I had high hopes for Caffeine from the beginning and I thought that a single prompt would actually be enough to generate a complete app and get it ready to use, but in reality what I’m getting is very far from that idea. Even when I give it something simple, it still only builds it halfway and never finishes it properly, and what bothers me the most is that when something is wrong it completely fails to understand where things went off, so I have to sit there and manually point out what needs to be fixed.
Another thing that is clearly noticeable is that it doesn’t really see the project as a whole. It feels like it breaks everything into fragments, and if you fix something in one place, something else breaks somewhere else, which gives the impression that it has neither logical understanding nor a stable memory of what it did in previous steps.
There is also this feeling that it is being overly careful with resources, as if it is trying to avoid consuming too many cycles, and because of that it doesn’t allow the AI to think deeply or repeatedly verify itself, which ultimately results in a product that works only on a superficial level.
Overall, the idea is very strong, but at the moment this is not a tool that a developer would rely on for serious projects. It feels more like something for generating early prototypes rather than a real app builder that can operate independently.
no its many good
u need buy credit more
I recommend to let ChatGPT make prompt, best is to tell it to make app that will be built with 3 prompts and make it in logical order. V3 might give prompt for you like GPT, V2 didnt do it well.
Sometimes we give short prompt and it does not fully understand as It has to modify many files to complete it.
V2 had lot of manually pointing out, I did it on V3 too, but it failed to fix as problem is caused other stuff, we point to wrong stuff. V3 can find problems themselves now, just tell it to look code and find issues.
Same problem on Bomberman, it want to use same code on different features that has to be independent, but it want to use same flow.
For serious projects, yes manual building is needed.
Im still happy it exists. I think people where hoping its new golden star that will bring ICP token to 100+ usd, but it wont. Only way price to go that high is if there is constant buyers who collect lot of tokens and dont sell at low prices. When there is demand, there will be price movement upside as demand dont go away.
Basically AI is tool, we need to learn to use it.
If people can remember first time they played NES games with joystick or first time used keyboard. We had to learn to use them so we become faster.
To get max out of AI, i think we have to use it a lot to understand little more about coding, get feeling what prompts it likes and learn new tricks what makes AI better tool for us.
That’s very insightful Henn91. I feel brighter already.
Actually I think that way ICP explodes in value is institutional adoption through storing important data at first, then slowly transferring the entire tech stack over using cloud engines, solely to protect from hackers and outages. People have to buy ICP to burn it which is even better than just holding it.