crunchysteve

I read "Where's Your Ed At," a newsletter about the interface between the tech and finance sectors, written and extensively researched by Ed Zitron, a finance journalist by trade. Like most of you out there, I've played a bit with AI and, probably like you, I've been greatly unimpressed with the results, not to mention the carbon emissions of the data centres needed to run Large Language Models that do the heavy lifting Ollama, GPTChat and others need.

I've used OpenAI, MidJourney and Ollama - online where I could access free trials, offline on my laptop - and I have never once been impressed with the results. This "tool" doesn't simply hallucinate, it positively trips balls. It makes Archer on a bender look sober! AI simply does not work reliably enough to use for mission critical stuff. Never mind the copyright theft involved in training it, the social prejudice biases in it because the AI sector is mostly run by rich, white snake oil salesmen - and they are mostly men.

If you want an example of AI hallucination to fully understand the term, here's my former blog masthead, "Buddha on a recumbent bicycle."

The image actually works for the purpose I wanted it for, the inaccuracies of the design are whimsical and esoteric, the bike appears to grow from Bodhisattva, himself. You'd hardly call it a useful image. Ponder, though - what if you needed a solution to a practical problem, say working program code for a mission critical application?

I was using Copilot on github late in its beta period, before Microsoft rolled it out to 365. Now, I'm an Arduino hobbyist, and not a very good programmer, but here is how I'd get an Arduino board to count to 10 by flashing its onboard LED (leaving out the setup code, just the counting and flashing)...

void setup(){
  pinMode(13,OUTPUT);          //  setup the builtin LED as an output
}

void loop(){
  for(int i=1;i<=10;i++){      //  Count from 1 to 10 using "i"
    for(int n=1;i<=i;n++){     //  Count from 1 to the value of to flash the LED
      digitalWrite(13,HIGH);   //  Turn the LED on
      delay(250);              //  Wait a quarter second
      digitalWrite(13,LOW);    //  Turn the LED on
      delay(250);              //  Wait a quarter second
    }
    delay(2000);               //  wait 2 seconds
  }
  delay(2500);                 //  wait 2 and half seconds
}

This code will flash the LED, once, twice then 3 times until it flashes 10 times, with 2 seconds between each set of flashes, then it'll wait 2 and a half seconds before starting from the beginning. It's an extension of Arduino's "Blink" tutorial and the basic "Blink" is where all Arduino benners begin learning the C language and how to make microcontrollers do stuff.

When I tried using Github Copilot to write code fragments for a project of mine (a kind of simple band-in-a-box guitar pedal for live performance), an algorithm that have really well worked out, it works in circuit simulations, I just haven't got it working reliably on real hardware, yet, because simulations to hardware is HARD.

I thought I might be able to refine the code - fewer lines can mean faster code, easier debugging, that kind of thing. I started by checking Copilot on the above advanced/extended version of "Blink." I used the prompt, "Make the Arduino count from one to 10 by flashing its LED on pin 13 using nested for() loops then repeat. Use 250ms delays between each flash on and off and 1s between each loop iteration" Copilot invented the commands "flash(n=1 to 10)" and "blink()", never created functions for those commands to call (they're not Arduino commands!), never use a "for loop" and never setup any of the variables in the global space, never initiated the output pin in the setup function. These missing things are mandatory things in the Arduino environment. I was using Microsoft's VS Code and the PlatformIO microcontroller plugin, invoking Copilot from via these tools, exactly per Microsoft's documentation.

Considering this was at the height of AI hype, probably 18 months ago as I write this, Copilot was supossedly pretty "refined, tried and tested" on C basics. Using in code completion mode wasn't much better and the completions for my comment lines (the "//" prefixed bits) bore no resemblance to the code lines they were supposed to be describing. I turned it off. I'm a shit coder, but a good technical communicator - was a workplace trainer and audio producer in radio, I can write an understandable prompt. It never once gave me more than an illusion of help, and even that was early, while I was learning how to "speak" to it. Check this screenshot of Reddit in a google search "does AI even work."

Now, my own experience aside, and nothing more needs to be said about the burgeoning carbon cost of the massive growth in data centres to run "Eliza on steroids," but AI is the 21st Century's Ponzi scheme. It's a bubble economy. Ed argues this the post I link at the top of this long post, and eloquently, linked, footnoted, the works. It's stuff from the finance and tech sectors, not from the anti-tech community. It's a thesis, it's the length of a non-fiction expose novel! Well, almost. It's numbers. OpenAI is running at a loss, they all are if you only measure their AI parts. It's economically unsustainable. And it doesn't work. It's not intelligent, it's a hyper complex version of throwing a dart at a board and saying, "greater than 10 equals yes, below 10 equals no." And because the big players are so tied into the finance sector, almost literally burning money with no profit in sight, there's going to be a crash that'll make 2008 look like dropping $20 on a losing lotto ticket.

My advice to anybody willing to listen is to buy a BIG hard drive, an SSD at least as big as your cloud provider's plan, if you value the data you keep. Move that data off the cloud onto the SSD. Regularly synchronise it. Move any investments or annuities you own into old world money sectors. Because, when the big boys fall, and 2025 is only a few months away, they will take a LOT of 21st Century infrastructure with them! Dreamwidth may be OK, but who backs DW's web services? My blogger, shinyhappyrainbows, backing that up is my project this week. I'm freezing my M1 Mac at Sonoma - yes, Apple have been suckered into OpenAI's heist of the century and have deeply embedded that house of cards into the very guts of their OS. Microsoft are seriously exposed to OpenAI's vulnerability. Facebook's manipulative advertisng business model is insulating them, but they're using many of the same data centers as all the other big players. Amazon are exposed. Heavily.

Pass the popcorn. It's about to get ugly. I hope my super isn't invested in anything close to tech, but the fireworks will be "fun," all the same.