Mapping AI

Mapping AI is an open-source map of the U.S. AI policy landscape, covering more than 3,000 entities and relationships, including collaborations, funding links, regulatory positions, AGI timelines, risk assessments, and threat models. Since launching in May 2026, it has drawn millions of page views, been featured by NPR Marketplace, and received requests for data access from researchers, think tanks, safety organizations, and forecasters.

The immediate bottleneck is capacity for maintenance and scaling. The people, organizations, and positions in the database change almost weekly, and the founders and volunteers cannot keep it current by hand. This grant would fund human-reviewed automated verification pipelines, source monitoring, new-entry seeding, and the engineering needed to maintain the public site and prepare it for a long-term organizational home. 

Longer version + more context:

Mapping AI is an open stakeholder-mapping tool for the U.S. AI policy landscape. The database includes 3000+ entities and relationships between them, covering funding streams and collaborations, alongside measures such as each actor's regulatory stance, AGI timelines, levels of x-risk, and threat-model focus. A BlueDot rapid grant helped support our launch in May.

With support from the first grant, we hosted our first community workshop (a “Mapping Party”) in San Francisco, created a community Discord, launched publicly with a thread on X that drew millions of page views, and were featured on NPR Marketplace shortly after. In June we worked with a volunteer to add a crowdsourced verification site where contributors can check individual claims. 

After the launch, we received inbound from dozens of people and orgs, like academics, think tanks, AI safety groups, civil society orgs, and forecasters, several asking for API or direct data access. Many of our new priorities to scale the tool have come out of the feedback we received. What those organizations tell us is that the map is useful as long as the data stays rich and current (which may also involve future design choices that narrow its scope). However, data about a landscape that changes weekly cannot sustainably be maintained by hand, by our volunteers or a host org’s FTEs.

As the tool continues to grow, the co-leads don't have the full-time capacity to maintain it long-term. We have worked on this nearly full-time since February alongside our day jobs, and we're hoping to free up more time to think deeply about the policy and governance issues the tool surfaces (see https://civiccompact.org/). We also believe the tool is more effective when, in addition to being a strategic awareness resource used and built by the public, it can be hosted by an organization with capacity to maintain and scale it. A non-negotiable is that the tool remains fully open-source and publicly accessible. This grant funds the costs in building an automated data seeding/verification layer and additional features that would allow the tool to be easily maintained by any host organization: agents that review crowdsourced submissions, read new sources, check claims against the database, recheck old entries on a schedule, and seed new organizations, people, and policy issues as they emerge. We would build the verification scripts and features around the adopting organization's needs and timeline, and run the system alongside them until the handover.

Links

• Main site: https://mapping-ai.org

• X launch thread: https://x.com/mapping_ai/status/2051334980144710112?s=20

• NPR Marketplace feature: https://www.marketplace.org/story/2026/05/19/can-an-ai-map-help-people-track-and-participate-in-ai-policy

• Map: https://mapping-ai.org/map

• Methodology: https://mapping-ai.org/methodology

• Initial research and analysis: https://mapping-ai.org/insights

• Repository: https://github.com/MappingAI/mapping-ai

The minimum, round-cap, and ideal budgets correspond to three operating levels. Detailed editable cost model: https://docs.google.com/spreadsheets/d/18xOZKJ4eWCbq_Nqbo_JVswuLymcyBAnSjnNjRVgD4aY/edit?usp=sharing

The same per-unit verification assumptions apply at each level. Each budget changes the volume of work and the amount of human review, engineering, and operating capacity.

Per-unit model, search, and compute assumptions

• Donated local GLM-5.2: $0.056 per entity-equivalent (EE) in search cash cost, plus 0.189-0.915 H200 GPU-hours per EE when compute is donated.

• Sail GLM-5.2 flex: $0.35-$0.50 per EE for the default hosted open-weight route.

• Frontier batch: $0.90-$1.10 per EE for uncertain records and non-urgent batch work.

• Blended frontier escalation: $1.20-$1.50 per EE for a higher-assurance mixed workload.

• Opus class model-first: about $5.00 per EE, reserved for benchmark adjudication and contested cases.

A comprehensive rich-record verification costs on average around $0.294 in the measured Sail-hosted run, including web search. We plan at $0.35-$0.50 per EE for routine Sail verification to cover harder records, retries, contradictory evidence, and pricing drift. Every workload total below includes a 15% rerun and adjudication allowance.

At $10,000: minimum maintenance

• $4,000 for about 100 hours of data review and source checking.

• $2,000 for limited engineering maintenance and incident response.

• $2,500 for model inference and web search.

• $1,500 for hosting, database, storage, monitoring, and backups.

This covers about 2,408 EE: one complete re-verification pass over the current 2,041 records, three months of weekly correction review at the lean operating volume, and the rerun allowance. Sail-hosted model and search use is estimated at $843-$1,204 for this workload. The $2,500 allocation also covers harder records, retries, selective escalation, and provider variance.

At this funding level, engineering capacity remains limited and some work continues on a volunteer basis. We would cap API spending and keep part of verification semi-manual, which means fewer recheck passes and slower coverage of new policy issues.

At $50,000: the Launch Round cap

• $25,000 for data and research maintenance, approximately 625 hours over 12 months.

• $11,000 for engineering maintenance, approximately 180 hours.

• $6,000 for model inference and web search.

• $4,000 for production services and monitoring.

• $2,000 for benchmark maintenance and adjudication.

• $2,000 as a maintenance reserve for growth, failed jobs, or incident recovery.

This funds 12 months of standard maintenance: weekly review of submissions and corrections, monthly additions, four full re-verification passes, and benchmark upkeep. The plan contains about 11,350 EE, including the current database, rolling additions, targeted updates, and the 15% rerun allowance.

At this volume, Sail-hosted model and search spending is estimated at $3,973-$5,675. The $6,000 allocation leaves about $325 of headroom at the high estimate. A frontier-batch route for the full workload would cost about $10,215-$12,485, so frontier models are reserved for uncertain or contested records. A person reviews every accepted change before it reaches the public database.

At $150,000: the ideal expansion plan

• $55,000 for data stewardship and research editing, supporting roughly 0.5 FTE across the year.

• $35,000 for engineering and product maintenance, supporting roughly 0.25-0.3 FTE.

• $15,000 for expanded source coverage and productionizing the existing claims and evidence layer.

• $12,000 for model inference, search, and frontier-model escalation on contested records.

• $10,000 for infrastructure, security, monitoring, backup retention, and higher usage.

• $8,000 for contributor support, documentation, and regular public verification sessions.

• $5,000 for an independent annual benchmark and adjudication.

• $5,000 for a security and accessibility review.

• $5,000 as a reserve for dataset growth, provider changes, or incident recovery.

This level supports the active-expansion workload of about 13,320 EE while maintaining quarterly verification. It adds roughly 100 people or organizations and 40 policy resources per month. Sail-hosted model and search spending is estimated at $4,662-$6,660. Frontier batch for the entire workload would cost about $11,988-$14,652; the $12,000 line therefore supports routine open-model processing with selective frontier escalation, plus room for retries and provider changes.

The additional funding also supports a stable public API or versioned data export, production use of the existing claims and source tables, stronger search and contribution workflows, and a documented system that a partner organization can operate. These are extensions of the current platform and verification pipeline.