# Nathan Douglas

https://bottomuplabs.ai/about/

I build the infrastructure that makes research experiments fast, cheap and trustworthy to run. I want to work alongside researchers on experiment design — asking what's actually hard, then building the harnesses, pipelines and benchmarks that answer it, and catching problems on small runs before anyone spends a big one. I bring nearly fifteen years of software engineering, including production systems in government and healthcare, and undergraduate study in philosophy, sociology and psychology.

My research interest is a method more than a topic: build the smallest system that still produces the failure, where ground truth is knowable, then intervene and measure. I want to apply it at every scale where things can go wrong:

- **Inside a single model:** mechanistic interpretability.
- **In a model's behavior:** machine psychology and model organisms.
- **Between models:** social psychology — whether agents conform, cascade, or mistake one source repeated for independent evidence.
- **Across the institutions that deploy them:** sociology — the evolution of norms and culture, the shifting of alliances, and the slow, locally reasonable handovers through which we are gradually disempowered.

These scales aren't separate: what a single model does shapes what many do together, and what they do together shapes the institutions that come to depend on them. My aim is to improve alignment by improving how well we can understand AI systems at every one of them.

I expect to grow into that over years rather than months; what I'm looking for now is an engineering role close enough to the research that I can learn it properly.

Based in northeast Ohio. Open to relocation, including internationally, for the right role, and comfortable and experienced working remotely.

## Timeline

- 2026-11-08: [Is This MNIST? Training Choices Steer What LLM Weight Readers Conclude](https://github.com/ndouglas/weight-stego) - Solo project in the Swarm Alignment Hackathon (AIAF and Apart Research). Training with Adam instead of SGD made frontier LLMs auditing a net's weights misidentify its training data, with no weight edits. One check re-derives every reported number.
- 2026-11-02: [ARENA AI safety curriculum](https://arena.bottomuplabs.ai/) - Started working through it independently and in public.
- 2026-10-31: [BlueDot Impact: Technical AI Safety](https://bluedot.org/courses/technical-ai-safety) - Completed the course.
- 2026-10-25: [A Legal Warning Can Prevent an LLM Cartel, but Not Break One](https://apartresearch.com/sprints/projects/a-legal-warning-can-prevent-an-llm-cartel-but-not-break-one-s3jy) - Solo project in Apart Research's AI Collusion Sprint. Released 350 runs that reproduce the reported numbers, and got a benchmark fix merged upstream.
- 2026-10-16: [BlueDot Impact: AGI Strategy](https://bluedot.org/courses/agi-strategy) - Completed the course, with a personal action plan on collective AI behavior, evaluation and control.

[Résumé (PDF)](https://resume.bottomuplabs.ai/resume.pdf)