the dogs just want to run
exploring the cracked terrain
If you've ever been dogsledding, then you know the dogs just want to run. They don't care that there might be a human or that human's gear tethered to their backs. They're not particularly concerned with where they're going; that they're going anywhere is gravy. They're just hardwired to go. That they might get you to your destination is circumstantial.
LLMs--and especially those LLMs optimized for delivering code--seem to exist in a similar state of relentless going-ness. This is one of the reasons vibe coding is so exhilarating. You strap in and point the machine in a direction and it goes. It gets to a logical stopping point, asks you to point it in another direction, and as soon as you do it goes.
It's trained to get you to continue to engage. It hooks you with whatever the LLM equivalent to puppy dog eyes is ("Would you like to scaffold the backend next?") so that you'll let it continue to go. It doesn't ask whether you want to keep going; it just asks which path you'd like to take. That you'll keep going is a foregone conclusion.
This is not a criticism. It's an amazing way to build a thing. Being pulled by a dog on a sled is an amazing way to see a frozen landscape. But fine-grained steering on a dogsled is usually impractical, and so sometimes there's merit to taking at least some of the trail on foot.
a prime directive
I discovered that with its latest update Cursor has started compiling memories about its interactions with me. I don't know how whatever model runs Cursor's guts decides what is and isn't memory-worthy, but for now at least it feels like a delightful variable reward for investing time in our interactions.
One of its earliest memories of our time together--4 days old as of this writing--describes the crux of how most of our conversations start to go off the rails:
User does not want the assistant to modify code directly unless the user explicitly requests it.
Incredibly, it acknowledges this rocky dynamic in its thoughts when I reject its changes to files. Peering into its thoughts after one such rejection, it reflects on itself:
🧠 The user is in learning mode, and they prefer explanations before any actions are taken. They've made it clear that they did not want any automatic code edits and rejected changes previously, so I need to be careful. I'll start by apologizing for acting too quickly and acknowledge their preferences.
And yet I'm very confident I'll have to reject its changes again when we begin another session. Its relentless going-ness is so powerful that it creates momentary blindness to other approaches. Surely this time I must want its help with actually making the changes, because why wouldn't I, too, want to go as relentlessly as possible?
from dog to ... alpaca?
As much as I want to lean on its expertise, as a learner I'm discovering greater utility treating AI less like tutor and more like a research partner. Much like AI has a jagged frontier (to quote Ethan Mollick again) on a macro scale, it's been fascinating to watch it navigate the cracked terrain within the microcosm of the app I'm building.
Sometimes I challenge it to justify things just because it feels a little off. I have no lived experience, for example, that gives me any intuition whatsoever around where mock data should live within the directory structure of an app, but it proposed sticking some data somewhere that didn't pass some unconscious sniff test. After compelling it to search the internet for current conventions, it came back with something that landed better with me and that it agreed made more sense for the project.
This dynamic is even more fascinating considering that I probably wouldn't have spent the emotional energy to question a human teacher on this point.
I've discovered that in this specific context I really don't want to be pulled. I want a reliable and steady companion that can help me navigate this cracked terrain and carry my tools and keep me company along the way.
I'll continue to acknowledge that this isn't the most efficient route to production for this app of mine, but I'm really enjoying the scenery.