But also AI cannot currently do everything, so you need someone to fill those areas.
And who is going to be able to fill those gaps? Probably not the person who “knows what I want to achieve but (…) don’t know how to actually implement it”.
Which ties in to
their capability to learn, their personality, will they mesh well with the existing team, have they got drive to make things better, do they have soft skills to position themselves to become better, is the person adaptable
is the bar for what is considered fundamental shifting?
If the bar is “I know how to ask a magic box to do my job for me” then there is genuinely no need for previous training and experience and a company won’t be hiring engineers or spreadsheet gandalfs or marketing experts. They’ll hire the cheapest “prompt engineer” they can, underpay them, and then replace them the moment they ask for a cost of living increase.
And… the companies considering that really aren’t the ones with any longevity. Yes, yes, any port in a storm. But they will RAPIDLY run into that wall and have no way to move past it. Whether that is getting the senior engineer in cargo shorts to do it or curating training data to improve the model.
but as time went on we got new levels of coding and so knowing how to write low level code is no longer a required skill.
And that is another barrier that MANY companies have run into.
The average coder? Yeah, they don’t need to understand how to optimize a loop. But when there are forty tools on the market that all just call pytorch? The one company that knows how to optimize a critical path function suddenly looks REALLY good with their 10% performance (and thus power) savings.
Again, these tools are incredibly powerful and I regularly use chatgpt et al to generate a first draft of a utility script. And I’ve been using editor plugins for… sweet Eothas over two decades now, to generate docstring stubs and even a lot of unit tests. And people SHOULD know how and when to use these tools.
But you also have to consider what you can get out of it. “AI” generated documentation is pretty much worthless outside of checking off a box that you have documented every function in the code. Your LLM won’t understand what that function was trying to achieve or why “it is wrong but that is because this library is wrong” and so forth. Any documentation that is actually meant to be referenced still needs a proper pass from whoever drew the short straw in Engineering.
Same with testing. AI can generate tautologies. AI won’t stress test your code because it doesn’t know what you think that code might do in the future. By all means, generate the boilerplate, but you are still going to be the one who has to go in and add that really weird corner case that TOTALLY didn’t break prod lats month.
And… you know who historically did those tasks? Interns and junior engineers. The same ones who are adamant that their entire job can be done by chatgpt and lamenting that they don’t know how to move from idea to implementation. And guess how you learn how to do that?
Thanks, it’s interesting to read your thoughts on this.
If you were a Jr entering the job market now, and have management encouraging vibe coding because they want quick results, how would you go about getting the experience and building the right skillset?
The first is how to develop skills. And you do that the exact same way everyone before you did it: you actually do the work. Calculators are awesome but you still learn how to do long division and the like because it gives you insight into how to approximate things. Same with sims/solvers versus actually solving PDEs.
The other is… if your boss wants you to feed everything into an LLM then you won’t have a job much longer. So you can either look for a new one or work toward more advanced tickets/tasks. Make it clear that LLMs have limitations and that some stuff will need a proper coder and that YOU are that proper coder.
And who is going to be able to fill those gaps? Probably not the person who “knows what I want to achieve but (…) don’t know how to actually implement it”.
Which ties in to
If the bar is “I know how to ask a magic box to do my job for me” then there is genuinely no need for previous training and experience and a company won’t be hiring engineers or spreadsheet gandalfs or marketing experts. They’ll hire the cheapest “prompt engineer” they can, underpay them, and then replace them the moment they ask for a cost of living increase.
And… the companies considering that really aren’t the ones with any longevity. Yes, yes, any port in a storm. But they will RAPIDLY run into that wall and have no way to move past it. Whether that is getting the senior engineer in cargo shorts to do it or curating training data to improve the model.
And that is another barrier that MANY companies have run into.
The average coder? Yeah, they don’t need to understand how to optimize a loop. But when there are forty tools on the market that all just call pytorch? The one company that knows how to optimize a critical path function suddenly looks REALLY good with their 10% performance (and thus power) savings.
Again, these tools are incredibly powerful and I regularly use chatgpt et al to generate a first draft of a utility script. And I’ve been using editor plugins for… sweet Eothas over two decades now, to generate docstring stubs and even a lot of unit tests. And people SHOULD know how and when to use these tools.
But you also have to consider what you can get out of it. “AI” generated documentation is pretty much worthless outside of checking off a box that you have documented every function in the code. Your LLM won’t understand what that function was trying to achieve or why “it is wrong but that is because this library is wrong” and so forth. Any documentation that is actually meant to be referenced still needs a proper pass from whoever drew the short straw in Engineering.
Same with testing. AI can generate tautologies. AI won’t stress test your code because it doesn’t know what you think that code might do in the future. By all means, generate the boilerplate, but you are still going to be the one who has to go in and add that really weird corner case that TOTALLY didn’t break prod lats month.
And… you know who historically did those tasks? Interns and junior engineers. The same ones who are adamant that their entire job can be done by chatgpt and lamenting that they don’t know how to move from idea to implementation. And guess how you learn how to do that?
Thanks, it’s interesting to read your thoughts on this.
If you were a Jr entering the job market now, and have management encouraging vibe coding because they want quick results, how would you go about getting the experience and building the right skillset?
There are two layers to that.
The first is how to develop skills. And you do that the exact same way everyone before you did it: you actually do the work. Calculators are awesome but you still learn how to do long division and the like because it gives you insight into how to approximate things. Same with sims/solvers versus actually solving PDEs.
The other is… if your boss wants you to feed everything into an LLM then you won’t have a job much longer. So you can either look for a new one or work toward more advanced tickets/tasks. Make it clear that LLMs have limitations and that some stuff will need a proper coder and that YOU are that proper coder.