Sorry, I know this point of view goes in the opposite direction from what Replit’s vision is clearly becoming.
But I simply do not agree that vibe coding is about letting the AI run blindly for 30, 60,… or upto 200 minutes.
Frankly, you might as well write a Word document outlining your entire app idea, and give it to a human developer, and wait for the completed app to be delivered back to you (admittedly, with just as many bugs as the Replit agent gives when it is in a higher autonomy mode).
200 mins “the AI will go from blank sheet to completed app” vision is utter nonsense.
I am starting to feel this is a hill I will do battle on. So throw your arguments at me as to why I am wrong. It’s a debate those in the AI dev world need to have, to ensure we don’t take this incredible new technology down the wrong road.
I think it’s both wrong and right. The response from me would be the stock answer you always get in tech, it depends…
There have been certain requests where it’s been amazing - give Agent very specific instructions, tell it to identify root causes and implement fixes based on best practices, then test to ensure it’s working properly (this cut my timeline for beta it feels by 1.5 months in just the past few days).
In other cases, ask for something easy like making sure design is responsive so you’re not ending up with a pixelated illegible logo and it runs off the rails making a completely new brand design.
I think your belief is both right and wrong, just depends on what you’re trying to accomplish, how much you’re willing to spend accomplishing it, and in your case probably weighing if you or the Agent will do a better job. I don’t think it’s too far off to say AI could go from blank sheet to completed app with the right documentation, some users have talked about doing just that. I’d always expect some cleanup after the initial build though.
Note: putting this message in the universe is definitely going to have something break this weekend and add 2 months
If there are certain patterns in how we work with the agent or how it behaves that are less than say 75% successful (insert your own figure) then I simply will not do them.
Most of us are not here to pay for the privilege of testing Replit’s system to help them improve it to the point where it can “go from blank sheet to completed app” 100% autonomously.
If they want me to do that then they can give me some Replit Inc shares.
So yes, whilst the occasional person pops up (usually on Youtube) to say “I rebuilt the Facebook system for $30 in a single 200 minute prompt”, the chances of that are 0.001%. Which is well below my safety net. So Max autonomy mode is off.
Testing mode: seems like 1 in a 100 people have had any luck with it. Switched off.
Once any tool becomes “real feedback shows that 80% of people find this tool amazing”, I will start using it.
And yeah, this debate is definitely going to get heard by the universe… expect breakages today
