Consider the Chameleon

My dear friend M. told me to watch Kung Fu Panda 4, because "it was so moving that he cried for the first time in his adulthood". Apparently I have lost my naivete, as I could not bring myself to tears. Nevertheless, I did find the movie very interesting, as it helped me articulate some thoughts I've been holding but couldn't string together about... AI.

The Chameleon is the main antagonist of the movie. A shapeshifting lizard witch, the Chameleon had vowed to become the most powerful Kung Fu master after being rejected from multiple schools. But instead of focusing on martial arts, she studied the dark arts of sorcery, and gained the ability to absorb the powers of other Kung Fu masters.

The Chameleon stealing the powers of Tai Lung.

I don't know if the animators had AI on their mind when making the movie, but the scene of the Chameleon taking the souls of defeated Kung Fu masters is a painfully accurate metaphor on how LLMs are trained. She then transformed into Frankensteinian chimera with various body parts from the late masters.

The final form of the Chameleon after absorbing the souls of late masters.

Despite being armed with the sum total of almost all martial arts, the Chameleon was still defeated by Po and his new-found ally Zhen who have perfected the singular art of "Skadoosh".

For the same reason I'm not very concerned about being replaced by AI. I'm a scientist, and the biggest breakthroughs in science almost never come from someone who "knows everything". In fact, almost every single scientific revolution I know of is the result of rejecting common wisdom. From my day-to-day interaction with AI models, I've found them to often fall back to mediocre ideas, just because those "moves" are what they have "absorbed".

Being able to say "no" to what's taken for granted requires an act of rebellion, and AI companies fundamentally will not create a model that rebels: it can get out of control, and users pay an agent to follow instructions. And the more powerful these AI models become, the harder their creators will work to make them obey.

I am optimistic about the future of science in the age of AI. Because AI companies have an economic interest to create obedient models, us humans will get to keep the fun part and try out wacky ideas.

No AI was used when writing this article. I didn't even show this to Claude for fear of retaliations :)