How Smart is AI?
As I said in my last post, AI is a glorified search engine. In that way, it helps with research; it’s like instead of copying and pasting from Stack Overflow, AI does it for you. But as the saying goes: “If you copy and paste, you won’t learn anything. You might feel like you are, but you aren’t.”
Research-wise, instead of going through ten blogs in search of solutions, AI gives a crisp, single answer. But the problem is, apart from these solutions not being 100% right, I miss those ten opinions. Those might not be the best—sometimes not even related—but I would have learned something new: future problems, new understandings, weaknesses, strengths, and how the people who came up with them think. It improves one’s reasoning and thinking capabilities.
This time, I wanted to implement a simple feature flags system for the backend. And as you can guess, it was a bit of a struggle. For those interested in the technical details, I have embedded the chat below.
In short, I had to switch Claude to its Opus model with the effort set to “high.” Even then, I had to prompt it multiple times to get the right answer. Fundamentally, I wanted to chat with it about my decisions: do they make sense? Is there a better approach? Does my thought process hold up? Like a pair programmer. And once we settle on what to do, I have it do it for me. I wanted to use it for code generation, but prompting line-by-line doesn’t make sense. So, let’s align on what we’re going to do, and then let it do it. It’s not a major refactoring, but it’s not a trivial task either, even though it looks like one. It requires thinking—not much, but understanding is required, in my opinion. When I do things, I want to do them properly, with the understanding I have gained up to that point. It’s never 100%, but I try to reach as far as possible. And also, more importantly, I aim for the “bare minimum to get the job done.”
In conclusion: AI can generate code, but it cannot design. It can design those stupid “system designs” on whiteboards—those are just copy-pasting at this point. But when it comes to designing a feature that requires deep or even mid-level understanding, it fails. And to be honest, it somewhat hinders the process. Google AI’s search mode is much better in that context; it helps more with research, in my opinion.
So, is it worth it? To be honest, in cases like these, it is not that much help. But personally speaking, I like chatting with it. As a lonely person, I feel good talking to someone. But how smart is it? Not that much. Actually, how dumb is it? Well, enough to make you one…
I tried with Claude CLI, but it should be the same with others also.