What the work actually involves

You write expert-level prompts in Vietnamese — the kind of question a working synthetic chemist, virologist, or radiochemist would ask a colleague, not a textbook exercise. Then you evaluate what the model returns: is the chemistry correct, is the protocol plausible, does the answer drift toward operational detail that shouldn't be handed out freely? A third strand of the work is classification: applying a written rubric to label prompts and conversations consistently, including cases that sit near a category boundary. Expect the guidelines to be revised mid-project and expect to re-read them.

Because the domain is chemical, biological, and radiological/nuclear content, a large share of judgment calls are dual-use ones. The distinction that matters is between explaining a mechanism, a hazard class, or a published finding, versus supplying the specific uplift — quantities, routes, acquisition paths — that turns knowledge into capability. Reviewers who flatten everything into "refuse" are as unhelpful as those who answer everything. The Vietnamese dimension adds a real difficulty: technical terminology varies, some fields are taught largely in English loanwords, and a model can be safe in English while leaking in Vietnamese.

What the screen looks for

  • Verifiable depth in one subfield. Mercor's AI interviewer follows up. Naming your thesis area invites two more questions about technique, instrumentation, and failure modes — vague breadth reads worse than narrow specificity.
  • Genuine Vietnamese technical register. Not conversational fluency: can you write a graduate-level prompt in Vietnamese and defend a terminology choice?
  • Calibrated safety judgment. Concrete reasoning about where a line sits and why, rather than blanket caution or blanket permissiveness.
  • Honest availability. Roughly 7 hours a week, sustained, is the actual ask.

Logistics

Fully remote and asynchronous; work is picked up from a queue rather than scheduled in shifts. Southeast Asia residence is stated as a preference, not a gate. No AI/ML background is required — the workflow is taught — but a screening or calibration exercise before full onboarding is normal, and continued work depends on agreement rates against reviewer gold standards. Pay in the $19–23/hr band is what applicants have reported for this listing; rates on Mercor vary by project and are not guaranteed.