Day four, the hot takes arrive and the founders reply

By Saturday, December 3, 2022, ChatGPT was three days old and the internet had split into two camps. Sam Altman described them in a tweet that day: the takes were either "this is AGI," which he called obviously not close, or that this approach couldn't go much further. His own line was "trust the exponential."

Elon Musk replied underneath, per the archived thread:

ChatGPT is scary good. We are not far from dangerously strong AI.

Altman answered him in the same thread. He agreed about being close to dangerously strong AI in one specific sense, an AI that poses a huge cybersecurity risk, and said he thought real AGI could come in the next decade, so the risk had to be taken extremely seriously.

That's the whole exchange. Three tweets, archived on a page you can still open.

Why the takes were so loud

Both camps were reacting to the same thing: a chatbot that sounded sure of itself on every topic. If you tried it on something it happened to be good at, it looked like magic. If you tried it on something it fumbled, it looked like a parlor trick. Most people tried five things and wrote a verdict.

I don't think either camp was being dumb. They were being unspecific. "Is it smart?" isn't a question anyone can test. "Can it draft the weekly report I write every Friday, well enough that I'd only edit it lightly?" is.

Make it give you a verdict

The same vagueness shows up inside the chat. Ask ChatGPT a broad question and you'll often get a balanced little essay that never lands anywhere. The quickest fix I know is to ask for the format you want back, before it starts. A table with a verdict column. A yes or no on the first line. A score out of five with one reason. When the answer has a fixed shape, there's nowhere for the hedging to hide.

Before:
Is ChatGPT good for writing emails?

After:
I write about 20 work emails a day: scheduling, follow-ups, and two or
three tricky ones. Make a table with columns: email type, would AI help
(yes/no/partly), what to watch out for. Then one line on whether it's
worth trying for a week.