3 ms·

> Currently, we are still far from a point where machines are able to abstract high-level concepts from data or engage in reasoning and reflection Of course wh

> Currently, we are still far from a point where machines are able to abstract high-level concepts from data or engage in reasoning and reflection Of course when an AI does that, we then say its just doing statistics, not reasoning. Until you have built a recommendation engine from scratch, it is hard to appreciate the complexity. I don't mean the complexity of the code or algorithm (ALS and Spark are straightforward enough) but the contextual problem. Models end up being large collections of models in a complex hierarchy, with hyperparams to tune higher level concepts such as "surprise" or business targets such as "revenue", "engagement" etc. TikTok have nailed this, as has Spotify.

Barrin92 4y ago
>Of course when an AI does that, we then say its just doing statistics, not reasoning. no, AI simply doesn't do that. Even Demis Hassabis of Deepmind fame in a recent interview pointed this out. Machine learning is great on averaging out a large amount of data, which is often useful, but it doesn't generate true novelty in any human sense. AI can play Go, it can't invent Go. In the same way today's recommender systems are great at averaging out my last 50 shopping items or spotify playlist but they can't take a real guess at what truly new thing I'd like based on a genuine understanding of say, my personality. Which is reflected in the quality of recommendations which is mostly "the thing you just bought/watched", which is ironically often incredibly uninteresting.
- teruakohatu 4y ago
  > but it doesn't generate true novelty in any human sense. Few humans do either, even great artists create from, or react to, the art they experienced in their life. At the end of the day this is a rather metaphysical question. A Go model is trained to play Go, not create new games. > on a genuine understanding of say, my personality That is exactly what they are trying to do. The purpose of many recommendation systems is to uncover the latent variables and categories that you might like. They are not just averaging what you listened to. It is feasible (but very unwise) to predict certain users like to listen to happy upbeat tracks on Fridays, and sad songs sung in Romance languages on Mondays. Each time you skip a track or relisten to a track (implicit feedback) or like a track (explicit feedback) you are giving it information, which might be more than just the interaction (the timestamp, the IP address location, how quick you were to skip the track etc.). On the other hand, the purpose of those systems is simply to optimize what generates the most short term income, in which case, you get 'boring'/simple shopping recommendations that on average make more profit. Systems have tried to predict model the human experience, but in the past there have been epic fails so nobody is trying to explicitly do that anymore. The classic example would be stores deciding teens were pregnant and inadvertently alerting their parents. Nobody wants to see ads for Amazon's in-house toilet paper brand just as they are about to get up to go to the toilet.
  - Barrin92 4y ago
    >to predict certain users like to listen to happy upbeat tracks on Fridays, and sad songs sung in Romance languages on Mondays. This is still entirely superficial. There's no recommender system that could even tell what a 'sad track' is other than by human labelling, or distinguish say, a satirically sad track from the genuine thing, or accurately tell what any individual even perceives as happy or sad which is highly subjective. It's not about metaphysics but about a very practical fact that these statistical systems fail at. Making genuinely novel human recommendations requires and understanding of human nature in general, individual nature of the person in question, and understanding of context and meaning. In short a system like this needs to be able to reason about what it is that it's recommending and who it's talking to. There is no system that could invent a game like Go because it has nothing to do with data in any direct sense. It'd require being able to reason about what makes games compelling and aesthetically pleasing to human beings and that's not a question of mapping one button click to another.

Barrin92 4y ago

>Of course when an AI does that, we then say its just doing statistics, not reasoning. no, AI simply doesn't do that. Even Demis Hassabis of Deepmind fame in a recent interview pointed this out. Machine learning is great on averaging out a large amount of data, which is often useful, but it doesn't generate true novelty in any human sense. AI can play Go, it can't invent Go. In the same way today's recommender systems are great at averaging out my last 50 shopping items or spotify playlist but they can't take a real guess at what truly new thing I'd like based on a genuine understanding of say, my personality. Which is reflected in the quality of recommendations which is mostly "the thing you just bought/watched", which is ironically often incredibly uninteresting.

teruakohatu 4y ago
> but it doesn't generate true novelty in any human sense. Few humans do either, even great artists create from, or react to, the art they experienced in their life. At the end of the day this is a rather metaphysical question. A Go model is trained to play Go, not create new games. > on a genuine understanding of say, my personality That is exactly what they are trying to do. The purpose of many recommendation systems is to uncover the latent variables and categories that you might like. They are not just averaging what you listened to. It is feasible (but very unwise) to predict certain users like to listen to happy upbeat tracks on Fridays, and sad songs sung in Romance languages on Mondays. Each time you skip a track or relisten to a track (implicit feedback) or like a track (explicit feedback) you are giving it information, which might be more than just the interaction (the timestamp, the IP address location, how quick you were to skip the track etc.). On the other hand, the purpose of those systems is simply to optimize what generates the most short term income, in which case, you get 'boring'/simple shopping recommendations that on average make more profit. Systems have tried to predict model the human experience, but in the past there have been epic fails so nobody is trying to explicitly do that anymore. The classic example would be stores deciding teens were pregnant and inadvertently alerting their parents. Nobody wants to see ads for Amazon's in-house toilet paper brand just as they are about to get up to go to the toilet.
- Barrin92 4y ago
  >to predict certain users like to listen to happy upbeat tracks on Fridays, and sad songs sung in Romance languages on Mondays. This is still entirely superficial. There's no recommender system that could even tell what a 'sad track' is other than by human labelling, or distinguish say, a satirically sad track from the genuine thing, or accurately tell what any individual even perceives as happy or sad which is highly subjective. It's not about metaphysics but about a very practical fact that these statistical systems fail at. Making genuinely novel human recommendations requires and understanding of human nature in general, individual nature of the person in question, and understanding of context and meaning. In short a system like this needs to be able to reason about what it is that it's recommending and who it's talking to. There is no system that could invent a game like Go because it has nothing to do with data in any direct sense. It'd require being able to reason about what makes games compelling and aesthetically pleasing to human beings and that's not a question of mapping one button click to another.