3 ms·

Driving a car has largely been democratized, yet most of the car driving research is done by highly credentialed phd scientists working for rich american tech c

Driving a car has largely been democratized, yet most of the car driving research is done by highly credentialed phd scientists working for rich american tech companies. Although it may sound like one, I honestly don't intend this as a strawman, but as a, "I pretty much ALWAYS expect research to occur in that way." I pessimistically think most very deep fields that require significant domain background are largely out of reach of "citizen scientists" nowadays. But even by the standard of doing research, I'm a rather uneducated (by the standards of a PHD AI researcher) programmer who has given multiple talks on AI models I developed, and uses relatively deep functionality on a day to day basis that I couldn't hope to implement the end to end of by hand. Depending on where you draw the line of democratization, I think we're already well into the promised land, and I'm only seeing continued positive improvements in this respect as well.

argonaut 10y ago
The analogy is a bit mixed up. Driving a car has been democratized. Actually going in and fixing/modifying a modern car is increasingly less accessible/democratized, though still possible with quite a bit of training. Car internals are getting more complex every generation. Likewise, using Google search / getting things recommended to you is obviously widely accessible. It is unclear how accessible ML research / going into the guts of a model actually is. While I don't doubt you've given talks about applying ML, I do have a little skepticism that the talks went any deeper than just application-related topics.
- existencebox 10y ago
  "While I don't doubt you've given talks about applying ML, I do have a little skepticism that the talks went any deeper than just application-related topics." Let me correct that misconception; every aspect of the algorithm was custom, from tokenization through classification through smoothing and postprocessing. It was a more simple model and process, but it was not exclusively application related. Far simpler than many of the deeper models I consume as a black box, for sure, but I'd defend my competencies in that I'm not exclusively a plug-and-play data engineer :) Let me also emphasize that I intentionally called out "driving" as opposed to "fixing". In fact I can't think of any car-fixing research going on just about anywhere, although I'm sure there is. My core point was that _using_ the tooling is key, and has been democratized in both driving and AI (the application layer aspect as you put it), the research point was merely an add-on that I've also seen a democratization of the knowledge needed to do even the basic level of novel algorithm development, even if I'm writing deep network papers or something of that ilk; I just find that to be the exception rather than the rule (per my "hardcore research is hard for the non-deeply-initiated once low hanging fruit has been pruned" comment)
  - argonaut 10y ago
    To be clear, many of the things you just mentioned are feature engineering-related and don't have much to do with ML. Car construction/modification research is going on. It's what car companies are doing... Driving is not a valid analogy because by that logic AI was democratized 10 years ago because everyone "uses" Google search.

argonaut 10y ago

The analogy is a bit mixed up. Driving a car has been democratized. Actually going in and fixing/modifying a modern car is increasingly less accessible/democratized, though still possible with quite a bit of training. Car internals are getting more complex every generation. Likewise, using Google search / getting things recommended to you is obviously widely accessible. It is unclear how accessible ML research / going into the guts of a model actually is. While I don't doubt you've given talks about applying ML, I do have a little skepticism that the talks went any deeper than just application-related topics.

existencebox 10y ago
"While I don't doubt you've given talks about applying ML, I do have a little skepticism that the talks went any deeper than just application-related topics." Let me correct that misconception; every aspect of the algorithm was custom, from tokenization through classification through smoothing and postprocessing. It was a more simple model and process, but it was not exclusively application related. Far simpler than many of the deeper models I consume as a black box, for sure, but I'd defend my competencies in that I'm not exclusively a plug-and-play data engineer :) Let me also emphasize that I intentionally called out "driving" as opposed to "fixing". In fact I can't think of any car-fixing research going on just about anywhere, although I'm sure there is. My core point was that _using_ the tooling is key, and has been democratized in both driving and AI (the application layer aspect as you put it), the research point was merely an add-on that I've also seen a democratization of the knowledge needed to do even the basic level of novel algorithm development, even if I'm writing deep network papers or something of that ilk; I just find that to be the exception rather than the rule (per my "hardcore research is hard for the non-deeply-initiated once low hanging fruit has been pruned" comment)
- argonaut 10y ago
  To be clear, many of the things you just mentioned are feature engineering-related and don't have much to do with ML. Car construction/modification research is going on. It's what car companies are doing... Driving is not a valid analogy because by that logic AI was democratized 10 years ago because everyone "uses" Google search.