Field notes on Arabic AI.
What we learn building frontier-quality Arabic data — written for the teams shipping Arabic models into production. No fluff, no vendor theater: dialects, compliance and evaluation, treated with the seriousness they deserve.
A fluent answer is not a completed transaction
An interactive, local-first crash test for what Arabic agents actually did—tool by tool, state by state.
Arabic annotation is not binary
78% on binary toxicity becomes 25.3% on the real taxonomy—and 22.4% across five dialect regions. The reproducible audit.
Model collapse speaks MSA: the limits of synthetic Arabic data
Collapse eats distribution tails first — and dialects are the tails. What the research means for Arabic, and the pipeline that avoids it.
The tokenization tax: the hidden cost of Arabic in English-first LLMs
Per-token pricing hides a 2–3× surcharge on Arabic. Where the tax comes from — and how to measure it on your own corpus.
Why dialect coverage decides your Arabic model's ceiling
MSA gets you reading. Dialect gets you understood. Where the real data gap sits — and what it costs in production.
PDPL & data residency: a checklist for MENA AI teams
What in-region really means under Saudi Arabia's PDPL, and the questions to ask any data partner before work begins.
How to benchmark an Arabic LLM you can actually trust
Beyond accuracy: why translated benchmarks mislead, and a five-dimension protocol for measuring what matters in dialect.
Working on one of these problems?
Tell us the task. We'll scope a pilot that proves the quality on your own data.
Talk to us