Average Intelligence

AI is an equalizer. Or I would say, averagizer. I know it’s probably not a real word. But it seems to be a word that fits what AI is.

I have seen a fresh-graduate using AI to make themselves look like a decent software engineer, and they might produce decent code as well if they know how to direct it.

At the same time, I’ve seen an expert using AI to go further and beyond, like Terence Tao who discussed with ChatGPT about the Jacobian conjecture. My math knowledge is very limited, so when I read the chat transcript, I don’t even understand what’s being talked and why it’s even a breakthrough. But at least it gives me a chance to shout at ChatGPT “WTH IS EVEN THIS” “WHY ITS LIKE THIS” “YES BUT WHY ITS IMPORTANT” or something like that.

After 30 mins or so on trying to understand. AI has bumped my math knowledge by 0.01%. Because now I know more about polynomials. Maybe if I shouted louder, I’d bumped it by 0.02%. If I do it 1000 times, maybe my math knowledge would gradient towards 1% understanding. Do it 10.000 times and maybe my error rate on trying to explain this conjecture to a stranger would now be 0.1%.

Yet, I’ve seen people that were capable of writing, capable of thinking - using AI to dumb down their output. Before AI, their output was superior. A highly readable code, a nicely designed design patterns - now their output is bunch of top level exports inside 10 scattered typescript files.

For people that have lower than average numbers, AI will help them move up and reach the averages. For people that are higher than the average, AI will drag them down to reach the averages.

For people that know what they’re doing, and use AI to expand their thinking capacity - like Terence Tao; will be able to move up above the average.

It’s just how an LLM is inherently. They are trained to guess the most probable token given the previous tokens. And what’s the most probable one?

The average.