Working on organizing the story for the next stage. Planning to emphasize the story aspect more in Stage 2.
Since I’m working alone, I find myself talking to AI a lot while developing the story. AI doesn’t create stories or come up with ideas for me, but just the act of discussing with AI helps me draw out ideas and thoughts from within myself.
I’ve been wandering between note apps for nearly 20 years. I think I went through Evernote → OneNote/others → Joplin → Evernote → Joplin. I don’t need many features, but they were all somewhat lacking. I wanted to build my own, but had zero web app experience and knew nothing about server configurations. Recently, I built a memo app using Gemini and I’m extremely happy and satisfied with it. Even though it was imperfect, Gemini created a migration app from Joplin. The best part is being able to add or modify features immediately when needed. It really made me think AI has genuinely improved.
Still, I recently canceled GPT-5 because it would give random responses without actually looking at the images I uploaded. I was trying to get design feedback on images, but something felt off, so I checked and found it was responding without viewing the images. The explanation was lengthy but ultimately didn’t reveal the real reason. Maybe it was for efficiency. After this happened several times, I tried Gemini again and found it had improved significantly and was even free, so I canceled GPT-5 and switched to Gemini. However, Claude still seems better for translation and writing reviews, so I continue using Claude.
This made me think that current AI approaches might be hitting their limits. AI works well as a table that converts A to B, but for tasks requiring contextual understanding and creating genuinely new outputs, it still seems unusable (and maybe always will be). This might be because I haven’t tried more expensive models.
Plus, it’s too slow, restrictive, and unreliable. Even with paid models, you hit limits quickly when using advanced features. Every model still has hallucinations, and the problem is you have to check everything to identify what’s hallucinated. In other words, you can’t trust it.
While dramatic miniaturization like ENIAC’s billions of times more computation in a phone-sized device hundreds of thousands of times smaller is unlikely to happen again (though quantum computers might be different, idk), unless there’s a paradigm shift allowing current AI to run real-time on smartwatch-sized hardware, or a change from understanding the world through text to understanding actual 3D space and physical laws (which might have already started, idk), I wonder if there’s much point in continuing the current approach of just throwing massive amounts of electricity and hardware at the problem.
I had hoped that in about 10 years, I could use AI to create games at a rate of one per month, whereas making all the games I want would be impossible in my lifetime using past or current methods. But that expectation is starting to fade.
Anyway, creating stories is still the most enjoyable and thrilling part.