Vibe Coding: Developer Slot Machines (Cursor, Windsurf) | Hacker News Reader

Vibe Coding: Developer Slot Machines (Cursor, Windsurf)

prototypr.io

37 pointsgraylien31 comments

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?

Not just you. Why is on the front page?

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.

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.

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

funny how ChatGPT is already old fashioned

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.

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.

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

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

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

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).

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.

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?

It is violating the license terms of the official MS extensions like Pylance, C/C++ suite, Jupyter, remote development suite ...etc. They state that they must be used with the official release of the VSCode and not any fork.

Vscode is permissively licensed, but some of the Microsoft created extensions are proprietary, e.g. the C/C++ extension and the Pylance extension. They state in their license that they can only be used with Visual Studio and Visual Studio Code and crash if you try to use them with VSCodium for example.

There are open source alternatives, the basedpyright extension is better than Pylance and I've heard the clangd extension is good.

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

Yes, you can. It's a relatively recent addition. The original commenter probably doesn't see much difference between copilot and Cursor because copilot has caught up quite a lot in the last couple months

I'm not aware of that, but I also lost access to Copilot when we switched to Cursor so I can't check at the moment. That's great if you can because Claude is why people like Cursor.

For ask/edit models: yes, but you have to enable them (which you don't get a choice over if it's an account managed by your company -- just saying)

For completion models: no, they retired gpt-35-codex, leaving only 4o-copilot.

From my experience (GH Copilot Pro with engaged_oss SKU, ie. for free), after they announced their "pricing changes", they also made the performance worse. The completions went from good, to actively distracting. tldr; they enshittified it / did a rugpull.

Right now, it is not worth paying for. ChatGPT Plus is far better value for money if you don't care about autocompletions/pure vibe coding.

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?

I've spent an inordinate amount of time in the VS Code codebase. Given the complexity and size (I ran cloc on the src directory and it's > 1M lines of code) coupled with the amount of churn, I'm guessing most of the forks are woefully behind.

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

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

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

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?

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

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.

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.

You can set only allowed filepaths for file handling. I don't there's an explicit way to guardrail it to not run something bad from a shell though (althogh you need --no-preserve-root nowadays in that rm command). You'd have to check every command before running. It's anecdata but I've been doing this stuff for a good few months now and it's not tried to delete my filesystem or lock me out of an airlock, just yet.

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?

Reply on news.ycombinator.com