Human-System Trust
Designing how humans shape, inspect, and govern a living computational system. When software is alive and continuously generated, new trust models are needed.
Current Frontier
Governance models for AI software: how do humans maintain meaningful control over a system that generates its own behavior?
Key Questions
How do you inspect and audit software that doesn't exist as static code?
What governance models work for living software?
How do you build trust in a system that behaves probabilistically?
Key Papers
Is Sora a World Simulator?
May 2024
Challenges the claim that visual quality = world understanding. Critical framing for trust calibration.
Critiques of World Models as Planning
Jul 2025
Honest assessment of failure modes — shallow coherence, error explosion, generality limits.
World-in-World: Closed-Loop WM Evaluation
ICLR 2026 Oral
Benchmarking infrastructure for evaluating world models. Trust requires measurability.
Current Insights
Trust in The Last Computer is fundamentally different from trust in traditional software. You can't read the source code because there is no source code.
The intent contract becomes the auditable artifact. Instead of inspecting code, you inspect the intent specification and verify the experience matches.