4 ms·

Your browsing behavior is being modeled as a language

36 points by jiwidi 4y ago

magic_hamster 4y ago
Huh. I feel as if the example given could be very easily solved with simple code instead of spinning up AI models to know you'd want to watch the next episode of the show you're currently watching (even if it's not the most substantial part of your viewing patterns). There could be an interesting implementation of NLP in the recommendations space but it also seems like giving more weight to what the user is currently interacting with is not really a reason to use NLP. Maybe I missed something.
aaaaaaaaaaab 4y ago
And it’s still mostly shit.
- ratww 4y ago
  And you're absolutely right, despite the downvotes, but here's the coldest of all hot takes: it doesn't matter if it's shit. It just has to be there. Doomscrolling and modern web products were never about quality, but rather about giving the user the hope of potentially seeing something of quality. Which means that my cousin magic_hamster who said "this can just be an algorithm" is also right [1]. In most cases there's no need to leverage AI here. 90% of everything is gonna be crap, and people historically often get addicted to low-quality crap anyway. [1] https://news.ycombinator.com/item?id=33150830 https://news.ycombinator.com/item?id=33150830
Ozzie_osman 4y ago
I usually love articles that make things simple, but this article was... Too simple. Anyway, afaik YouTube was one of the first large-scale applications of this, where they used Recurrent Neural Networks to model what video you might watch next in their recommendations (similar to how an NLP model might predict what word comes next).
- aaaaaaaaaaab 4y ago
  This is not an article. This is a so-called content marketing piece.
visarga 4y ago
I'd like to see that applied to general web browsing, as the title suggests, not just for content recommendations. Can we automate the repetitive parts of our work with a language model? Something like this is being developed by Adept.ai but they are tight lipped.
lupire 4y ago
Clickbait blogspam goop. Here's the article it's talking about: https://arxiv.org/pdf/2206.02254.pdf https://arxiv.org/pdf/2206.02254.pdf It's not "modeled as a language", it's "language problems are modeled as prediction problems, and "your next action is obviously a prediction problems."
- mattkrause 4y ago
  It’s amazing that they managed to write so much before finally hitting the key word, *sequence* at the very end. That’s really all it is: given a sequence of tokens, figure out the most likely continuation.
ourmandave 4y ago
So C-3PO isn't allowed to translate Sith, but he's totally fluent in Daves-Porn-Preferences. How is that fair?

36 points by jiwidi 4y ago

Huh. I feel as if the example given could be very easily solved with simple code instead of spinning up AI models to know you'd want to watch the next episode of the show you're currently watching (even if it's not the most substantial part of your viewing patterns). There could be an interesting implementation of NLP in the recommendations space but it also seems like giving more weight to what the user is currently interacting with is not really a reason to use NLP. Maybe I missed something.

aaaaaaaaaaab 4y ago

And it’s still mostly shit.

ratww 4y ago
And you're absolutely right, despite the downvotes, but here's the coldest of all hot takes: it doesn't matter if it's shit. It just has to be there. Doomscrolling and modern web products were never about quality, but rather about giving the user the hope of potentially seeing something of quality. Which means that my cousin magic_hamster who said "this can just be an algorithm" is also right [1]. In most cases there's no need to leverage AI here. 90% of everything is gonna be crap, and people historically often get addicted to low-quality crap anyway. [1] https://news.ycombinator.com/item?id=33150830 https://news.ycombinator.com/item?id=33150830

Ozzie_osman 4y ago

I usually love articles that make things simple, but this article was... Too simple. Anyway, afaik YouTube was one of the first large-scale applications of this, where they used Recurrent Neural Networks to model what video you might watch next in their recommendations (similar to how an NLP model might predict what word comes next).

aaaaaaaaaaab 4y ago
This is not an article. This is a so-called content marketing piece.

visarga 4y ago

I'd like to see that applied to general web browsing, as the title suggests, not just for content recommendations. Can we automate the repetitive parts of our work with a language model? Something like this is being developed by Adept.ai but they are tight lipped.

lupire 4y ago

Clickbait blogspam goop. Here's the article it's talking about: https://arxiv.org/pdf/2206.02254.pdf https://arxiv.org/pdf/2206.02254.pdf It's not "modeled as a language", it's "language problems are modeled as prediction problems, and "your next action is obviously a prediction problems."

mattkrause 4y ago
It’s amazing that they managed to write so much before finally hitting the key word, *sequence* at the very end. That’s really all it is: given a sequence of tokens, figure out the most likely continuation.

ourmandave 4y ago

So C-3PO isn't allowed to translate Sith, but he's totally fluent in Daves-Porn-Preferences. How is that fair?