Vibe Coding: Developer Slot Machines (Cursor, Windsurf)

Vibe Coding: Developer Slot Machines (Cursor, Windsurf)(prototypr.io)

37 points by graylien 1 year ago | 31 comments

bilekas 1 year ago |

This article is really difficult to read or its just me? The breaks with the adds are really distracting and I don't know what the author actually is saying here, that he will try windsurf? Do I need to watch the video?

stavros 1 year ago | |

I had the same thought, basically here's a summary, as far as I can tell:

> AIs are sometimes good for code, sometimes not so good. I put some money in but don't know what code will come out, and it's so easy to put more money in. My friend told me to use Windsurf instead of Cursor, and it's easy to switch IDEs, they have no moat. Where's the moat?

AFAICT, that's the whole article.

graylien 1 year ago | |

Hey, to clarify it's about the user experience of IDEs - how it feels to jump from one to the other, how the UI prompts you to add more credits (so you can end up spending a lot if you're not careful with prompts).

This has often drawn similarities to a slot machine, the way it is not deterministic what code will come out, and what changes to which files will be made.

ratatoskrt 1 year ago | |

Not just you. Why is on the front page?

softwaredoug 1 year ago |

Maybe I'm a luddite, but I like having ChatGPT generate my code the old fashioned way...

hu3 1 year ago | |

This is a way to bypass alleged hidden prompt trimming by IDEs.

Some people say AI IDEs might summarize or even trim prompt to reduce token input.

graylien 1 year ago | |

funny how ChatGPT is already old fashioned

rgoulter 1 year ago |

Since LLMs are sometimes wonderfully useful, and sometimes not, I'd suggest effective use involves figuring out in which cases it's likely to succeed, which it's likely to fail.

For example, the mentioned graph has "initial prompt with iterative tweaks", followed by iterations of 'starting from scratch'. -- I don't understand why you'd think "this is an ineffective way of doing things", and then keep doing it.

Describing LLMs as "slot machines" seems like the author has no curiosity about the shape of what LLMs can/can't do.

okamiueru 1 year ago | |

Answer is useful as a suggestion, and doesn't need to be factually correct: Good

Answer is useful as is, and needS to be factually correct: Bad

nonethewiser 1 year ago | |

I do think its a bit slot machine like because the output isnt deterministic.

graylien 1 year ago | | |

and they prompt you to add more credits, it's seamless to spend more - slot machine is referring to the experience using it, not what is happening technically

Arn_Thor 1 year ago |

I was coding with VSCode for a while and just this week decided to try Cursor which everyone have been raving about. Was quite disappointed to find no meaningful difference with VSCode (so far, after 30 minutes of working with it).

jmisavage 1 year ago | |

It's a fork of VS Code that still uses the official extensions, despite that violating the license terms. Functionally, it's pretty similar to using VS Code with Copilot, but with the added bonus of choosing from multiple models. I’ve only just started using it at work, so my experience is also pretty limited, but so far the results seem slightly better.

graylien 1 year ago | | |

Do these forks keep updating with the new versions of VS Code? I wonder if once they're forked, they are stuck with that version they forked from?

On the other hand, I think MS are pretty much cloning the good stuff from Cursor, like agent mode: https://www.youtube.com/watch?v=dutyOc_cAEU

Leaves me thinking VS Code might be the best in the long long run, but for now they're all kinda similar?

bitpush 1 year ago | | |

How is it violating license terms? As far I know, vscode is released under a permissive open source license.

Isn't the situation similar to Brave et al built on top of Chromium but supports Chrome extensions?

almostdeadguy 1 year ago | | |

Can you not choose from multiple models in Copilot? My understanding is you can.

Arn_Thor 1 year ago | | |

I’ve been juggling different models with Copilot in VSCode already

graylien 1 year ago | |

Were you using Copilot? I haven't tried VS Code for a while, I guess I can have that one open too. You can even have Windsurf inside VS Code with the plugin I think..

Maybe Cursor/Windsurf could've just been plugins? Only Zed.ai seems really different of the popular IDEs

Arn_Thor 1 year ago | | |

I was using Copilot within VSCode and it seemed very well integrated, making changes in files for me and letting me choose between AI models

hu3 1 year ago |

Speaking about agentic coding I have a question about Copilot in VSCode.

In agent mode (need to enable it manually in settings), did anyone test forcing Copilot to run unit tests after code changes and fix code to pass tests if they break?

_joel 1 year ago | |

That's what I do as standard in Windsurf. I set memories for it to run the tests too. That's standard, right?

hu3 1 year ago | | |

I just tested and yes, it works.

My issue was my instructions.md file telling it to think too much before writting files and running tests. So it was in a rabbit hole of ethernal thinking.

Now I can tell it to create crud pages and it will generate and run tests for those pages as well.

hu3 1 year ago | | |

How do you solve the risk of AI writting and executing unit tests with something like `rm -rf /` and wiping your files?

I'm looking for solutions. Docker looks like one but I like to keep things simple.

zerosleep 1 year ago |

I find VS Code + Cline extension and it's memory-bank pattern far easier to control while still feeling like I'm "vibing". Anyone else use that setup or is it already antiquated?