Why aren't implementation based questions welcome on this stack?

Question

I've seen many times questions get asked to find another stack (usually Data-Science or Cross-Validated) because they're focused on either implementation or coding specific questions which are clearly stated as Not to be asked on this site by the site's acceptable topics. My question is why this decision was made, along with is it worth reconsidering?

Recently it has been asked when this stack will get out of beta? The response was that there is not enough traffic / helpful content being created. Right now, because of the AI boom of the last decade along with large companies placing bottomless resources into higher-level frameworks (tensorflow, torch, keras, sonnet, etc..) making entering this field as a developer very attractive/easy to the younger generation, they will have a lot of questions on these topics!

This increased push in traffic also means more answers. If we get more people coming to ask questions, more will come to answer. And if they're already answering one, who knows, they may answer another, and so on (Dominoes).

The counter-claim to this idea is probably along the lines of: Implementation based questions go against the purpose of this stack which is meant to be driven by theory, ethics, and societal impacts

My argument to this is: For the longest time, experimentation followed this track:

1. Learn
2. Research
3. Experiment (implementation is a subset of this)
4. Wash Rinse Repeat

But with these resources flowing in, and open-sourced projects are becoming popularized, we see people becoming ML practitioners without a clue of what they're doing. Using object detection api people can train up a ~SoTA model on their own dataset right out of the box almost and place something like that on their resume. From there, there exist 2 types of people: The ones who are complacent with what they achieved (by just using others' code on their data) or the ones that start to think, How does this work?, or how could I adjust this to also do that?

The latter people I feel are definitely a category of people this site wants to attract (if I've understood this site's purposes correctly). Through implementation, they are trying to understand the process. I understand the counter-claim to this is that questions should be generalized to try to assist as many people as possible, but the number of people this category would invite would make up for this difference (I have no proof/ pure speculation). Granted you want to make sure not to attract questions from the former group (people who just want to use code out of the box and go on their merry way), so you could say implementation questions have to be focused on achieving certain levels of functionality that they are struggling to achieve rather than debugging/package-specific questions.

I.E. I can extrapolate why the rule was made the way it was, but I think it's not taking into account this new, kind of odd line of work that has been born

1. Experiment
2. Learn
3. Research
4. Wash Rinse Repeat

Answer 1

If we want to accept all or most questions that are related to AI and are also on-topic on the Data Science SE, Cross Validated SE, and Stack Overflow websites, then we'd better just merge the websites. We focus on the theoretical and philosophical aspects of AI, but I would not say that all implementation-related questions are off-topic here (but we need to define precisely which ones can be on-topic, which has not yet been done, AFAIK). However, questions that involve the debugging of source code (like "Why am I getting this TypeError in this machine learning program?") should be considered off-topic, because there is already Stack Overflow for these. AFAIK, this website was created because there wasn't yet a website dedicated to the philosophical (and, partially, theoretical) aspects of AI.

(There are a lot of questions on Stack Overflow, Data Science SE and Cross Validated SE that would be better asked here, including some of the questions I had asked there. For example, What exactly is bootstrapping in reinforcement learning?. I remember I had asked it there because, at the time, I had almost no hope in this website and I thought it would not have had a future, given the number of poor questions and answers that I used to see and the small number of competent regular users. I suppose that, if I had asked that question on this website, it would not have received so much attention. There are still users on this website that degrade its quality, because they do not follow the SE standards or because they are just trolling. Furthermore, I think that moderators on this website are too slow and are hesitant to take action regarding certain questions that are not compliant with the supposed goal of the website. For example, there are a lot of broad questions on this website, which could have been avoided, if we had more active moderators that follow the rules. During this year, I've invested a lot of time on this website, so I believe that, in general, the quality of the website (answers and questions) has increased (but this is just my perception of the situation). By the way, I believe that this website still lacks more competent people in certain areas, such as geometric deep learning, POMDP, hierarchical RL, swarm intelligence, etc. People that give good answers on this website are the usual suspects. We need more diversity and perspectives.)

Which implementation-related questions should be on-topic here? I believe that this is a question that should be asked on this meta (if it wasn't already asked). In any case, I believe we should still focus on the theoretical and philosophical aspects of AI. If you also like to answer questions related to implementation issues, then you'd better also use other dedicated websites, such as Stack Overflow and Data Science.

From there, there exist 2 types of people: The ones who are complacent with what they achieved (by just using others' code on their data) or the ones that start to think, How does this work?, or how could I adjust this to also do that?

The latter people I feel are definitely a category of people this site wants to attract (if I've understood this site's purposes correctly). Through implementation, they are trying to understand the process. I understand the counter-claim to this is that questions should be generalized to try to assist as many people as possible, but the number of people this category would invite would make up for this difference (I have no proof/ pure speculation)

There is Data Science SE for these people. However, maybe some implementation-related questions that also involve theoretical or philosophical aspects could also be on-topic here. For example, "How is this concept usually implemented?".