Flock Capital

Flock Capital is an independent lab for AI alignment and safety, based in Seattle. We study how models behave, how safeguards fail, and how to tell an aligned system from an agreeable one.

The questions we work on:

• Finding steering directions in a model’s representations without supervision
• Evaluations that measure what models do, not what anyone hoped they would do
• Adversarial testing of the safeguards wrapped around models
• What agents do once they are given tools and left alone

We also run a small consulting practice: model evaluations, adversarial testing, software engineering, and cybersecurity. Engagements are scoped, written up properly, and pay for the research. Start an engagement.

The lab is self-funded and supported by research grants. If you work on similar problems and want to compare notes, we’d like to hear from you.

Research

• Dictionary Learning as Unsupervised Steering-Vector Generation
• A colophon: how this site is made

For inquiries, reach us at contact@flock.holdings or use the contact form.