reply (I work on Sundays, yeah :). Thanks for the conversation, I hope we can get to pick it up again sometime.