Dhruvik Parikh

AI Flow States

I've noticed that whenever I need to write something new these days, getting the first few words on the page is much more painful than it used to feel. I worry that I am starting to develop a reliance on AI for activation energy. Whereas pre-ChatGPT, you had to initiate, edit, and polish anything you're making all by yourself, now for many things you can provide very little guidance to an AI tool and it will do an excellent job at initiating it for you. You still have to edit and polish fairly carefully, for now, but for tasks like coding a new project, getting the v0 boilerplate down is so easy if you can write a decent prompt. On the other hand, there is no great AI tool I've found that helps with writing. I have tried to use ChatGPT to help me ideate how to frame an inkling, but I've found that once I see it written that way, it's very difficult for me to write it very differently from that. Even if I don't love the suggestion, the AI manages to anchor me. For that reason, I've deferred ChatGPT usage for writing to the very end, when I want to polish something and check for typos. This also helps me preserve my voice and message, I think. For obvious reasons, I won't be sending this one to ChatGPT because I don't want it to feel too perceived.

Although having to initiate all by myself is a little annoying, I've found that once I've gotten started, the feeling of completing a thought and finishing my writing is extremely satisfying. It's a level of flow state that I don't get into that much these days. I think there's something about the ideas starting and finishing within my own mind that makes the whole process very fluid and enjoyable.

Relatedly, I've found myself getting pretty bored while vibe coding. The other day, I did some Leetcode and took an AMC exam in the middle of the day just to feel mentally stimulated (10/10, would recommend). I've talked about this with a few friends who feel similarly. Vibe coding and related AI paradigms are simply not that interesting and intellectually challenging. If you were previously used to thinking very deeply and solving problems every day, you will find yourself very bored very quickly if you start using AI tools to do your job. I hope that this is just a function of the current state we're in with AI, where it's good enough to try everything but not good enough to do exactly what you want, rather than a pattern intrinsic to human/AI interfaces. I'm not advocating for constraining AI to copilot-like usecases, but I do think that there are way better ways for creative and intelligent humans to use AI than the current vibes-based agentic flows.

Part of this relies on the underlying technology improving. From a complexity perspective, it is much easier to propose a solution when the scorer has a vague rubric versus when the scorer has a detailed rubric that is unknown to you. This explains why AI does an astonishingly impressive job at tasks like "generate a Studio Ghibli version of this picture" or "code a watermelon merge game in pygame". After all, there could be thousands of different pictures it could produce that you'd be happy with and thousands of variations of a watermelon merge game that would impress you. But if you know exactly what you want? With pixel-perfect accuracy or very specific functionality? There is no chance AI can do it today. GPT-4o does an embarrassingly bad job at editing photos, it warps the existing faces in the photo and forgets details. Despite all the work on AI for coding, even the best models and IDEs still hallucinate changes to random files along their way towards completing an instruction. We'll need AI to fundamentally improve before we can expect it to consistently solve these harder problems, which are the ones that we need it to solve to use for stimulating work.

Beyond that, I think there is a lot of creative work that is yet to happen on the interface layer for AI. Chat is still the primary interface and the next best thing is... voice? Surely we can do better than that. Where's my megamind dashboard with 10 windows open each running a different task that I can see perfectly visualized summaries of and intervene naturally with keyboard shortcuts and context-aware instructions and generated input fields? I don't want a copilot, I want to fly solo in an airplane cockpit that has all these bells, whistles, and knobs for me to control and express.