Framing: Open World Development — the second of the two theses Sadra organises his work around; the first is Research Agenda, framed on his site as AI Integration Challenges. Organisation: co-founded OpenSciLab with a group of friends. Separately, Sadra co-leads an open-source team that builds scientific tools.

Summary

Open-sourcing NLP research is what got the field to things like ChatGPT. The same openness now makes it very cheap to produce work that looks convincing and is quietly wrong. Sadra calls this the Frankenstein-Trojan problem: the monster is one we built ourselves, and since there is no reliable way to evaluate it, it walks straight into our own science unnoticed. The threat is not an outsider attacking the field’s standards — it is the field’s own output, assembled from legitimate-looking parts, arriving inside the gates unchallenged.

Auditable Science

The response is to write artifacts that people can rerun. Software is one of the few domains where nobody has to take anyone’s word for anything: no lab, no samples, no budget — you run the thing on your own machine and see whether you get what the author got. That makes it an unusually auditable science, and Sadra’s position is that the field does not lean on this nearly enough. It follows that reproducibility and transparency are properties of the tooling rather than of the paper: software that makes its own workings inspectable is what lets the check actually happen.

OpenSciLab

OpenSciLab was co-founded with a group of friends, to build those rerunnable artifacts.

The Work

The output is grouped by topic — OpenSciLab projects, datasets, and things Sadra built on his own:

OpenSciLab

Open Science Laboratory (OpenSciLab) — the organization and its output. O — the clearest expression of the idea: non-executable, inspectable model serialization.