The secret to keeping your sanity with AI

The secret to keeping your sanity with AI is the same as the secret to a happy marriage.

Low expectations.
(In the short-term at least)

My wife would not be happy if she expected me to be competent overall… in fact she’d be very disappointed. Constantly.

Instead, by having low expectations of me, she is thrilled when I perform well even in a very underwhelming task.

It’s all about perspective.

Same with AI (Large Language Models specifically). If you expect AI to be great, it will disappoint you. Hallucination. Confusion. Plain old stupidity. They are guaranteed (for me and AI).

But, if you have low expectations of AI, you’re likely to build into your thinking that this stuff will be wrong every so often.

That’s good, because you’ll build scaffolds and systems around it. You’ll support and monitor it a bit more closely when it matters. You’ll choose it for specific tasks in specific ways - it might write drafts, but never send without you checking. You will expect it to mess things up, and it will not be catastrophic when it does.

When it does do things well though, you’re pleasantly surprised. Like my wife when I actually finish putting some shelves up instead of just saying I will.

When you really want to ramp up your AI game, you can then blend those short-term low expectations with lofty long-term high expectations. Stupid today, but AGI soon.

AI researchers are like parents in that way. They can watch their AI models do some pretty stupid stuff today, but can still look at it, smile, and think “one day, when it grows up, that little scamp is going to be President of the United States”

Do you have low expectations of AI short term? Where do you think it will get to long term? Let me know.