Putting together a couple of your predictions, I wonder if we’ll start to see a shift towards pragmatism in dating and marriage. Anti-romance to go along with anti-Romanticism. Loneliness might drive people to find companions that don’t fit the mold of passionate love that so much of our media tends to idolize. When people can satisfy sexual needs through largely destigmatized services, marriage can look more like the Puritan ideal that’s oriented around comfort and family.
Your emphasis on whatever comes after transformers was also really interesting to me. Rather than a new architecture, do you think there’s any potential for mixture of experts to extend the impact of AI?
> do you think there’s any potential for mixture of experts to extend the impact of AI?
Great question. I think that specialized modules are absolutely a path forward. However, I’m more optimistic about Toolformer than MoE to make that happen [https://arxiv.org/abs/2302.04761].
I’m probably saying something you’re already aware of, but just to frame my thoughts…When MoE first got popularized people were using it as an ensemble. Typically, when you do an ensemble you do something like split your data into A/B sets and train on A and test on B to get an accuracy for how much you’ll weight that particular model within your ensemble and then that weight will stay fixed. With MoE the original intent years ago was to make that weight dynamic, based on the input, but still use all of the experts / ensembles models. More recently, people said, let’s take that and use that MoE to route the data and not use most of the model (the experts). There’s some elegance to MoE because it’s end-to-end differentiable, which always does better in peer review. But I’m not convinced MoE is really the right tool for combining specialized modules. In my mind, it’s still just a tool for dynamically weighting multiple models.
I would bet good money that the first AGI system includes soft attention. It’s just way too effective to discard completely. But there has to be something more than making the context length longer. And adding in more experts that also use attention just seems to be putting a bandaid on the deeper problem, which is we don’t know how to really connect vectors to some sort of knowledge graph. Soft attention scales quadratically and we can’t keep doing it forever. Sure, there’s LongFormer, which is just a sliding window of attention, but empirically, it just doesn’t work that well.
However, I think that ToolFormer could be a better path forward to connecting various specialized modules / experts. It’s manual and hacky and inelegant. But having the model use ToolFormer to read/write to memory or a knowledge graph still seems like a promising option.
Hey Mr. Smith - great content as always. Looking forward to seeing how these predictions unfold over the next few years.
Would like to hear your thoughts about how social media is producing less risk. The examples you give in your post seem plausible, but there are some interesting counter arguments. YouTube prankster, only fans content, general clout seeking behavior… there’s a case to be made where social media is creating riskier/ more reckless behaviour.
In terms of the lack of originality in art, I think it has more to do with the fact that there is less filter processes through traditional gatekeepers. 50 years ago we might see a good piece of content make it to the front page. Now we will still see that piece of content, as well as 10 derivatives and remixes.
> In terms of the lack of originality in art, I think it has more to do with the fact that there is less filter processes through traditional gatekeepers
I’m glad you brought this up. It might be worth it at some point to write up a full post on this. But for now, I’d summarize my thoughts by saying that the issue is that the traditional gatekeepers have become ideologically captured. I think that elite universities have a lot of power as gatekeepers and have become very unwilling to hear anything that is outside of their thinking. They’ll “cancel” old writers because their writing from 75 years ago isn’t consistent with contemporary views. This is a relatively common thing to happen in art. For example, during the 1800’s, France had strict regulations on painting with officials only wanting to see certain styles and having an official art display for their approved work. They wouldn’t accept any work outside of that style, and the first impressionist painters, like Monet, had to show their work in rogue art shows.
I have a lot more to say on this, but for now, I might point you to a couple other articles on this. I’m not sure if they’re paywalled or not unfortunately. Alex Perez [https://www.thefp.com/p/the-fight-for-the-future-of-publishing] argues that publishing has become both risk averse and ideologically captured. An article in New York Times argues that this is also the least innovative period in art in 500 years [https://www.nytimes.com/2023/10/10/magazine/stale-culture.html]. I think that’s a little more extreme than my thinking. I just think it’s been the past couple decades.
> Would like to hear your thoughts about how social media is producing less risk…there are some interesting counter arguments. YouTube prankster, only fans content, general clout seeking behavior
I don’t have a link handy, but I do remember seeing a study that teenagers who use a lot of social media are more passive and risk averse. More fearful of real world interaction to the point where some people in their early twenties are apparently struggling to order food at a restaurant because their social anxiety has grown so high. This is driving a rise in a desire for things like safe spaces and safety culture. I might say that while people are producing more OnlyFans content, that’s probably a far safer job than being an actual sex worker in the real world. But you’re right that there are some people doing really dangerous things online to get attention and they’re willing to take some extreme risks to get them. I’ll have to find those studies I remember seeing.
Yeah it would be an interesting thing to explore, look forward to any further posts that you have on the subject.
In terms of social media, I guess having a camera in front of your face ill unless it’s different reactions. Some people shy away, while other people will act over the top.
In terms of the growing rigidity in gatekeeping institutions, I’m cautiously optimistic. I think we’re starting to turn a corner where even a lot of “woke” people are realizing the lack of pragmatism in their belief systems.
Putting together a couple of your predictions, I wonder if we’ll start to see a shift towards pragmatism in dating and marriage. Anti-romance to go along with anti-Romanticism. Loneliness might drive people to find companions that don’t fit the mold of passionate love that so much of our media tends to idolize. When people can satisfy sexual needs through largely destigmatized services, marriage can look more like the Puritan ideal that’s oriented around comfort and family.
Your emphasis on whatever comes after transformers was also really interesting to me. Rather than a new architecture, do you think there’s any potential for mixture of experts to extend the impact of AI?
> do you think there’s any potential for mixture of experts to extend the impact of AI?
Great question. I think that specialized modules are absolutely a path forward. However, I’m more optimistic about Toolformer than MoE to make that happen [https://arxiv.org/abs/2302.04761].
I’m probably saying something you’re already aware of, but just to frame my thoughts…When MoE first got popularized people were using it as an ensemble. Typically, when you do an ensemble you do something like split your data into A/B sets and train on A and test on B to get an accuracy for how much you’ll weight that particular model within your ensemble and then that weight will stay fixed. With MoE the original intent years ago was to make that weight dynamic, based on the input, but still use all of the experts / ensembles models. More recently, people said, let’s take that and use that MoE to route the data and not use most of the model (the experts). There’s some elegance to MoE because it’s end-to-end differentiable, which always does better in peer review. But I’m not convinced MoE is really the right tool for combining specialized modules. In my mind, it’s still just a tool for dynamically weighting multiple models.
I would bet good money that the first AGI system includes soft attention. It’s just way too effective to discard completely. But there has to be something more than making the context length longer. And adding in more experts that also use attention just seems to be putting a bandaid on the deeper problem, which is we don’t know how to really connect vectors to some sort of knowledge graph. Soft attention scales quadratically and we can’t keep doing it forever. Sure, there’s LongFormer, which is just a sliding window of attention, but empirically, it just doesn’t work that well.
However, I think that ToolFormer could be a better path forward to connecting various specialized modules / experts. It’s manual and hacky and inelegant. But having the model use ToolFormer to read/write to memory or a knowledge graph still seems like a promising option.
Hey Mr. Smith - great content as always. Looking forward to seeing how these predictions unfold over the next few years.
Would like to hear your thoughts about how social media is producing less risk. The examples you give in your post seem plausible, but there are some interesting counter arguments. YouTube prankster, only fans content, general clout seeking behavior… there’s a case to be made where social media is creating riskier/ more reckless behaviour.
In terms of the lack of originality in art, I think it has more to do with the fact that there is less filter processes through traditional gatekeepers. 50 years ago we might see a good piece of content make it to the front page. Now we will still see that piece of content, as well as 10 derivatives and remixes.
> In terms of the lack of originality in art, I think it has more to do with the fact that there is less filter processes through traditional gatekeepers
I’m glad you brought this up. It might be worth it at some point to write up a full post on this. But for now, I’d summarize my thoughts by saying that the issue is that the traditional gatekeepers have become ideologically captured. I think that elite universities have a lot of power as gatekeepers and have become very unwilling to hear anything that is outside of their thinking. They’ll “cancel” old writers because their writing from 75 years ago isn’t consistent with contemporary views. This is a relatively common thing to happen in art. For example, during the 1800’s, France had strict regulations on painting with officials only wanting to see certain styles and having an official art display for their approved work. They wouldn’t accept any work outside of that style, and the first impressionist painters, like Monet, had to show their work in rogue art shows.
I have a lot more to say on this, but for now, I might point you to a couple other articles on this. I’m not sure if they’re paywalled or not unfortunately. Alex Perez [https://www.thefp.com/p/the-fight-for-the-future-of-publishing] argues that publishing has become both risk averse and ideologically captured. An article in New York Times argues that this is also the least innovative period in art in 500 years [https://www.nytimes.com/2023/10/10/magazine/stale-culture.html]. I think that’s a little more extreme than my thinking. I just think it’s been the past couple decades.
> Would like to hear your thoughts about how social media is producing less risk…there are some interesting counter arguments. YouTube prankster, only fans content, general clout seeking behavior
I don’t have a link handy, but I do remember seeing a study that teenagers who use a lot of social media are more passive and risk averse. More fearful of real world interaction to the point where some people in their early twenties are apparently struggling to order food at a restaurant because their social anxiety has grown so high. This is driving a rise in a desire for things like safe spaces and safety culture. I might say that while people are producing more OnlyFans content, that’s probably a far safer job than being an actual sex worker in the real world. But you’re right that there are some people doing really dangerous things online to get attention and they’re willing to take some extreme risks to get them. I’ll have to find those studies I remember seeing.
Yeah it would be an interesting thing to explore, look forward to any further posts that you have on the subject.
In terms of social media, I guess having a camera in front of your face ill unless it’s different reactions. Some people shy away, while other people will act over the top.
In terms of the growing rigidity in gatekeeping institutions, I’m cautiously optimistic. I think we’re starting to turn a corner where even a lot of “woke” people are realizing the lack of pragmatism in their belief systems.