Researcher Profile
About
Research Motivation
Why This Work
This platform documents independent research on AI alignment, governance, and forecasting. The work is motivated by a straightforward observation: the development of advanced AI systems is proceeding faster than our institutional capacity to govern it, and faster than our technical capacity to verify its safety.
Closing these gaps requires researchers who can move between technical and policy domains — who understand both the empirical claims being made about AI capabilities and the governance mechanisms available to respond to them. This research attempts to occupy that space.
The platform is explicitly not neutral. It takes seriously the possibility that advanced AI systems pose catastrophic risks, evaluates evidence for and against that view rigorously, and advocates for governance frameworks proportionate to the risk profile.
Intellectual Trajectory
How I Got Here
Designed and built TQC — a three-component quantitative trading system using Hidden Markov Model regime classification, GARCH volatility scaling, and lock-free concurrency in C++20. Building it confronted me with emergent system behaviour I could not predict from component-level understanding, which led me directly to AI safety research.
Began systematic engagement with the AI safety literature across three tracks: forecasting, governance, and mechanistic interpretability. The connection between HMM belief states in TQC and how transformers represent latent structure internally became the bridge between my systems background and alignment research.
Current focus: the question of whether understanding AI system internals enables more reliable predictions about deployment behaviour, and what governance frameworks can remain legible under rapid capability change. Building this research platform as a long-term knowledge system.
Methodology
Forecasting Approach
Forecasts on this platform use a modified Bayesian updating approach with explicit reference classes. The process for each forecast:
- 01
Establish a reference class of similar historical events and derive a base rate.
- 02
Adjust for inside-view factors specific to the question — technological trajectory, institutional capacity, geopolitical dynamics.
- 03
Set explicit resolution criteria before publishing the forecast to prevent post-hoc revision.
- 04
Update on material new evidence (model releases, regulatory developments, empirical results) with logged rationale.
- 05
Track calibration over time. Resolved forecasts are archived with their resolution outcome.
All probability estimates are personal views, not institutional positions. The goal is calibration — not confidence for its own sake. A 40% probability means the event is considered more likely to not occur than to occur, with meaningful uncertainty in both directions.
Research Agenda
Current Interests
AI Timeline Forecasting
Reference class forecasting for transformative AI milestones with explicit calibration tracking and Bayesian updating.
Compute Governance
The structural limits of hardware-based AI regulation and what complementary mechanisms are needed.
Institutional Coordination
How international AI governance institutions can be designed to avoid race dynamics and remain coherent under rapid capability change.
Mechanistic Interpretability
Whether internal structure — circuits, belief states, residual stream geometry — can serve as a foundation for meaningful third-party oversight.
Specification Gaming
Whether governance frameworks that rely on complete behavioural specification are structurally inadequate, and what replaces them.
Deployment Behaviour
The gap between designed and actual AI system behaviour, and whether interpretability research can close it for forecasting purposes.
Long-Term Vision
Where This Is Going
This platform is designed to evolve. Near-term goals include establishing a systematic paper review pipeline covering the core AI safety and governance literature, a forecasting section with explicit resolution criteria, and deeper governance analysis of the international AI regulatory landscape.
The long-term goal is to contribute to governance frameworks that are technically grounded, institutionally realistic, and adequate to the risk profile of advanced AI systems. This requires both the technical analysis done here and engagement with policymakers — which is the next phase of this work.