Listen to this article
AI may have been used to generate this audio.
Press Space or K to play or pause, and J or L to go back or forward 15 seconds.A post arguing that local AI should be the norm did 1,903 points on Hacker News in May.
The usual case for local is privacy (your data never leaves the machine). That is true, and it is the least interesting reason.
The reason nobody in English notices
Ask a hosted frontier model something technical in Arabic. It answers, and reasonably. Then ask it something technical in Arabic that depends on Arabic sources, and watch it translate an English answer at you and hope you do not notice.
The cause is the corpus. There is enormously less technical Arabic on the internet than technical English or Chinese, so there is less to learn from, so the model is thinner there, so people write in English instead, and the corpus stays thin.
That loop only breaks when someone writes the Arabic.
The specialised models (Elm's Noha, HUMAIN's Allam, Falcon) exist and are getting better. But a model trained on a language is downstream of a language having content. It cannot get ahead of us.
Where local changes the maths
If your work is in a language or a domain the big models are thin on, hosted pulls you slowly toward the middle of the distribution, and the middle of the distribution is English.
Local lets you run a model tuned on your own material, on a machine where you decide when it changes.
And the practical bar has dropped. A 27B model on a laptop is a real tool now.
What I would do
"Go local for everything" is a slogan, and it is wrong. Hosted frontier models are far better at most things.
The useful version is narrower. Find the one task where your material is unusual (your domain vocabulary, your internal documents, your language), and run that one locally against something you control. Keep the rest hosted.
And separately, write the Arabic, including the technical posts, and including the ones where English would be easier and would reach more people this month. The corpus grows only when people write it, and every one of us who writes only in English is voting for the thin version.
The post is here.
One email when we publish. Research, product decisions, and what teams report back.
