I like my AI with flaws - By Graeme Devine

Local tech icon, and computer game designer, Graeme Devine (Magic Leap, Apple, Microsoft, GRL Games, QXR Studios) shares his thoughts on AI via LinkedIn.

I like my AI with flaws.
In the 1980s I was fascinated with Eliza. For those who are too young to remember Eliza is a very simple pattern recognition response system that was an early conversation engine. It was really basic, if you typed in a sentence with the word “mother” in it, it would reply with “tell me more about your mother?” or something along those lines.

I wrote a version of it that if it didn’t see any words it recognized allowed the user to add to the database and give some responses. I worked in an office in London at the time and a few people there got quite addicted to adding to it’s repertoire of terrible responses.

If there was more than one word in there it could respond to it would choose one randomly unless the previous word was already in there (I think, I’m going on a 40 odd year old memory here). But it had a lot of bugs and the database writing to a hard drive sometimes took longer than the next response and the whole thing got messed up.

Except it got messed up and made a character that was like a goofy soothsayer. You’d say things like “I’m having a terrible day”, “terrible day” had lots of great responses, but often it would choose something like “you smell like fish and chips!” which is the last line in a contrived set of answers to a joke you could make it tell you.

I added Eliza to my 1987 game Metropolis and you converse with all the robots in the game that way to solve the crimes. I added many many weird references to Douglas Adams and as many long words from my newly acquired book of long and strange words and it would often quote them. It was beautifully flawed. So much so that many people just stayed on the first screen.

This brings me to my current experiences with ChatGPT-3 and ChatGPT-4. ChatGPT-4 is brilliant, amazing, incredible, intelligent, and more perfect than ChatGPT-3. Apparently it can score really highly on bar entrance exams (although I thought you only had to walk in the door .. ba boom doof). But it’s too perfect I think when it comes to being a personality.

ChatGPT-3 was much better at being your friend and it had flaws. Not just “I forgot what I said two responses ago”, they both do that, but genuine it can’t guess a riddle right because neither can most humans because we’re not that clever flaws.

And it makes me wonder. Where is the balance in this? It’s clear that a more perfect tool is required for say, writing code, but you need something different, more human, when it comes to an actual character.

That’s why ChatGPT-4 is great news for creatives. It’s so good it’s not that creative anymore, at least, not in the right way. At least not in the right productive way that matters to people like me. I think. Me and my buddy ChatGPT-3 agree 11/10 times.