Google gifts a Free AI Coding Assistant to the developer community

But that's IMO exactly what your parent commenter says. Use LLMs only after you actually have a clue what are they producing. So if you are a beginner, basically don't because you'll not have any understanding.

I'd argue that coding assistants for a SWE hinders their ability to build subject matter expertise.

But at the same time, having personalised StackOverflow without the negative attitude at your fingertips is super helpful, provided you also do the work of learning.

We are forgetting what general purpose name means to learning and real practical usage as a tool.

BTW the best AI and Computer Science discussions are happening on bluesky

What I've found after developing software for many decades and learning many languages is that the concepts and core logical thinking are what is most important in most cases.

Before the current AI boom I would still have had a problem doing some tasks in a vacuum as well. Not because I was incapable, but because I had so much other relevant information in my head that the minutia of some tasks was irrelevant when I had immediate access to the needed information via auto-complete in an IDE and language documentation. I know what I needed to look up because of all that other knowledge in my head though. I knew things were possible. And in cases where I didn't _know_ something was possible, I had an inkling that something might be possible because I could do it in another language or it was a logical extension of some other concept.

With the current rage of AI Coding Copilots I personally feel like many people are going down a path that degrades that corpus of general knowledge that drives the ability to solve problems quickly. Instead they lean on the coding assistant to have that knowledge and simple direct it to do the tasks at a macro level. On the surface this may seem like a universal boon, but the reality is they are giving up that intrinsic domain knowledge that is needed to be functional at understanding what software is doing and how to solve the problems that will crop up.

If those two paragraphs seem contradictory in some manner, I agree. You can argue that leaning on IDE syntax autocomplete and looking up documentation not foundationally different than leaning on a coding assistant. I can only say that they don't _feel_ the same to me. Maybe what I mean is, if the assistant is writing code and you are directly using it, then you never gain knowledge. If you are looking things up in documentation or using auto-complete for function names or arguments, you are learning about the code and how to solve a problem. So maybe it's just, what abstraction level are we, as a profession, comfortable with?

To close out this oddly long comment, I personally use LLMs and other ML models frequently. I have found that they are excellent at helping me formulate my thoughts on a problem that needs to be solved and to surface information across a lot of sources into a coherent understanding of an issue. Sure, it's possible that it's wrong, but I just use it to help steer me towards the real information I need. If I ask for or it provides code, that's used as a reference implementation for the actual implementation I write. And my IDE auto-complete has gotten a boost as well. It's much better at understanding the context of what I'm writing and providing guesses as to what I'm about to type. It's quite good. Most of the time. But it's also wrong in very subtle ways that require careful reading to notice. And I'll sum this paragraph up with the fact that I'm turning to an LLM more and more as a first search before I hit a search engine (yet I hate Google's AI search results).

Most of us do not remember the exact syntax for everything despite having coded in that language/framework for years.

I agree that it makes sense to use these currently but IMHO the ultimate programming will be human readable code-free, instead the AI will create a representation of the algorithm we need and will execute around it.

Like having an idea, for example if you need to program a robot vacuum cleaner you should be able to describe how you describe it to behave and the AI will create an algorithm(an idea, like "let's turn when bump into a wall then try again") and constantly tweak it and tend it. We wouldn't be directly looking at the code the AI wrote, instead we can test it and find edge cases that a machine maybe wouldn't predict(i.e. the cat sits on the robot and blocking the sensors).

I still feel more comfortable with the chat interface where I talk to LLM and make it generate the code that I end up putting together in a dumb editor because I'm still writing code for human based analysis that will be interpreted by machine and my claim is that if the code is actually to be written by machine for the consumption of a machine, then the human should be out of the code creation loop completely and fully assume the role of someone who demands stuff and knows when its done right and doesn't bother with the code itself.