Five years ago, I was under the impression that most “machine learning” jobs were mostly just data cleaning, linear regression, working with regular data stores, and debugging stuff. Or, that was at least the meme that I heard from a lot of people. That didn’t surprise me at the time. It was easy to imagine that all the fancy research results were fragile, or hard to apply to products, or would at the very least take a long time to adapt.
But at this point it’s been quite a few years since there have existed machine learning systems that immensely impressed me. The first such system was probably AlphaGo—all the way back in 2016! AlphaGo then spun off in to multiple better faster cheaper systems that I didn’t even keep track of them. And since then I’ve lost track of the number of unrelated systems that immensely impressed me. And their capabilities are so general that I feel sure that they must be convertible into enormous economic value. I still believe that it takes a long time to boot up a company around novel research results, but I’m not actually well calibrated on how long that takes, and it’s been long enough that it’s starting to feel awkward, like my models are missing something. Here are examples of AI products that I wouldn’t have been surprised if they existed by now, but which I don’t think do. (I can imagine that many of these examples technically exist, but not at the level that I mean).
Spotify playlists that are actually just procedurally generated music of various genres
A tool that helps researchers/legislators/et cetera by summarizing papers, books, laws on demand
Tools that help people (like writers) brainstorm, flesh out ideas by generating further details, asking questions, etc
A version of photoshop but with tons of AI tools
Widely available self-driving cars
Physics simulators that are way faster
Paradigmatically different and better web search
So what’s the deal? Here’s a list of possible explanations. I’ve love to hear if anyone has evidence for any of them, or if you know of reasons not on the list.
The research results are actually not all that applicable to products; more research is needed to refine them
They’re way too expensive to run to be profitable
Yeah, no, it just takes a really long time to convert innovation into profitable, popular product
Something something regulation?
The AI companies are deliberately holding back for whatever reason
The models are already integrated into the economy and you just don’t know it.
[Question] Why hasn’t deep learning generated significant economic value yet?
Or has it, and it’s just not highly publicized?
Five years ago, I was under the impression that most “machine learning” jobs were mostly just data cleaning, linear regression, working with regular data stores, and debugging stuff. Or, that was at least the meme that I heard from a lot of people. That didn’t surprise me at the time. It was easy to imagine that all the fancy research results were fragile, or hard to apply to products, or would at the very least take a long time to adapt.
But at this point it’s been quite a few years since there have existed machine learning systems that immensely impressed me. The first such system was probably AlphaGo—all the way back in 2016! AlphaGo then spun off in to multiple better faster cheaper systems that I didn’t even keep track of them. And since then I’ve lost track of the number of unrelated systems that immensely impressed me. And their capabilities are so general that I feel sure that they must be convertible into enormous economic value. I still believe that it takes a long time to boot up a company around novel research results, but I’m not actually well calibrated on how long that takes, and it’s been long enough that it’s starting to feel awkward, like my models are missing something. Here are examples of AI products that I wouldn’t have been surprised if they existed by now, but which I don’t think do. (I can imagine that many of these examples technically exist, but not at the level that I mean).
Spotify playlists that are actually just procedurally generated music of various genres
A tool that helps researchers/legislators/et cetera by summarizing papers, books, laws on demand
Tools that help people (like writers) brainstorm, flesh out ideas by generating further details, asking questions, etc
A version of photoshop but with tons of AI tools
Widely available self-driving cars
Physics simulators that are way faster
Paradigmatically different and better web search
So what’s the deal? Here’s a list of possible explanations. I’ve love to hear if anyone has evidence for any of them, or if you know of reasons not on the list.
The research results are actually not all that applicable to products; more research is needed to refine them
They’re way too expensive to run to be profitable
Yeah, no, it just takes a really long time to convert innovation into profitable, popular product
Something something regulation?
The AI companies are deliberately holding back for whatever reason
The models are already integrated into the economy and you just don’t know it.