28 ms·

Recent Carmack YouTube interview with him saying the code for AGI will be simple: https://m.youtube.com/watch?v=xLi83prR5fg https://m.youtube.com/watch?v=xLi83

Recent Carmack YouTube interview with him saying the code for AGI will be simple: https://m.youtube.com/watch?v=xLi83prR5fg https://m.youtube.com/watch?v=xLi83prR5fg

raverbashing 4y ago
I'm sure he the one that could write it in only a few blocks of x86 assembly and off you go
- O__________O 4y ago
  My understanding is that his point was that it’s if you knew what to write, as a single person, it is doable and compared to anything else at this point in time would have an impact on humanity like no other.
nomel 4y ago
> saying the code for AGI will be simple To be fair, it will most likely be some python imports, for the most of it, with complex abstractions tied together in relatively simple ways. Just look at most ML notebooks, where "simple" code can easily mean "massive complexity, burning MW of power, distributed across thousands of computers".
- O__________O 4y ago
  No, not what he means, he means code will be simple enough that a single person would be able to write it, if then knew what to write and will bootstrap itself into existence for that simple code and vast amounts of external resources viable via humans, data, etc.
  - gwern 4y ago
    One interesting paper on estimating the complexity of code: http://www.offconvex.org/2021/04/07/ripvanwinkle/ http://www.offconvex.org/2021/04/07/ripvanwinkle/
  - nomel 4y ago
    > a single person would be able to write it I don't understand the distinction. A single person can't really write a code now, if you'll disallow ML libraries. You need a complier, libraries, an OS, etc. The vast majority of us work in the highest levels of abstraction possible. That's where the practicality and productivity is. I don't think he's saying "Anyone will be able to sit down and write it in assembly". I think he's saying, it will make sense when it's written, with the abstractions as they are, with hindsight clarity of "we just had to...".
    - mdonahoe 4y ago
      Pretty sure that was Carmack's point. He said the code for AGI would be simpler, in terms of number of statements, than say the code for a web browser.
cgrealy 4y ago
I tend to think Carmack is right in that the “seed” code that generates an AGI will be relatively small, but I think the “operating” code will be enormous.

Recent Carmack YouTube interview with him saying the code for AGI will be simple: https://m.youtube.com/watch?v=xLi83prR5fg https://m.youtube.com/watch?v=xLi83prR5fg

raverbashing 4y ago

I'm sure he the one that could write it in only a few blocks of x86 assembly and off you go

O__________O 4y ago
My understanding is that his point was that it’s if you knew what to write, as a single person, it is doable and compared to anything else at this point in time would have an impact on humanity like no other.

nomel 4y ago

> saying the code for AGI will be simple To be fair, it will most likely be some python imports, for the most of it, with complex abstractions tied together in relatively simple ways. Just look at most ML notebooks, where "simple" code can easily mean "massive complexity, burning MW of power, distributed across thousands of computers".

O__________O 4y ago
No, not what he means, he means code will be simple enough that a single person would be able to write it, if then knew what to write and will bootstrap itself into existence for that simple code and vast amounts of external resources viable via humans, data, etc.
- gwern 4y ago
  One interesting paper on estimating the complexity of code: http://www.offconvex.org/2021/04/07/ripvanwinkle/ http://www.offconvex.org/2021/04/07/ripvanwinkle/
- nomel 4y ago
  > a single person would be able to write it I don't understand the distinction. A single person can't really write a code now, if you'll disallow ML libraries. You need a complier, libraries, an OS, etc. The vast majority of us work in the highest levels of abstraction possible. That's where the practicality and productivity is. I don't think he's saying "Anyone will be able to sit down and write it in assembly". I think he's saying, it will make sense when it's written, with the abstractions as they are, with hindsight clarity of "we just had to...".
  - mdonahoe 4y ago
    Pretty sure that was Carmack's point. He said the code for AGI would be simpler, in terms of number of statements, than say the code for a web browser.

cgrealy 4y ago

I tend to think Carmack is right in that the “seed” code that generates an AGI will be relatively small, but I think the “operating” code will be enormous.