#convictions

Training frontier models is cathedral work. Running useful inference is not. My bet: serious builders will own more of their intelligence than the API era wants them to.

Machines can make competence look abundant. They cannot make a person keep their word. In the post-AI era, integrity becomes the scarce infrastructure.

A revisit of my 2011 machine-awareness essay through Leopold Aschenbrenner's Situational Awareness, unhobbling, and the uncomfortable question of who owns the stack.

A blunt look at India's domestic angel and VC ecosystem: good headlines, weak ownership, currency drag, and too little global ambition.

Books are not files, and treating them that way has made digital publishing duller than it needed to be.