“Charlie Holtz believes that AI models can enter a negative feedback loop where they see bad code and consequently write more bad code.”