LLMs reward expertise

> What an ode to exceptionalism

I've had people criticise me for having had the opportunity to learn in that way, as they did not.

> You also think that you are smarter than the people flooding Ceuta streets these days, don't you?

No, why would I think that? I don't know them, the only thing I can say is in their favour: moving country to better your situation is difficult and them getting as far as they did is a demonstration of putting in a lot of effort of the exact type I praise by default.

I suck at writing code, so strictly speaking: Everything I do is vibecoded.

The stuff that I produce in this way would probably be considered by many to be unusable trash. But it solves the problems I have, and it does so with exactly the amount of precision that I demand.

When I built a PWM fan controller for a pro audio amplifier, I was very particular about some aspects. I wanted maximum resolution from the DS1820B temperature sensors (which is a relatively slow mode where reads take ~750ms, and often the bot is primarily interested in fast), and resolutely-consistent PWM output (so software PWM was a non-starter).

It was very important to me that the fan speed ramp smoothly and without audibly-discernible steps, so the target output goes through a low-pass filter to smooth things out and the final PWM value gets recalculated at a completely-overkilled rate of 1KHz.

Power consumption was a very deliberate non-concern: The power used by the MCU is ~nothing compared to that of the whole of the system, so optimizing towards reducing it was never my goal.

At the end, it's a rewarding little project that is all wrapped into a state machine that burns clock cycles like they're free (they are free!), and it works very well.

There's parts of this thing that I do not understand at all, and that I have no desire to understand.

But if I hadn't been so particular about the parts I did care about, then: An underspecified one-shot prompt seems like it would probably have just produced a loop with a lazy 1-second sleep at the end, since being sleepy and power-efficient was a feature that the bot kept working to reintroduce.

I spent a lot of time working to dismantle the bot's proclivities to be this way, and I probably would not be happy with the end result if I had just let it do its thing.

Differently-stated: It could have been an unsupervised one-shot prompt, and the result almost certainly would have done the job of keeping the amplifier cool. (I just would not like it.)

This reminded me of Gaussian Processes. You start out with n-dimensional unconstrained (but strongly correlated) gaussians. As soon as constraints (data) are added (mathematically it's called conditioning), the thing goes more and more into shape.

Prompting feels a lot like this conditioning phase to me. You start with an LLM in unconstrained mode, basically just a "soup" of knowledge. If you prompt wisely, you immediately condition the LLM into "your space of (domain) knowledge".

What comes out is an extended version of your existing knowledge.

You can see the same thing with restaurants generating their own menus/pictures. Some of them look absolutely terrible, visually ugly, way too information dense, the classic piss filter, etc. idk how they do it, even the most basic prompt I can come up with makes something 10x better, and when I put in my amateur photography knowledge/keywords in it gets pretty close to what I'd consider a good pre LLM quality menu. Literally just adding "make it look nicer/cleaner" seems to get rid of most problems, but people just don't care apparently.

I ran a similar test and got completely different results. My girlfriend (hair stylist/artist) with zero coding background mentioned a Telegram bot idea. I asked "Why not build it yourself?" I gave her a Windows laptop, but she said she wanted what I have instead. So I handed her a USB stick and told her she was on her own now. Fast forward: she now runs Arch Linux with Hyprland (I use Xorg/i3 though), fully riced with cats. I only interfered on partitioning to preserve my data. She even installed Steam and got Portal running (that was her "watch this" flex after I told her not to even try). She pulled all of this off using a free-tier Gemini chat.

For the bot, a friend gave her a Kimi 2.7 key. She set up their harness and built a working bot in a matter of days. She even got a free Oracle VPS for deployment, though I stopped her there to check security first (still haven't had time, unfortunately). She uses that laptop daily now and says she enjoys it over Windows by a mile.