how elon musk uses ai daily
His seven rules
Seven things he actually does, and the clip where he says each one. Tap play on any of them.
He asks every new model the same three questions.
Not benchmarks. Three questions he already knows the shape of, so he can grade the answer instead of trusting it. Pick your own three, and where a model fails on what you know cold, stop trusting it on everything you do not.
- The Fermi paradox
- the universe is huge and old, so why have we never met anyone?
- Rocket engine design
- how to build an engine that does not blow up
- Electrochemistry
- what is actually happening inside a battery
His verdict on all three, in the clip below: terrible.
Photograph the thing. Do not describe it.
He points a camera at a circuit board and asks what is wrong with it. He sent his own MRI scan to Grok. Writing out a description throws away almost everything the model could have seen for itself.
When the answer is weak, push the same thread.
Do not rewrite the prompt and start again. Stay in the same conversation and say more, further, again. He gets his best output by escalating one thread, not by restarting a better one.
Ask what is false, half true, or missing.
Then tell it to rewrite the whole thing. That two step prompt is what his own company runs across Wikipedia and the rest of the internet. Point it at your doc, your page, your old proposal.
Never ask it to be right.
Ask for the assumptions, what follows from them, and how sure it is at each step. Then you decide. He calls it critical thinking 101, and it is the whole difference between checking a model and believing one.
It is most confident exactly where it is most wrong.
The harder and more important your question, the more likely the answer comes back polished, sourced and false. Confidence is not evidence. Check hardest where it sounds surest.
The smartest model is not his.
In July 2026 he said Anthropic's Fable is still clearly the smartest model. He owns a competitor. He also says no lab stays ahead by more than about six months, so never build your work around one of them.
Start with rule one.
Write down three questions you already know the answers to. That is your test for every model that comes out from now on.
Every clip on this page is him saying it, uncut, at the second it was said. Nothing here is a paraphrase.