3 ms·

Call me a cynic (many have, especially on this topic) but I can't help but think that the majority of what AI will "successfully" replace in terms of craftsmans

Call me a cynic (many have, especially on this topic) but I can't help but think that the majority of what AI will "successfully" replace in terms of craftsmanship is going to be stuff that would've never been produced the "correct" way if you will. It's going to be code created for and to suit the interests of the business major class. Just like AI art isn't really suitable for anything above hobby fun stuff like generating your D&D character's avatar, or product packaging stock photo junk or header images for LinkedIn blog posts. Anything that's actually important is going to still need to be designed, and that goes for creative work like design, and proper code-work for development too, IMO. Like sure, these AI's can generate code that works. Can they generate replacement code when you need to change how something works? Can they troubleshoot code that isn't doing what it's meant to? And if you can generate the code you want but then need to tweak it after to suit your purpose, is that... really that much faster than just writing the thing in your style, in a way you understand, that you can then change later as required? I dunno. I've played with these tools and they're neat, and I think they can be good for learning a new language or framework, but once I'm actually ready to build something, I don't see myself starting with AI generation for any substantial part of it.

Oioioioiio 3y ago
The question is not about what AI can do today but what we assume AI will be able to do tomorrow. All of what you wrote in your second paragraph will become something AI will be doing better and faster than you. We never had technology which can write code like this. I prompted ChatGPT to write a very basic java tool which renders an image from an url and makes it bigger on a click. It just did it. Its not hard to think further and a lot of technology is already going into this direction. Alone last week devin was showne. Gemini has a window token of 1 Million tokens. Groq shows us how it will feel to have instant response. Right now its already good enough that people with Copilot like to keep it when asked. We already now pay billions for AI daily. This means the amount of research, business motivation and money flowing into it now is probably staggering in comparision to what moved this field a few years ago. Its not clear at all how fast we will progress but i'm pretty sure, we will hit a time were every junior is worse than AI which will force people of rethinking what they are going to do. Do i hire an junior and train him/her? Or do i prefer to invest more into AI? The gap will widen and widen, a generation or a certain amount of people will stay longer and might be able to stay in development but a lot of others might just not.
- dartos 3y ago
  > We never had technology which can write code like this. I prompted ChatGPT to write a very basic java tool which renders an image from an url and makes it bigger on a click. It just did it. It's worth noting, that it can do things like that because of the large amount of "how to do simple things in java" tutorials there are on the internet. Ask an AI to _make_ java, and it won't (and will continue to not) be able to. That's the level that AI will fail at, when things aren't easily indexed from the internet and thus much harder / impossible to put into a training set. I think the technology itself (transformers and other such statistical models) have exhausted most of their low hanging fruit by now. Sora, for example, isn't a grand innovation in the way latent space models, word2vec, or transformers are, it's just a MUCH larger model than DALLE-3. which is great! but still has the limits inherit to statistical models. They need the training data.
  - ToucanLoucan 3y ago
    > It's worth noting, that it can do things like that because of the large amount of "how to do simple things in java" tutorials there are on the internet. Much like the same points made elsewhere with regard to AI art: It cannot invent. It can remix, recombine, etc. but no AI model we have now is anywhere close to where it could create something entirely new that's not been seen before.
    - Oioioioiio 3y ago
      I have not seen crabs made out of food before. What level do you think 'invention' have to be to count as something AI can't do? The only thing AI needs is a feedback loop and a benchmark/cost function. If the cost function is page impressions, thats easy. If its running unit tests based from business requirements, thats easy too.
      - dartos 3y ago
        > The only thing AI needs is a feedback loop and a benchmark/cost function. You’re forgetting about data. AI needs data. It’s arguably the most important thing any statistical model needs. That data must come from somewhere. There’s no free lunch.
__loam 3y ago
I think the question is whether we're going to plateau at 95% or not. It's possible that we just run into a wall with transformers, or they do iron it out and it does replace us all.

Oioioioiio 3y ago

The question is not about what AI can do today but what we assume AI will be able to do tomorrow. All of what you wrote in your second paragraph will become something AI will be doing better and faster than you. We never had technology which can write code like this. I prompted ChatGPT to write a very basic java tool which renders an image from an url and makes it bigger on a click. It just did it. Its not hard to think further and a lot of technology is already going into this direction. Alone last week devin was showne. Gemini has a window token of 1 Million tokens. Groq shows us how it will feel to have instant response. Right now its already good enough that people with Copilot like to keep it when asked. We already now pay billions for AI daily. This means the amount of research, business motivation and money flowing into it now is probably staggering in comparision to what moved this field a few years ago. Its not clear at all how fast we will progress but i'm pretty sure, we will hit a time were every junior is worse than AI which will force people of rethinking what they are going to do. Do i hire an junior and train him/her? Or do i prefer to invest more into AI? The gap will widen and widen, a generation or a certain amount of people will stay longer and might be able to stay in development but a lot of others might just not.

dartos 3y ago
> We never had technology which can write code like this. I prompted ChatGPT to write a very basic java tool which renders an image from an url and makes it bigger on a click. It just did it. It's worth noting, that it can do things like that because of the large amount of "how to do simple things in java" tutorials there are on the internet. Ask an AI to _make_ java, and it won't (and will continue to not) be able to. That's the level that AI will fail at, when things aren't easily indexed from the internet and thus much harder / impossible to put into a training set. I think the technology itself (transformers and other such statistical models) have exhausted most of their low hanging fruit by now. Sora, for example, isn't a grand innovation in the way latent space models, word2vec, or transformers are, it's just a MUCH larger model than DALLE-3. which is great! but still has the limits inherit to statistical models. They need the training data.
- ToucanLoucan 3y ago
  > It's worth noting, that it can do things like that because of the large amount of "how to do simple things in java" tutorials there are on the internet. Much like the same points made elsewhere with regard to AI art: It cannot invent. It can remix, recombine, etc. but no AI model we have now is anywhere close to where it could create something entirely new that's not been seen before.
  - Oioioioiio 3y ago
    I have not seen crabs made out of food before. What level do you think 'invention' have to be to count as something AI can't do? The only thing AI needs is a feedback loop and a benchmark/cost function. If the cost function is page impressions, thats easy. If its running unit tests based from business requirements, thats easy too.
    - dartos 3y ago
      > The only thing AI needs is a feedback loop and a benchmark/cost function. You’re forgetting about data. AI needs data. It’s arguably the most important thing any statistical model needs. That data must come from somewhere. There’s no free lunch.

__loam 3y ago

I think the question is whether we're going to plateau at 95% or not. It's possible that we just run into a wall with transformers, or they do iron it out and it does replace us all.