AI – the Toddler that’s Getting Dumber Every the Day.

(as at July 2023. Since then I’ve become more impressed, at least with Dall.E-3)

AI – artificial intelligence – may have a bright and/or scary future when it grows up. With apologies to the proud parents, though, the particular AI that I’ve played with over the last few weeks (June-July 2023) shows no signs of being gifted. And unlike many other toddlers, it appears to be getting dumber as it ages.

That’s a curious progression and I’ll give my thoughts on why in a moment. But let’s take this one step at a time.

The AI toddler I’m chiefly talking about is Microsoft’s Bing Chat. It’s free with Microsoft’s browser, which is why I gave it a trial. It’s similar in some ways to its brother, Chat GPT.

One noticeable difference is that Bing Chat can be referred to without offense in French as “Bing Chat”. “Chat GPT” on the other hand is pronounced in French as “Chat j’ai pété” which means “Cat, I farted.” Discussions of the future uses, threats to human existence, and legal issues arising from “Cat, I farted” are routinely carried in the French media with mostly straight faces, but still, you see the advantages of “Cat Bing.”

Cat Bing (or to be Anglo, “Bing Chat”) prides itself on taking questions in natural sounding English (or maybe French, or whatever) and replying in natural conversational language. This epic computing leap (OK, that was sarcasm) obscures the fact that Bing Chat, while it speaks fluently, spews junk. In this regard it resembles most politicians. We should kill it now, trample it with elephants, strew salt on its ashes, and put a stake through its heart.

Here are some of my prime interactions with this little monster.

1. When the car-transport ship, Fremantle Highway, caught fire off the coast of the Netherlands in July 2023, I wanted to know where it had loaded its cargo.

My conversation with Bing Chat went like this:

Pretty embarrassing response considering that a search on Google revealed the answer IMMEDIATELY:

2. My related question, I admit, was ambiguous. I hadn’t intended it that way, but the ambiguity was revealing. Most adults (politicians excepted) would have asked for clarification before answering. The toddler AI, Bing Chat, didn’t.

R/C means radio-controlled. So miniature zoom-zoom cars for kids and hobbyists holding little joysticks.

This goes to the heart of my fear that it’s not AI – artificial intelligence – that will kill us all, but that AS – artificial stupidity – will. [Or maybe human stupidity or cupidity.]

3. By now, I was intrigued enough to try a trick question on this child-prodigy-of-dumbness/laziness. Here’s what happened when I threw Bing Chat a typo. If I did the equivalent of this in conversation, you’d assume I’d misspoken or you’d misheard and you’d ask me to repeat the crucial word. Maybe you’d even say “Did you mean Beethoven?” Instead, the toddler AI threw a tantrum worthy of any two-year-old (or politico dealing with an overly astute journalist):

4. I gave the little monster a topic for a rhyming poem. In general AIs do this type of task quite well – try it for yourself. In this case, though, Bing Chat plagiarized its opening straight from a poem by poet Strickland Gillilan. Clearly, in addition to other AI problems, some parental guidance on ethics is lacking.

Etc. etc. I won’t show the whole poem because Shakespeare it ain’t. And it’s image generation wasn’t Van Gogh either:

5. When I asked it a simple question about arithmetic, Bing Chat got the theory right but the answer wrong. (Shades of singer/songwriter/satirist Tom Lehrer’s lovely song “New Math”). I would love to show Bing Chat’s silly answer, but today for a change, as I write all this, it’s getting both the answer and theory right. Tomorrow? Who knows. With Bing Chat, as with some politicians, the answer to 2+2 depends on the day of the week.

---

This leads me to the progressive dumbing down of AI, and the subject of toddlers in bars.

Several writers plus a recent piece of research out of Stanford ⁽¹⁾ have tracked the increasing dumbness of certain AIs.  While this dumbing down may come as a surprise to many, I’m NOT surprised.

The way I understand it, AIs educate themselves by reading prodigiously everything they can find in digital form. That means they’re absorbing wisdom from the internet, which, as we know is 88.2% garbage. (OK. I made up the figure 88.2%. But once I post this, you’ll be able to find the same figure on the internet, which gives authority to my number and justifies my statement, and likely some AI will soon absorb it as gospel.) Anyway, letting AIs educate themselves from the internet is like sending your toddler to gather wisdom from drunks in the local bar instead of sending him/her to school. Sure, he/she may overhear some valid opinions there, but eh, he/she will end up with a skewed view of the world. And surprise, surprise: the longer he/she spends in this bar of boozers, loudmouths, and extremists, the dumber his/her views of the world will become.

***

Footnotes:

⁽¹⁾ https://www.dexerto.com/tech/chatgpt-is-actually-getting-dumber-2217094/