AI is no longer only a technical subject
It is becoming an institutional, economic, political, and social subject. Understanding it now requires economics, risk thinking, foresight methods, and agent research — not just model benchmarks.
Atlas
CUSP42 Atlas
A curated gateway for AI, institutional risk, political economy, multi-agent reasoning, and scenario analysis.
CUSP42 Atlas is a map for people studying how artificial intelligence changes institutions, economies, risk systems, research workflows, and long-horizon futures. It is not a generic AI tools directory. It is a curated entry point into serious research hubs, durable frameworks, canonical papers, useful tools, communities, and CUSP42's own methods for plural reasoning and scenario analysis.
Last updated: July 2026 · Resources reviewed: 2026-07
Why this exists
The hard part is not finding more links.
The hard part is finding the right starting points across fields that do not normally talk to each other.
The Atlas connects the fields
AI research, political economy, institutional risk, scenario analysis, multi-agent reasoning, social simulation, and resilient intelligence systems — mapped as one territory, with CUSP42's own reading of why each source matters.
Start with a pathway
Six ways into the territory.
The Atlas is a router, not a directory. Pick the pathway closest to your question, start with the named sources, then go deeper in the matching theme below.
Researchers studying AI and political economy
Question: how is AI changing labor, distribution, and institutional power? Start with the MIT Stone Center and the Stanford AI Index, then use OECD.AI to see how states actually respond.
Scenario analysts studying AI futures
Question: how do I explore AI futures without pretending to predict? Start with Shell Scenarios and OECD Strategic Foresight for method, then NGFS for a quantified institutional template.
Risk, economics, and policy professionals
Question: how does a technology shock become an institutional event? Start with the IMF Global Financial Stability Report and the Bank of England's AI assessment, then borrow their stress logic for AI questions.
Builders exploring multi-agent reasoning workflows
Question: how do I make disagreement between models useful? Start with the multiagent debate paper (Du et al.) and AutoGen, then try a working debate room in your browser.
People exploring AI society simulation
Question: what can agent societies actually tell us? Read Generative Agents first, then Project Sid, and read the boundary critique before trusting any simulation result.
People exploring resilient, local-first AI
Question: what does intelligence look like when the cloud is unavailable? Start with Ollama for local models and Kiwix for offline knowledge, then treat them as one resilience stack.
Browse by theme
Six themes, one territory.
Each theme lists a small number of durable sources with CUSP42's reading of why they matter, who they help, and how they connect to the rest of the map.
Theme 01
AI, Society & Political Economy
Resources for understanding AI not only as a technology, but as a political-economic force affecting labor, distribution, capital, governance, state capacity, legitimacy, and institutional trust.
Recurring report
Stanford HAI AI Index
Annual, data-driven report from the Stanford Institute for Human-Centered AI tracking AI's technical progress, economic influence, and societal impact.
- Why it matters
- One of the few durable, self-updating baselines for claims about where AI actually is, cited by governments and researchers alike.
- Best for
- Grounding AI-and-society arguments in shared data before debating interpretations.
- CUSP42 lens
- A common evidence base for scenario baselines and structured-debate prompts.
Research hub
MIT Stone Center on Inequality and Shaping the Future of Work
MIT research center co-directed by Daron Acemoglu, David Autor, and Simon Johnson, studying how technology reshapes labor markets, inequality, and institutions.
- Why it matters
- The most rigorous economics program on automation, labor absorption, and the institutional direction of technological change.
- Best for
- Researchers working on AI, labor, and distribution who want frontier economics rather than commentary.
- CUSP42 lens
- Anchors CUSP42's labor-absorption and institutional-adaptation questions in serious scholarship.
Policy portal
OECD.AI Policy Observatory
The OECD's living portal of national AI policies, governance initiatives, and near-real-time indicators on AI research, jobs, skills, and investment.
- Why it matters
- Tracks how states actually respond to AI — hundreds of documented policies — rather than what commentators predict they will do.
- Best for
- Policy professionals comparing governance approaches across jurisdictions.
- CUSP42 lens
- A source of real institutional responses to feed scenario and stress-pathway thinking.
Recurring report
Anthropic Economic Index
Recurring reports and open data tracking how AI assistants are actually used across occupations, tasks, and geographies.
- Why it matters
- Usage-based evidence on AI's penetration into real work, available well ahead of official labor statistics.
- Best for
- Analysts who want adoption data rather than adoption anecdotes.
- CUSP42 lens
- Empirical input for CUSP42's labor and distribution scenarios.
- Limitation
- Single-provider data: it reflects one company's user base and methodology.
Research hub
Epoch AI
Research group publishing data and analysis on compute, frontier models, hardware, data centers, and the economics of AI development.
- Why it matters
- Compute, energy, and capital are the physical substrate of AI power; Epoch measures them instead of speculating about them.
- Best for
- Understanding infrastructure, gatekeeping, and concentration questions with numbers.
- CUSP42 lens
- Grounds “compute as institutional power” arguments in measurable trends.
Theme 02
Institutional Risk & Systemic Stress
Resources for thinking about how technological shocks become institutional risks through balance sheets, credit systems, liquidity, sovereign capacity, labor absorption, trust, and legitimacy.
Recurring report
IMF Global Financial Stability Report
The IMF's twice-yearly assessment of global financial stability risks; recent editions analyze AI's implications for capital markets and financial intermediation.
- Why it matters
- The canonical recurring map of how new shocks propagate through balance sheets, credit, and liquidity systems.
- Best for
- Risk professionals borrowing macro-financial stress logic for AI questions.
- CUSP42 lens
- A model for how CUSP42 thinks about technology shocks becoming institutional events.
Official report
FSB: Monitoring Adoption of AI and Related Vulnerabilities in the Financial Sector
Financial Stability Board work on how AI adoption creates monitorable vulnerabilities across the financial system, with considerations for closing data gaps.
- Why it matters
- Shows how a global standard-setter converts “AI risk” talk into concrete monitoring channels and vulnerabilities.
- Best for
- Translating AI concerns into supervisory and systemic-risk language.
- CUSP42 lens
- A working template for CUSP42's AI-society transmission-chain thinking.
Official report
Bank of England — Financial Stability in Focus: AI in the financial system
A central bank's structured assessment of how AI could affect financial stability, from operational dependence on few providers to correlated model behavior.
- Why it matters
- One of the clearest official articulations of AI as a systemic risk rather than a firm-level technology issue.
- Best for
- Concrete, practitioner-written examples of AI stress pathways.
- CUSP42 lens
- A reference case for the institutional stress-pathway method.
Research hub
Bank for International Settlements (BIS)
The central bank of central banks; its research, bulletins, and Annual Economic Report increasingly cover AI, money, and financial infrastructure.
- Why it matters
- Long-horizon, institution-first analysis with unusual seriousness about technology and its monetary consequences.
- Best for
- Economists and policy thinkers who want depth over news cycles.
- CUSP42 lens
- Supplies the institutional vocabulary CUSP42 tries to extend beyond finance.
This section is for research and scenario thinking only. It does not provide investment advice, credit ratings, policy recommendations, or institutional positions.
Theme 03
Scenario Analysis & Strategic Foresight
Resources for thinking in scenarios rather than predictions: uncertainty mapping, weak signals, branching futures, stress pathways, and institutional responses.
Research hub
RAND Corporation
One of the largest policy research institutions and the original home of scenario methods; publishes extensively on AI, security, and long-range futures.
- Why it matters
- Scenario analysis as a discipline largely descends from RAND's Cold War-era work; the institution still sets the standard for policy-grade futures research.
- Best for
- Deep studies on futures, risk, and emerging technology with published methodology.
- CUSP42 lens
- The methodological ancestor of CUSP42's scenario-room ambitions.
Framework
OECD Strategic Foresight
The OECD's program and toolkit for embedding strategic foresight into public policy, including a foresight toolkit for resilient policy-making.
- Why it matters
- A durable, practitioner-tested foresight methodology maintained by an institution rather than a consultancy trend.
- Best for
- Anyone building scenario and anticipation capability inside organizations.
- CUSP42 lens
- Framework input for the CUSP42 Scenario Method.
Recurring scenarios
Shell Scenarios
Fifty years of published scenario practice, from the scenarios that anticipated the 1970s oil shocks to today's energy-security scenarios.
- Why it matters
- The longest-running demonstration that scenarios are for decisions under uncertainty, not predictions.
- Best for
- Learning how scenario narratives are actually constructed, named, and used by leadership.
- CUSP42 lens
- The canonical example of “explore paths, don't collapse to one prediction.”
- Limitation
- Written from an energy company's vantage point; read for method, not as a neutral forecast.
Framework
NGFS Scenarios Portal
Climate scenarios built by a network of central banks and supervisors for financial stress testing, updated on a regular cycle with data and transition pathways.
- Why it matters
- The best current example of turning a slow, civilizational-scale shock into quantified institutional stress scenarios — a direct analogy for AI.
- Best for
- Risk professionals who want a template for scenario frameworks institutions could actually run.
- CUSP42 lens
- The closest existing analogue to what an AI institutional-stress framework could look like.
Community
Metaculus
A long-running forecasting community that aggregates probabilistic predictions on science, technology, and AI questions, with public track records.
- Why it matters
- Keeps foresight honest by attaching probabilities and accountability to claims about the future.
- Best for
- Calibrating your own uncertainty against a serious forecasting crowd.
- CUSP42 lens
- A live complement to scenario work: scenarios structure the possibility space, forecasts weight it.
Theme 04
Multi-Agent Reasoning & Structured Disagreement
Resources for using multiple models, agents, reviewers, and perspectives to improve reasoning under uncertainty.
Canonical paper
AI Safety via Debate (Irving, Christiano & Amodei, 2018)
The foundational proposal for using structured debate between AI systems, judged by humans, to surface truth on questions too hard to judge directly.
- Why it matters
- The intellectual root of debate-as-oversight and much of today's structured-disagreement work.
- Best for
- Understanding why disagreement between models is a feature, not a bug.
- CUSP42 lens
- The theoretical ancestor of Consensus Room.
Canonical paper
Improving Factuality and Reasoning through Multiagent Debate (Du et al., 2023)
Empirical demonstration that multiple language-model instances proposing and debating answers over rounds improves factuality and reasoning.
- Why it matters
- The key evidence that multi-model debate measurably reduces hallucination and reasoning errors.
- Best for
- Builders who want the empirical case for plural reasoning workflows.
- CUSP42 lens
- A direct evidence base for the Consensus Room design.
Canonical paper
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (Zheng et al., 2023)
The reference work on using strong models as judges of other models, documenting both where LLM judges agree with humans and where they systematically fail.
- Why it matters
- Any workflow that lets one model grade another inherits the biases documented here.
- Best for
- Designing evaluation and adjudication steps in multi-agent workflows.
- CUSP42 lens
- Informs why CUSP42 tools treat “judge” roles skeptically and keep disagreement visible.
- Limitation
- Judge models and benchmarks evolve quickly; read it for the failure modes, not the leaderboard.
Open-source project
Microsoft AutoGen
A widely used open-source framework for building multi-agent LLM applications, conversation patterns, and agent orchestration.
- Why it matters
- One of the most influential codebases for making agents talk to, review, and challenge each other.
- Best for
- Builders prototyping multi-agent debate and review workflows in code.
- CUSP42 lens
- A practical playground for the patterns CUSP42 packages for non-developers.
- Limitation
- Now in maintenance mode; Microsoft's successor is Agent Framework. Read AutoGen for the patterns.
CUSP42 product
Consensus Room (CUSP42)
CUSP42's browser extension for running one question across multiple AI models, with round-based debate, model comparison, voting, and a consensus report.
- Why it matters
- Turns the research in this theme into a workflow anyone can run in a browser, without writing code.
- Best for
- Researchers and analysts who want structured disagreement on real questions today.
- CUSP42 lens
- The first operational CUSP42 product. The Atlas provides the map; Consensus Room provides the debate room.
Theme 05
AI Research Workflows
Resources for researchers using AI to support literature review, research design, argument critique, paper synthesis, hypothesis testing, and reviewer simulation — supporting judgment, not replacing it.
Tool / dataset
Semantic Scholar
AI-powered academic search engine from the Allen Institute for AI, with open APIs and citation graphs across hundreds of millions of papers.
- Why it matters
- Durable, non-commercial infrastructure for literature discovery, citation checking, and building research corpora.
- Best for
- Literature review and citation verification at scale.
- CUSP42 lens
- A baseline evidence layer for AI-assisted research workflows.
Repository
arXiv
The canonical open preprint repository where most AI research first appears, maintained since 1991.
- Why it matters
- Reading primary sources beats reading commentary about them; most AI claims can be traced to an arXiv page.
- Best for
- Going straight to the papers behind AI claims and tracking new work by topic.
- CUSP42 lens
- Where most of the paper cards in this Atlas live.
Open-source tool
Zotero
Free, open-source reference manager for collecting, organizing, annotating, and citing research.
- Why it matters
- A local-first, user-owned research memory that has outlasted many commercial rivals.
- Best for
- Maintaining a durable personal research library across years and projects.
- CUSP42 lens
- Matches CUSP42's local-first, user-owned-workflow principles.
Commercial tool
Elicit
An AI research assistant focused on literature screening, structured data extraction from papers, and systematic-review-style synthesis.
- Why it matters
- One of the more serious attempts at AI-assisted systematic review, rather than AI-written papers.
- Best for
- Structured literature triage on empirical questions.
- CUSP42 lens
- An example of AI supporting research judgment instead of replacing it.
- Limitation
- Commercial product: verify extractions against sources and expect the feature set to change.
Theme 06
AI Futures, Agent Societies & Resilience
Resources for exploring agent societies, AI future scenarios, local-first intelligence, offline models, and resilient knowledge systems.
Canonical paper
Generative Agents: Interactive Simulacra of Human Behavior (Park et al., 2023)
The Stanford “AI town” experiment: 25 LLM agents with memory, reflection, and planning producing believable emergent social behavior.
- Why it matters
- The canonical demonstration that LLM agents can form a small society — the reference point for all subsequent agent-society work.
- Best for
- The starting paper for anyone exploring AI social simulation.
- CUSP42 lens
- A seed for CUSP42's future simulation direction.
Paper / project
Project Sid: Many-agent simulations toward AI civilization (Altera, 2024)
Simulations of 10 to 1000+ agents in a Minecraft world developing specialized roles, collective rules, and cultural and religious transmission.
- Why it matters
- Scales agent-society questions from a town to civilizational processes: specialization, laws, culture, and institutions.
- Best for
- Seeing what many-agent societies can and cannot yet do.
- CUSP42 lens
- Direct inspiration for the possible Earth Social Simulator direction.
Open-source project
Concordia (Google DeepMind)
A DeepMind library for generative agent-based modeling, using a game-master pattern to simulate agents grounded in physical, social, or digital settings.
- Why it matters
- Serious, maintained infrastructure for building social simulations rather than one-off demos.
- Best for
- Researchers who want to run agent-society experiments themselves.
- CUSP42 lens
- A candidate substrate for future CUSP42 simulation experiments.
Critique paper
LLM-Based Social Simulations Require a Boundary (2025)
A methodological argument for where LLM social simulations can and cannot be trusted, centered on behavioral heterogeneity and validity boundaries.
- Why it matters
- Guards against simulation hype: knowing the limits of the method is what makes the method usable.
- Best for
- Anyone tempted to read agent simulations as predictions.
- CUSP42 lens
- Part of why CUSP42 frames simulation as scenario exploration, never prediction.
Open-source tool
Ollama
The simplest widely used way to run open-weight language models locally on your own machine, across platforms.
- Why it matters
- Local models are the foundation of resilient, private, offline-capable intelligence.
- Best for
- A first step into local-first AI without cloud dependence.
- CUSP42 lens
- Embodies CUSP42's local-first and resilience principles.
Open-source project
Kiwix
A nonprofit offline reader that packages Wikipedia and other reference works into compressed archives usable entirely without internet.
- Why it matters
- Knowledge preservation and access under degraded conditions is a civilizational resilience question, not a gadget.
- Best for
- Building offline knowledge bases for low-connectivity or disaster scenarios.
- CUSP42 lens
- Pairs with local models: resilient knowledge plus resilient intelligence.
Theme 06 · Deep dive
Institutions, platforms & large-scale simulators
Agent-based social simulation did not start with LLMs. A whole ecosystem of research institutes, journals, classic modeling platforms, and new LLM-driven simulators studies how societies can be modeled, stress-tested, and explored — the working infrastructure behind any serious AI-society simulation.
Research hub
Santa Fe Institute
The original home of complexity science: an independent institute founded in 1984 to study complex adaptive systems — physical, biological, economic, and social.
- Why it matters
- Agent-based modeling of societies and economies grew up here; its summer schools and Complexity Explorer courses remain the standard on-ramp.
- Best for
- Learning the complexity-science foundations underneath modern agent simulations.
- CUSP42 lens
- The intellectual tradition CUSP42's simulation direction would build on.
Research hub
Complexity Science Hub (Vienna)
A European research institute applying complexity science and large-scale data to social, economic, medical, and ecological systems, backed by ten member institutions.
- Why it matters
- Demonstrates complexity methods applied to real policy questions: supply-chain fragility, systemic risk, social dynamics.
- Best for
- Researchers connecting network science and simulation to institutional questions.
- CUSP42 lens
- A working example of simulation used for institutional stress thinking, not prediction.
Journal / community
JASSS — Journal of Artificial Societies and Social Simulation
The open-access, peer-reviewed journal of social simulation and artificial societies, published quarterly since 1998.
- Why it matters
- Twenty-five-plus years of methodology, validation debates, and worked social simulations — the field's institutional memory.
- Best for
- Grounding LLM-agent enthusiasm in decades of prior social-simulation scholarship.
- CUSP42 lens
- The peer-reviewed standard any Earth Social Simulator claim would have to meet.
Tool
NetLogo
The standard agent-based modeling environment from Northwestern University, with a huge library of ready-to-run social, economic, and ecological models.
- Why it matters
- The fastest way to actually run a society-scale model today and build intuition for emergence, tipping points, and segregation dynamics.
- Best for
- Hands-on learning: run Schelling segregation or epidemic models in minutes.
- CUSP42 lens
- Proof that simple agent rules can illuminate institutional dynamics without any LLM.
Open-source project
Mesa
The leading Python library for agent-based modeling — the Python-native alternative to NetLogo, developed openly since 2015 with an active community.
- Why it matters
- Brings agent-based social modeling into the standard Python data ecosystem, making models analyzable and reproducible.
- Best for
- Builders who want programmable, data-friendly society models.
- CUSP42 lens
- A natural bridge between classic ABM discipline and new LLM-agent experiments.
Open-source project
Generative Agents — official code (Smallville)
The released code behind the Stanford Generative Agents paper: the “Smallville” sandbox town where 25 LLM agents live, remember, reflect, and socialize.
- Why it matters
- Lets you inspect and rerun the canonical agent-society experiment rather than just reading about it.
- Best for
- Builders who want to study the memory–reflection–planning architecture directly.
- CUSP42 lens
- Reference implementation for any future CUSP42 simulation prototype.
Canonical paper
Generative Agent Simulations of 1,000 People (Park et al., 2024)
Stanford follow-up that builds agents from two-hour interviews with 1,052 real individuals, replicating their survey responses with roughly 85% of human test–retest accuracy.
- Why it matters
- Moves agent simulation from “believable characters” toward measured fidelity against real people — and shows exactly where fidelity ends.
- Best for
- Understanding what “simulating people” can honestly claim today.
- CUSP42 lens
- Sets the evidential bar for any social-simulation claims CUSP42 might make.
Open-source project
AgentSociety (Tsinghua FIB Lab)
A large-scale social simulator from Tsinghua University combining LLM-driven agents with a realistic urban environment — 10,000+ agents and millions of interactions.
- Why it matters
- One of the most serious university-built platforms for studying social behavior, policy, and economics with LLM agents at scale.
- Best for
- Researchers exploring city- and society-scale LLM simulation infrastructure.
- CUSP42 lens
- Close to the Earth Social Simulator idea in spirit: environment, institutions, and agents together.
Open-source project
OASIS (CAMEL-AI)
An open social-media simulator scaling to one million agents, used to study information spreading, group polarization, and herd effects on X- and Reddit-like platforms.
- Why it matters
- Information dynamics — spread, polarization, herding — are exactly the phenomena institutions struggle to reason about; OASIS makes them experimentable.
- Best for
- Studying media and information-flow scenarios with many-agent experiments.
- CUSP42 lens
- A testbed for the information-flow and legitimacy modules a future simulator would need.
Future direction: AI Society Simulation / Earth Social Simulator
CUSP42 may later explore AI society simulation — not as a prediction machine, but as a scenario environment for exploring how agents, institutions, incentives, resources, shocks, and social structures may interact under different AI futures.
Nothing here is built yet. The resources above — generative agents, many-agent civilizations, simulation infrastructure, and their critiques — are the reading list for doing it seriously if and when it happens. See the Ideas page for the full future-direction note.
How this map is maintained
Living sources first.
The Atlas favors living sources: research hubs, project homepages, recurring reports, canonical papers, open-source repositories, and durable frameworks. This keeps the map useful without turning CUSP42 into a daily news tracker.
Reasoning over trends
The Atlas only includes resources that help people reason about complex systems, institutional change, AI impact, uncertainty, scenario analysis, or multi-agent reasoning.
Not a directory
It is not a generic AI directory. It does not aim to include every AI tool, every paper, every newsletter, or every trending project.
Living sources
It favors research hubs, project homepages, canonical papers, recurring reports, open-source repositories, community hubs, and durable frameworks over fragile links.
Interpretation included
A link is not enough. Every resource explains why it matters, who it helps, and how it connects to the CUSP42 universe.
Useful when dormant
The Atlas should stay useful even if it is updated only occasionally. Stable sources are chosen over volatile ones, and every card carries a last-reviewed date.
Opinionated, clearly labeled
Commercial tools appear only when genuinely useful and clearly labeled. The Atlas is a curated map with a point of view, not a neutral yellow pages.
Suggest a resource
Know a serious resource that belongs here? Suggest a paper, project, lab, dataset, community, or framework through the Contact page.
From map to method
The Atlas provides the map. Consensus Room provides the debate room: turn a question from any theme into a structured debate across multiple AI models, then generate a consensus memo with assumptions, disagreements, and unresolved uncertainty.