8 ms·

I'm also not entirely convinced by "huge" context models just yet, especially as it relates to fuzzy knowledge such as overarching themes or writing style. In

by fzliu 3y ago

I'm also not entirely convinced by "huge" context models just yet, especially as it relates to fuzzy knowledge such as overarching themes or writing style. In particular, there are 0 mentions of the phrase "machine learning" in The Great Gatsby, so adding one sentence that introduces the phrase should be easy for self-attention to pick out.

EGreg 3y ago
This sounds like all the other skepticism about what AI can do. And then it can spot 200x more than any human and correlate it into common themes, and you’ll say what?
- devmor 3y ago
  Doing more than a human can isn't impressive. Most computer problems for any purpose can do more of something, or something faster than a human can. A better comparison would be if it can pick out any differences that can't be picked out by more traditional and simple algorithms.
  - EGreg 3y ago
    Or course it can very soon, since those were also written by humans. Like AlphaZero vs Rybka
  - chaxor 3y ago
    It does, using this method. My immediate thought as well was '... Yeah, well vimdiff can do that in milliseconds rather than 22 seconds' - but that's obviously missing the point entirely. Of course, we need to tell people to use the right tool for the job, and that will be more and more important to remind people of now. However, it's pretty clear that the reason they used this task is to give something simple to understand what was done in a very simple example. Of course it can do more semantic understanding related tasks, because that's what the model does. So, without looking at the details we all know that it can summarize full books, give thematic differences between two books, write what a book may be like if a character switch from one book to another is done, etc. If it doesn't do these things (not just badly, but can't at all) I would be surprised. If it does them, but badly, I wouldn't be surprised, but it also wouldn't be mind bending to see it do better than any human at the task as well.
    - Frost1x 3y ago
      >Of course it can do more semantic understanding related tasks, because that's what the model does. The problem is that marketing has eroded maintaining any such faith. Too often, simple examples are given to the consumer to extrapolate intended functionality because there's no false advertising involved then. It's used over and over again in products, the examples are well selected and don't actually show a good representation of perceived functionality. As such, I personally can't make the leap to of course it can do more semantic understanding related tasks like a diff that's not as simple, one where perhaps a characters overall personality over the course of the book is shifted, not just a single line that defines their profession. This isn't to say the demonstrative example isn't neat on its own accord given whats going on here, it is, I'm just saying I can't make such leaps from examples given by any products. When I work with vendors of traditional software, this happens all the time people dance around a lack of functionality you obviously want or need to make a sell. It's only when you force them to be explicit on the specific cases, especially in writing, that I have any faith at all.
      - xcv123 3y ago
        LLM's are specifically designed and trained for semantic understanding related tasks. That's the entire point. They are trained to solve natural language understanding tasks and they build a semantic model in order to solve the tasks.
      - chaxor 3y ago
        I did say it may do certain tasks with low performance. What you're saying is not really understandable (or simply wrong). The fact that it does understand how to do a 'simple' task such as finding where text is different is actually somewhat impressive given the typical training data for these models. But I suppose you need to actually understand the field of NLP to understand why that is. If you're expecting perfection or magic then you will be disappointed.
    - ehnto 3y ago
      I think it's fair to say most people don't know what a diff tool is, but do know how to ask questions. That is the democratizing factor that AI is introducing I feel, giving high powered computing to people without the need for specialized knowledge.
- tanseydavid 3y ago
  >> And then...you’ll say what? USER: It's a stochastic parrot. GPT: I know you are, so what am I?
lumost 3y ago
I'd be more impressed if it could rewrite Mr. Carraway as an ML engineer in the entire novel. However it's not intrinsically clear that it cannot do this... It'll be tough to find good benchmarks on long context windows. A human cannot label using 100k tokens of context.
- zooch 3y ago
  My thoughts exactly - rewrite the novel with Mr. Carraway as an ML engineer while maintaining themes/motifs (possible adding new ones too). I'm guessing what's impressive is that these are the first steps towards something like this? Or is it already possible? Someone please correct me here.
  - gcanyon 3y ago
    Or rewrite The Count of Monte Cristo as a science fiction novel and get The Stars My Destination. Or rewrite Heinlein's Double Star into present-day and get the movie Dave.
- saagarjha 3y ago
  Wonder if they’d set it in SF then…

by fzliu 3y ago

EGreg 3y ago

This sounds like all the other skepticism about what AI can do. And then it can spot 200x more than any human and correlate it into common themes, and you’ll say what?

devmor 3y ago
Doing more than a human can isn't impressive. Most computer problems for any purpose can do more of something, or something faster than a human can. A better comparison would be if it can pick out any differences that can't be picked out by more traditional and simple algorithms.
- EGreg 3y ago
  Or course it can very soon, since those were also written by humans. Like AlphaZero vs Rybka
- chaxor 3y ago
  It does, using this method. My immediate thought as well was '... Yeah, well vimdiff can do that in milliseconds rather than 22 seconds' - but that's obviously missing the point entirely. Of course, we need to tell people to use the right tool for the job, and that will be more and more important to remind people of now. However, it's pretty clear that the reason they used this task is to give something simple to understand what was done in a very simple example. Of course it can do more semantic understanding related tasks, because that's what the model does. So, without looking at the details we all know that it can summarize full books, give thematic differences between two books, write what a book may be like if a character switch from one book to another is done, etc. If it doesn't do these things (not just badly, but can't at all) I would be surprised. If it does them, but badly, I wouldn't be surprised, but it also wouldn't be mind bending to see it do better than any human at the task as well.
  - Frost1x 3y ago
    >Of course it can do more semantic understanding related tasks, because that's what the model does. The problem is that marketing has eroded maintaining any such faith. Too often, simple examples are given to the consumer to extrapolate intended functionality because there's no false advertising involved then. It's used over and over again in products, the examples are well selected and don't actually show a good representation of perceived functionality. As such, I personally can't make the leap to of course it can do more semantic understanding related tasks like a diff that's not as simple, one where perhaps a characters overall personality over the course of the book is shifted, not just a single line that defines their profession. This isn't to say the demonstrative example isn't neat on its own accord given whats going on here, it is, I'm just saying I can't make such leaps from examples given by any products. When I work with vendors of traditional software, this happens all the time people dance around a lack of functionality you obviously want or need to make a sell. It's only when you force them to be explicit on the specific cases, especially in writing, that I have any faith at all.
    - xcv123 3y ago
      LLM's are specifically designed and trained for semantic understanding related tasks. That's the entire point. They are trained to solve natural language understanding tasks and they build a semantic model in order to solve the tasks.
    - chaxor 3y ago
      I did say it may do certain tasks with low performance. What you're saying is not really understandable (or simply wrong). The fact that it does understand how to do a 'simple' task such as finding where text is different is actually somewhat impressive given the typical training data for these models. But I suppose you need to actually understand the field of NLP to understand why that is. If you're expecting perfection or magic then you will be disappointed.
  - ehnto 3y ago
    I think it's fair to say most people don't know what a diff tool is, but do know how to ask questions. That is the democratizing factor that AI is introducing I feel, giving high powered computing to people without the need for specialized knowledge.
tanseydavid 3y ago
>> And then...you’ll say what? USER: It's a stochastic parrot. GPT: I know you are, so what am I?

lumost 3y ago

I'd be more impressed if it could rewrite Mr. Carraway as an ML engineer in the entire novel. However it's not intrinsically clear that it cannot do this... It'll be tough to find good benchmarks on long context windows. A human cannot label using 100k tokens of context.

zooch 3y ago
My thoughts exactly - rewrite the novel with Mr. Carraway as an ML engineer while maintaining themes/motifs (possible adding new ones too). I'm guessing what's impressive is that these are the first steps towards something like this? Or is it already possible? Someone please correct me here.
- gcanyon 3y ago
  Or rewrite The Count of Monte Cristo as a science fiction novel and get The Stars My Destination. Or rewrite Heinlein's Double Star into present-day and get the movie Dave.
saagarjha 3y ago
Wonder if they’d set it in SF then…