Adventures in AI: Is it smart until it’s dumb?

Last week, a client and I embarked on putting artificial intelligence (AI), notably OpenAI’s ChatGPT (Generative Pre-Trained Transformer), through its paces.

Let’s preface all this by saying I’m not a computer programmer by any stretch. I dabbled in Turbo Pascal and Visual Basic in high school, which was over twenty years ago. I’m sure we’ve moved on since then.

The task was simple: write a series of 500-word articles about a given topic. Some had to have five FAQs (answers to frequently asked questions), and others had to have three. Sometimes they had to be two long FAQs (100 words or more) or three short (50 words or fewer.) Or four of equal length. They also needed to have links inserted at 100-word intervals, the keywords in the URL substantially or closely matching words on the page.

Seems like a piece of cake for a million CUDA core behemoth, right?

Well, sort of.

I “programmed” the GPT with articles I’ve written before so it would mimic my style of writing, as well as the information it needed to generate the copy. I also asked the GPT if it understood my rules, which it said it did.

The first few articles were written as per my instructions. Not exactly a thousand monkeys writing Shakespeare, but adequate. Satisfactory.

Then, it would start to produce stuff that I didn’t ask for. Instead of writing 500 words, it would give me 200 words instead. Sometimes it decided to omit the FAQs. If I prodded the GPT to try again using the rules, it would get three out of four right. Then two. Then I would have to stop and re-train it completely.

I wasn’t alone. The GPT was forgetting rules, hallucinating, and doing all sorts of weird things. That’s because a GPT is an automated intelligence. It isn’t something that can reason and intuit like a human can.

In Smart Until its Dumb, an eye-opening book by AI researcher Emmanuel Maggiori, posits that we’ve already experienced an AI boom during the late 1950s and through to the 60s; and an AI “winter” of disinterest in the tech during the late 70s through to the 1990s. Billions of dollars were poured into development and didn’t result in a significant return. Now we have Generative AI, Large Language Models, and other incredible automation tech at our disposal – and we’ve all heard about the epic fails it can conjure up. That’s because computers aren’t creating meaning from semantic signifiers; they’re breaking words down into letters down into mathematical tokens and manipulating them using probabilities based on pre-training. It “reads” thousands of documents to figure out “Is in India” follows a certain set of words, because thousands of sentences begin with “The Taj Mahal.”

AI could be likened to American philosopher John Searle’s Chinese room argument; that a Chinese person sends a question written in Mandarin into a sealed room; inside is a non-Mandarin speaker with a look-up table and computer that tells him what to write and send back to the Chinese speaker. This leads to the mistaken belief by those outside a Mandarin speaker must be inside the room. The man in the room has no idea what he’s reading or sending back; he’s just manipulating symbols and probabilities based on his training.

Although I’ll probably be roasted in the comments for “not understanding the AWESOME COSMIC POWER of AI” well enough – that may well be true. But I’m also of the belief that every technological innovation is a Faustian bargain, by way of media ecologist Neil Postman. “For every advantage a new technology offers, there is always a corresponding disadvantage. The disadvantage may exceed in importance the advantage, or the advantage may well be worth the cost.”

That said, I could pump out 100 articles in a day where it would’ve taken me several weeks (and my sanity.) It just might be that AI (either automated or artificial) isn’t the panacea for absolutely everything – at least not yet.

In my experience, this type of Generative AI is sort of like working with world’s smartest five-year-old: you kind of have to be vigilant of its every move, or it ends up playing with Lego and throwing cups of Milo across the room. Get your mop and bucket ready, is all I have to say.

Join the soloist movement. Whether you are new to Flying Solo or looking to grow your business, our membership options will help you attract more leads, grow your network and sharpen your business skills.  Sign up to our newsletter to get the latest news and advice straight to your inbox.

Now read this

//corestage-cdn.flyingsolo.com.au/working-smarter/time-management/godaddy-report-reveals-power-of-ai-to-save-soloists-time-and-money/