Doctoral Thesis:  Latent Levers: Representation-Guided Control For AI Supply Chains

By Aspen Hopkins

Details

• Date: Friday, May 1
• Time: 11:00 am - 1:00 pm
• Location: Star

Additional Location Details:

AI development is shifting from standalone models to AI supply chains—networks of interdependent models, datasets, and services spanning multiple actors. While these systems reduce cost and expand access, they introduce new challenges: components can be opaque, non-modular, non-deterministic, or evolve without disclosures.

This thesis provides the first in-depth analysis of AI supply chains as both technical and sociotechnical systems. I show that properties such as fairness and explainability emerge from interactions across the network, not from any single mode, and complement this with a stakeholder analysis that traces the organizational frictions and power asymmetries inhibiting coordination across supply chain boundaries. These findings point to a gap: existing tools are not designed for system-level control.

To address this, I introduce latent levers—derived from internal representations—that act as interfaces for controlling AI systems. Locally, latent levers can direct or control model behavior, either by surfacing intervention points for humans or by acting directly on a model’s internal state. I demonstrate this through interventions on data (to “design data”) and on a model trained on Othello sequences (to modulate behavior). Extending this to the supply chain, I show how latent levers support selective adoption of upstream model updates.

Together, this work argues for a new paradigm of AI infrastructure and proposes latent levers as a foundation for building controllable, efficient, and resilient AI supply chains.

Host

• Aspen Hopkins
• Email: dataspen@mit.edu