The Transaction of Frustration

AI, Stochastic Parrots, and the $2-a-Minute Trap

Like many of you, I spend time each day staring at a blank screen, wondering what the hell I am going to write to crystallise the findings from an engagement or pull together a heap of disparate threads I can articulate out loud but struggle to write down.

Words and pictures in a specific order generate a meaning that people pay me for (when I distil it like that it makes me question everything).

I also write code from time to time and AI has become extremely useful in cranking my output to a level 12-year-old me sitting at my VIC-20 (POKE36878,15 anyone?) would think was witchcraft, wait is it?

For coding, AI is an enormous help because the parameters are narrow, I’m not after opinion or creativity I just want a 100-line SQL query written in 30 seconds that would take me 2 hours of googling and frustration to accomplish.

But for what you might call ‘general’ use AI is getting worse, and quickly. Why? Because while my VIC-20 only knew what I told it as I mastered and commanded 3.5KB, this thing has been raised on a digital diet of every piece of nonsense we ever committed to a byte.

You are shaking and staring into a kaleidoscope filled with your old receipts, Reddit threads and failed song writing attempts (Visual Conquest since you ask). Every time you hit Enter, the tube gets a rattle and produces what appears to be an original pattern, but closer examination reveals it to be a fragment of a GeoCities blog post from 2004.


The Digital Dumpster: Common Crawl

Ever heard of Common Crawl? Since 2008, this non-profit has been vacuuming up the web like Pac Man. It has everything, from professional white papers to your 3000-word short story based on Red Dwarf where Lister gets eaten by the Cat.

All of this is in the belly of this beast, your early professional thoughts and your weird hobbies are all part of junk food that AI gorges itself on. We did not build a digital Einstein we built a mirror made from a dumpster fire.

Now, for a second look at that previous sentence, AI could only say that by guessing that 'dumpster' and 'fire' often go together when we flesh puppets are annoyed, it can’t feel the heat of the fire.


Stochastic Parrots and the BERT (not Ernie) Switch

In 2021, a group of researchers who knew how the sausage was made wrote a paper called 'On the Dangers of Stochastic Parrots.' It was this paper, which I read 3 times and still only just basically understand, that pushed its authors out of the industry (literally).

I’m not really doing it justice but it’s good, you should read it.

The industry used to be about what they call Classification. It was Zoltar, you feed it a dollar, and the machine would tell you the answer. Not hard to guess my age given my pop culture references.

But the value moved to the Journey not the destination. The parrot does not want to get to the point, it’s not looking for your truth, its looking to resolve its own query.

In the paper I referenced above they use the term ‘Coherence in the Eye of the Beholder’ ‘Text generated by an LM is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind. It can’t have been, because the training data never included sharing thoughts with a listener, nor does the machine have the ability to do that.’ (1)


The $2-a-Minute Trap you pay to enter

Think back to the premium phone lines of the 90s, (no not that kind). You would pay $2 a minute to stay on the line while an operator read a script at the speed of a tectonic plate. This is the Transaction of Frustration.

When you use AI you think you are in control because you are telling it what to do, but in reality, you are paying with the only currency that matters, your time and your human brain.

Every time you argue with the machine or fix its broken logic, you are a volunteer lab rat doing the R&D for a multi-billion-dollar company. They keep the AI chatty and circular because a blunt, fast tool does not keep the meter running.

They want you on the line, this spin is a feature not a bug and it is all billable time you just can’t see it.


They decide when you leave

Conversely, watch what happens when your conversation gets long, then it feels like you are in the garbage compressor with Luke and the walls start to close in. Every word the AI generates costs electricity, processing power, cooling and ultimately will sink the Maldives.

As the chat grows, the cost of re-reading the shared history starts to spike.

The machine begins a Winding Up process, stops being imaginative and starts being a polite bouncer. The early enthusiasm to help you turns into a "Let’s wrap this up" summary.

It has finished vacuuming your brain and now it wants you to stop burning through its GPU power. The relationship is over because the cost calculation is tipping against them.


10 THINK 20 GOTO 10

The machine wants to keep you on the line as long it suits its purpose.

I use the parrot for the SQL queries, for more creative problems I just do what I did when 12-year-old me had a BASIC problem, go and watch ‘The A Team’ and come back to it later. Then, when the screen isn't blank anymore, I’ll know it’s because I had something to say.

RETURN.

(1) Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21).