The ASCII cheetah interview
How do you check, during an interview, whether someone uses AI efficiently? In the past, the answer would have been trivial: spend an hour or two building something together using an agentic harness of their choice. But the new generation of models is heavily optimized for one-shot prompts, which makes this trickier. Asking a candidate to build XYZ when they could just forward the prompt to a model and get decent results is futile. There is, however, a funny set of niche problems that models find surprisingly difficult. One of them is drawing animals using ASCII art. Below, you can see some cheetahs drawn by different models.
Muse 1.3
/\_/\ ___
( o.o ) / \\____
> ^ < | cheetah \
/| |\\ \\ \\ \\
(_| |_) \\_____/ /
/ ___ \\ / /
/ / \\ \\ / /\\ /\\
/ / \\ \\ / \\_/ \\
/ / __ \\ \\/ ___ \\
/ / / \\ \\ \\__/ \\ \\
/ / / \\ \\ \\ \\
( ( ( ) ) spots ) )
\\ \\ \\__/ / . . . / /
\\ \\_______/ . : . / /
\\ . : . / /
\\_________. .: . /__/
| |
_| |_
/ \\
/ /\\ /\\ \\
| | | | |
| | | | |
| | | | |
_| | | | |_
(___) (___)(___)
Luna 6.0
/\_/\
( o.o ) >==3
> ^ <
Grok 4.7
.--.
/ \
| .--. |
| | | |
| '--' |
\ /
'--'
___ ___
/ \ / \
| o | | o |
\___/ \___/
\ /
\ __ /
\/ \/
/ \
| CHEE |
| TAH |
\ /
\____/
/ \
| spots |
\ /
'----'
Just ask the candidate how they'd improve the quality of the cheetahs being drawn.
For reference, this is as far as I was able to push Luna 6 for $0.03 with a single prompt. Clearly, these are AGI-worthy results. Why we aren’t seeing RSI yet is beyond my understanding.
/\ /\
/ \___/ \
| o\ /o |
| \_/ |__
\ . . / `-.
___..---..__________/ . . . / . . . \
==~~~~\ ____/ . o . . . . . . . . . . . . . . . . |
`--' . . . . . . . . . . . . . . . . . . . . /
\ . . . . . . . . . . . . . . . . . . . /
`--..____ . . . . . . . . . . ____..--'
/| /| |\ |\
/ | / | | \ | \
/__|/__| |__\|__\
Share any ideas how to improve further with @lazilyevaluated