Multilingual AI evaluation · Adversarial probes · Building
Tarek Gara.
I work in the seam where language meets machine.
Multilingual AI quality at Innodata, across Arabic, Hebrew, and English. Adversarial probes where non-English alignment breaks down. Small focused tools: Folio, Tayyar, Thagrah.
Finishing an MA in Cyber Politics & Government at TAU. Writing essays at The Loop.
Currently
what's holding my attention this week- Leading
- multilingual AI quality at Innodata
- Building
- Folio, Tayyar, Thagrah
- Writing
- Deadly Sand, sequel to Mortal Past
- Editing
- The Loop, monthly
- Finishing
- an MA in Cyber Politics & Government at TAU
Writing
long-form on multilingual AI, alignment, and how products drift as they grow Feb 2026 Feb 2026 Jan 2026
Why AI thinks in English and forgets in Arabic
On tokenization, context windows, and what gets lost when Semitic languages are treated as bloated data.
How AI killed the soul of the world's favorite language-learning app
On Duolingo's pivot from learning to engagement, and what 850 days of streaks failed to teach me.
The sanitization of intelligence
On the alignment tax, corporate risk-aversion, and what "safe" AI gives up on its way to being safe.
The rest
books, ventures, playgroundsA letter, once a month.
Essays land here first, then I republish them as The Loop on LinkedIn.
Follow The Loop
About 385 readers across both. RSS at /feed.xml.