Writing.

The MVP began as a survival tactic for startups with no money and no users. Now the richest companies in the world can use it to ship below a bar they could clear, let machines grade machines, and turn users into unpaid quality assurance.

Saudi Arabia is buying AI infrastructure at historic scale, and a frustrated Saudi user just explained better than any analyst why the chips are the easy part: the Arabic training corpus is thin, half of it is translated English, and the model's own rules forbid it the conversations its users actually want.

Russia is flooding the open web with propaganda engineered to be scraped into AI training data. The tactic is real and inexpensive; whether it actually changes what the models tell you is a harder and still unsettled question.

The fear is about who can switch the models off. The real power is quieter, and harder to take back: who decides which version of the truth your AI serves you first.

Western left-right doesn't fit the region. So I built Tayyar — a panel of language models that scores MENA parties on sixteen axes — then used it to dig into Israel's fault lines and a live Knesset coalition simulator. The data says the Middle East barely splits on the economy.

A look at Duolingo's user-(retention)-centered approach.

An audit of the architectural friction and reasoning bottlenecks across Semitic languages.

On the alignment tax, corporate risk-aversion, and what "safe" AI gives up on its way to being safe.

On the moment a product still works but its logic stops cohering.

كتبتُه في خضمّ الجائحة، حين انتقلت المدارس فجأةً إلى الشاشات: التعليم عن بُعد ليس وليد كورونا ولا الإنترنت، بل حلقةٌ في سلسلةٍ بدأت من البريد. وربّما كان التعلّم الذاتيّ، لا المدرسة، هو مستقبل المعرفة.