Tl;dr

The AMTAIR project addresses the critical coordination failure in AI governance by developing computational tools that automate the extraction of probabilistic world models from AI safety literature using frontier language models. Starting with a World Model Extraction and Analysis Tool and building toward a comprehensive framework for a grand strategy for AI safety, we're creating an end-to-end pipeline that integrates dynamic Bayesian networks with live forecasting data to quantify risks, evaluate policy impacts across diverse scenarios, and generate adaptive strategic recommendations. The system makes implicit models explicit, facilitates cross-domain coordination, and provides policymakers with actionable insights based on formalized causal reasoning. Our team combines expertise in Bayesian modeling and AI governance—precisely the interdisciplinary skills needed to develop this strategic "operating system" for aligning disparate efforts before the window for establishing effective governance closes as AI capabilities continue to accelerate.

HTML Prototype Demonstrations

Project Description

AMTAIR: Automating Transformative AI Risk Modeling

The Automating Transformative AI Risk Modeling (AMTAIR) project addresses a critical coordination failure in AI governance: despite unprecedented investment in AI safety, we lack the strategic "operating system" needed to align disparate efforts across technical, governance, and policy domains. We are working on scaling up the research work done by the team of the original MTAIR Project. 

We're developing computational tools—starting with a World Model Extraction and Analysis Tool—that automate the extraction of probabilistic world models from AI safety literature using frontier language models. These tools will form the foundation for a comprehensive, adaptive AI Grand Strategy framework.

Our system architecture implements an end-to-end pipeline from unstructured text to actionable insights through five components:

1. Text ingestion and preprocessing using ArgDown derived syntax

2. LLM-powered extraction of world models, argument structures and probability distributions

3. Bayesian network construction with directed acyclic graphs

4. Live forecasting integration from prediction markets

5. Interactive visualization and analysis interface

Key deliverables include a World Model Extraction and Analysis Tool for quantifying existential risks, an AI Grand Strategy framework for coordinating responses across domains, an AGI Doomsday Clock for public communication, and cross-model comparison tools for identifying agreements and cruxes of disagreement.

The project creates impact by improving research directions, enhancing decision-making for policymakers, and facilitating coordination across domains:

Enhancing Decision-Making for Policymakers: Our tools provide structured frameworks for assessing policy impacts across multiple scenarios, explicit quantification of uncertainties, comparison of expert perspectives, and identification of robust interventions.

Facilitating Cross-Domain Coordination: Our tools create a shared epistemic infrastructure that facilitates communication between technical and governance researchers, alignment of strategies across organizations, and representation of diverse philosophical perspectives within a unified system.

Our 12-month implementation plan follows a phased approach with clear milestones, risk mitigation strategies, and early stopping points that would still yield valuable outputs. The core team combines expertise in Bayesian modeling, AI governance, strategic planning and community engagement—precisely the interdisciplinary skills needed for this challenge.

The window for establishing effective governance is narrowing as AI capabilities accelerate. AMTAIR creates the epistemic infrastructure necessary for coordinating humanity's response to what may be its most consequential technological development.