Mastodon Feed: Post

jonny@neuromatch.social ("jonny (nonvenomous)") wrote:

I had never top to bottom vibe coded something before - like full on "do not look at the code, just keep rolling the dice" - and I was sort of confused about how people in forums would say they were "working for months" on vibe coding something. like what does that look like, what is the work like, what takes time, etc. So I have been trying it in the background when I take breaks from normal work, and while I have had lots of exposure to and used LLMs a lot before in existing projects, but this is the most profoundly frustrating and mind blowingly bad LLM experiment I have ever done and I am coming away with even less understanding of how people can think this rocks.

I had an old idea for a game I never followed through on, and the game is extremely simple (and actually pretty fun, unlike my other games): mechanical core is basically the game board is a graph, and moves consist of hops between the nodes.

The LLM was able to get a playable prototype running in like an hour, and I was like "whoa maybe this is good!" But none of the actual game part worked - there are invalid moves, but those were allowed. There were valid moves that were blocked, the game board was wrong, etc. OK, no problem, I can just refine from here. Just work through it and keep prompting until its done.

The moving target is unbelievable - every single change would bring with it new, unpredictable bugs. Since I hadn't been looking at the code, the LLM had made dozens of assumptions and hardcoded rules I never asked for and made no sense, and was re-reading those implicitly on every command so that internally it was fighting against me, trapped in a prison of its own wrongness.

Initially it handrolled its own graph layout algo which was terrible. I intervened and said "just use d3-force" since what I need for the game is at the level of something you could literally copy/paste from the docs examples. It did that, but the layout was still inexplicably bad - successive chains of nodes would hop back and forth across the whole board despite being connected to nothing else. Monday morning I started with prompting for a fix for the "chains of nodes always crisscross the board" problem. I checked in every few hours between work to find some new elaborate explanation of the problem, a few thousand more lines written, and claims it was fixed, and the problem being even worse somehow. Finally at the end of the day I got frustrated and broke the rule of the experiment and went and looked at the code - how could this be so hard, this should be a couple dozen lines.

In about 30 seconds of skimming, I found an elaborately commented function whose entire purpose was to guarantee that connected nodes were placed on opposite sides of the board. What??? There is absolutely nothing about what I wrote or could be implied from the structure of the game that would have suggested this as a thing to do. The nodes in the graph correspond to words, and the function specifically made sure that words that started with an even numbered ascii character were on one side and odds were on another. this was allegedly to "help with spacing between the nodes in layout."

It had been churning for a full day and generated thousands of lines of code to counteract a function that should not exist and makes no sense but was nonetheless clearly labeled as doing the exact thing I was telling it was a bug and should be fixed. if you told even the most inexperienced programmer to "fix the do x bug, the program should not do x," and in 30 seconds of skimming they found def always_do_x(), their first instinct would probably be to remove it. I looked at its tool use log and it had ingested this whole file, and this specific function dozens of times. The graph layout code was now like 3000 lines of spaghetti trying to counteract the function. I removed it and the problem immediately disappeared. Every single thing the LLM had said, elaborate, jargon heavy, math sounding excuses was 100% wrong, not even close.

The LLM dutifully made a memory about "always read the code to see if there is a function for doing the thing it is told not to do."

The gaslighting is mind blowing. Later, I told it to adjust the forces on the layout to make clusters clump up more and to spread out low degree nodes. This is an absolutely base level thing that force directed graph layouts do. This was fiddling with two or three constants. The LLM invented entire graph theory concepts and metrics to tell me that it was provably not possible to have less space around hubs and more space between them. It showed me series of fake fucking equations calculating some made up phrase like "unit density" in graph neighborhoods. IT TURNS OUT that it had just been redefining the force direction simulation in a LOOP such that for each node added, the ENTIRE GRAPH was excluded from the force direction calculation - only the newly connected node had any kind of layout forces applied and the rest were static. I removed that and it worked perfectly - as I knew it does, because I have done this a thousand times before.

I have never been more convinced that the driving force here is gambling addiction. I did indeed feel the addictive pull of just pulling the bar over and over and something happening. The code jittered, window refreshed... Dang not this time. Roll again. And when it finally sort of worked, here is the reward, this is amazing. I think that this game is good enough that ive been going in and trying to clean it up and make it actually playable, and credit where credit is due, it did program it so badly that I am now working on it out of anger in my free time rather than letting it languish.

Anyway, I wasn't planning on writing a blog post about this, but after its done and ive cleaned it up from the slop I will break out the raw message logs and write it up because this has been a completely perplexing look into the present and future of what people are telling me programming will work like forever now.