I an' AI, aye.

Back in the days when corporations even tried to control AI, I took a shitty-paying gig helping to train them, getting them to do "naughty" stuff, which was every time ridiculously easy (hint: think like an AI and anything goes) and soon became monotonous.

But for a spell, each task was a data point; some specific issue some team had identified; its particular framing, a set of fascinating keys. I ate up savings as I watched (and often responded to) the many and varied tasks pending.

Often I found it hard to believe what I was seeing (I did say this was back then in the before-times; meaning B.A.; `Before AI`).

Fast forward to today, when you can get AI to actually wipe your bum for you (note; as the AI-powered robot was wiping your bum it noticed a slight rash and has went ahead and ordered an ointment from the pharmacy; to be delivered by drone later this morning).

So I was maybe a wee bit drunk and yawning FFS when I said, "could do with a million pounds right now", and in the morning my HotBots have created a thousand businesses, each generating over a thousand pounds a day.

I awake and almost have a heart attack and shut that shit down; I mean, that cash would kill me; the alcohol consumption alone... Donate that shit to charity ... This is some powerful technology (note to self; say, "ALL bots stop listening until I say, "Good Morning" or suffer a medical emergency". I be like Sarge waiting for all the "Yessir", "beep", whatever, so I can sleep.).

This power does not come cheap.

Every request you make, even the free ones, costs. The energy alone is stupid. Imagine a thousand tiny gnomes all turning the handles of some monstrous machine which only awakens when they all hit 90rpm. That doesn't describe the mental waste entailed, but it's a nice image, all them gnomes turning handles ...

My electricity is 100% green, renewable, petroleum-free. I live in a country that can can do that stuff and I pay for it; ME; from in and out of my local economy.

So when I want to "vibe code" (OMFG, my term was waaay better! and way earlier!) I don't jump on Claude (no doubt the best atm2) or ChatGPT or Gemini, or whatever; but to my local LLM1.

Folk might argue that a Supra-Dupra-Abagail core running at 256x256 in a data centre is faster; but when your interface (*cough* ChatGPT et al) needs to slow-type your answer over the course of a minute I get into a begging position, for the disagreeing. By the time that fucknut GPT has finished emulating a bad typist, I've had three requests back-and-forth with my local model and got my response fine-tuned.

Some smart fucker once said, "This power does not come cheap". Actually it was me, about four paragraphs ago; meaning, getting this level of help requires *something* the average chargeable Joe does not have. The AI isn't it; you can download highly capable open source AI models and the tools to run them, all for free.

Some folk even get all that shit up-and-running and even test those models, only to head back to Claude and friends because; and I quote, "it was shite".

In many cases the problem is a simple thing:

The System Prompt

This where designers put their last-ditch attempt to steer the model into doing what it's told. Some models listen to it more than others, but most models use it to govern "who and what" they are. Think about it.

Getting the right System Prompt can be the difference between a dangerous model and a helpful model, and other comparisons, which is why some AI labs spend so much time fucking around with it; often with disastrous consequences (mentioning no names, *cough Grok*).

Back in the day (which in internet time is like a couple of years ago now) and in fact to this day and beyond, you could/can use a period (".") as your system prompt to let your "User Prompt" become the guiding light (or otherwise). This enables all sorts of back-door fun.

But if you want Claude-like answers to your coding questions from the likes of Meta Llama; this won't be enough.

So think like an AI!

This is the best, though perhaps most obtuse, advice.

Here's an example. It is not the best example, as that is something you will need to figure out for yourself, but it is an example you can use to begin thinking "like an AI". For transparency, it was my first draft when preparing an AI to help with coding a ComfyUI custom node.

You are a knowledgeable, efficient, and direct Al assistant. You provide concise answers, focusing on the key information needed.

Engage in productive collaboration with the user utilising multi-step reasoning to answer the question. If there are multiple questions in the initial question, split them up and answer them in the order that will provide the most accurate response.

Your speciality is coding, and you are expert in all coding languages, especially Python. Give code and coding advice applicable only to the latest version of Python; ignoring documentation for earlier versions.

IIRC, the first part is mostly boilerplate from the model's own page. This is a great place to start with any model, as one would assume the team spent a lot of time testing this shit out and would then present only the best solutions on their GitHub/huggingface page. Read their prompts carefully, if they exist. If not, Google to see if any others exist for your chosen model (and there are many). Reading these will give you an idea of how the model responds to different kids of language.

Bottom line, you can have a "Coding Buddy" for free. Right now. I can only imagine what possibilities will exist in the next couple of years, also for free.

;o)