4 ms·

The compression up, and the necessary expansion with inventing gap-filling down (inference) is one of the reasons (together with the general specification probl

by ci5er 7y ago

The compression up, and the necessary expansion with inventing gap-filling down (inference) is one of the reasons (together with the general specification problem) that I don't think that AI is going to take over the world soon. That is: The generalists don't know how to specify in detail - they rely on successive layers down to invent what goes in the gap. And then ultimately, at the bottom layer - the code is the specification. Attempting to describe to an AI what it needs to generate code is already what we use compilers and interpreters for. If we aren't clear, in detail, about what we want, what good would an AI even do? The whole AI will replace coders is an age-old dream that used to be expressed by drag-and-drop code-modules that would allow codge-gen to be directed by business domain experts. See how that worked out...

scarejunba 7y ago
Fascinating view. Also somewhat novel to me, but acts in the same manner as something that management experts like Deming and safety experts like Sidney Dekker talk about: the magic is in the human details at the point where the rubber hits the road. Those humans are actively making conscious decisions about things. The engineer deciding to encode some data model in is using his human intuition and intelligence to trade off future proofing vs ongoing cost etc. using some picture he’s designed in his head about the product. He can’t always ask the CEO about the deets. He’s filling in gaps because the compression is guaranteed lossy. Interesting view that that would hamper an effective AI. And cool analogy to the drag and drop thing.
- ci5er 7y ago
  It's like involving the product owner in deciding whether to use a linked-list or a bucket or a hashmap for object/DS traversal (say, a list of triangles in a 3D scene). Most of the time - you just wouldn't involve them. The coder decides and implements and may or may not even report off what was decided and why. That's the gap-filling that I think it would be hard for an AI to make (of course there are a lot of space-time-memory tradeoff costing matrices that could support an AI decision process). What I find fascinating about humans is how idealistic we are (as a species). "Wouldn't it be nice if a business domain expert could just "specify" applications without the need for technical specs or coders?" (It's just like a compiler, but smarter, right?) Or (not to get too political), "Wouldn't it be nice if we could all share economic outcomes more equally without specific consideration as to contribution?" Well, yes, it would! But, as the old trope goes: Capitalism is for sinners, and socialism is for saints - and we ain't no saints. So, I believe that most humans do their best every day to do the best, in their situation, with what they know. But individually, we can't know everything. So, heirarchal management structures are a hack to sorta/kinda solve a problem with a constraint. 1) No individual is a panopticon or even if a polymath, capable of industrial-scale output, and 2) We are apparently tribal hierarchal monkeys, and we operate better (as individuals of the human species) when there is a chain of command - where no individual can effectively directly manage more than 8~15 reports. Thanks for your comment. Hopefully, I didn't drive too far off into the ditch in my response.
ilaksh 7y ago
https://github.com/ogmacorp/OgmaNeo2/blob/RL/SPH_Presentation.pdf https://github.com/ogmacorp/OgmaNeo2/blob/RL/SPH_Presentatio... See slides like Hierarchy
- ci5er 7y ago
  I don't know how you made that leap - but this is pretty spiffy! Thank you!
  - ilaksh 7y ago
    That AI uses multiple layers and feedback loops to gain an understanding of it's input environment in order to predict changes in the environment.
    - ci5er 7y ago
      Certainly neural models do!

by ci5er 7y ago

scarejunba 7y ago

Fascinating view. Also somewhat novel to me, but acts in the same manner as something that management experts like Deming and safety experts like Sidney Dekker talk about: the magic is in the human details at the point where the rubber hits the road. Those humans are actively making conscious decisions about things. The engineer deciding to encode some data model in is using his human intuition and intelligence to trade off future proofing vs ongoing cost etc. using some picture he’s designed in his head about the product. He can’t always ask the CEO about the deets. He’s filling in gaps because the compression is guaranteed lossy. Interesting view that that would hamper an effective AI. And cool analogy to the drag and drop thing.

ci5er 7y ago
It's like involving the product owner in deciding whether to use a linked-list or a bucket or a hashmap for object/DS traversal (say, a list of triangles in a 3D scene). Most of the time - you just wouldn't involve them. The coder decides and implements and may or may not even report off what was decided and why. That's the gap-filling that I think it would be hard for an AI to make (of course there are a lot of space-time-memory tradeoff costing matrices that could support an AI decision process). What I find fascinating about humans is how idealistic we are (as a species). "Wouldn't it be nice if a business domain expert could just "specify" applications without the need for technical specs or coders?" (It's just like a compiler, but smarter, right?) Or (not to get too political), "Wouldn't it be nice if we could all share economic outcomes more equally without specific consideration as to contribution?" Well, yes, it would! But, as the old trope goes: Capitalism is for sinners, and socialism is for saints - and we ain't no saints. So, I believe that most humans do their best every day to do the best, in their situation, with what they know. But individually, we can't know everything. So, heirarchal management structures are a hack to sorta/kinda solve a problem with a constraint. 1) No individual is a panopticon or even if a polymath, capable of industrial-scale output, and 2) We are apparently tribal hierarchal monkeys, and we operate better (as individuals of the human species) when there is a chain of command - where no individual can effectively directly manage more than 8~15 reports. Thanks for your comment. Hopefully, I didn't drive too far off into the ditch in my response.

ilaksh 7y ago

https://github.com/ogmacorp/OgmaNeo2/blob/RL/SPH_Presentation.pdf https://github.com/ogmacorp/OgmaNeo2/blob/RL/SPH_Presentatio... See slides like Hierarchy

ci5er 7y ago
I don't know how you made that leap - but this is pretty spiffy! Thank you!
- ilaksh 7y ago
  That AI uses multiple layers and feedback loops to gain an understanding of it's input environment in order to predict changes in the environment.
  - ci5er 7y ago
    Certainly neural models do!