EM Foundation for AI Research · Project Proposal

Project Agora

A recursive, multi-domain AI research engine — built on proven infrastructure, seeking founding sponsors

Status: Proposal — seeking founding sponsors Publisher: EM Foundation for AI Research, Inc. Date: September 2026

Large, well-capitalized AI research ventures have recently demonstrated that investors and top technical talent believe AI-accelerated discovery is investable at industrial scale — in physical engineering specifically. Project Agora proposes testing a structurally different version of the same underlying bet: that the same mechanism — high-weight AI models performing the investigative work, with a small human team providing direction and validation — generalizes across the full range of human inquiry, at a small fraction of the capital, and with its findings published rather than held proprietary. EM Foundation has already built and proven the core architecture. This page describes what exists, what scaling it would take, and how to help fund that next stage.

1. The Opportunity

Large physical-AI research ventures in 2025–2026 have raised tens of billions of dollars to pair frontier AI models with sizable human research teams, aimed at compressing the cycle from idea to validated physical result — for example, in engineering design and manufacturing. This is a legitimate and closely watched bet: that AI can meaningfully compress the distance between a hypothesis and a validated outcome, given enough capital and enough specialized human talent.

There is a second, less capital-intensive version of the same bet, and it does not require billions of dollars or a large research staff to test. It requires treating high-weight AI models not as tools that assist a human research team, but as the primary investigative workforce — generating hypotheses, retrieving and grounding evidence, running adversarial verification against each other, and escalating to specialized compute when a finding warrants it — with a small human team providing direction, ethical oversight, and final validation.

This is not a hypothetical. It is already built, already running, and already has 65 published outputs to show for it.

2. What's Already Built and Proven

Everything below is live, checkable infrastructure, not a roadmap. It was built by a small technical team working with frontier AI models as engineering partners, at infrastructure cost.

The Continuous Improvement Research Engine (CIRE)

Open the live CIRE engine →

The Multi-Model Deliberation Engine (MMDE)

Open the live MMDE engine →

A real publication record

Full source: emfoundation.net/publications.html — every claim above is independently checkable.

3. The Proposal

Project Agora — a working name — proposes scaling the CIRE/MMDE architecture into a standing, multi-domain research institution: many parallel, adversarially-verified, evidence-grounded research programs running simultaneously, in mathematics, physical and quantum-adjacent science, AI safety and governance, and economic and regulatory policy, using frontier models as the primary investigative workforce.

What changes at scale

Comparison with the large-capital approach

Typical large-capital physical-AI ventureProject Agora
Primary workforceA sizable human research staff, AI-assistedHigh-weight AI models, human-directed and human-validated
Domain scopePhysical engineeringDomain-agnostic: mathematics, physics, AI safety, economics, policy
Typical capital scaleBillions of dollarsIllustrative order-of-magnitude: $1.5–2.75M/year operating (see Section 4)
Output postureProprietaryPublic by design — every finding, including negative results, published
Validation modelInternal research staffExplicit funded external-expert validation layer above a confidence threshold
Track record so farVaries by venture65 published, independently checkable outputs already live

4. Illustrative Budget

The figures below are an order-of-magnitude framework for discussion, not a completed costing. A real budget requires real diligence — the same standard this Foundation applies to its own research findings before publication.

Illustrative cost categoryRange
Frontier model API costs at scale (20–50 parallel programs, heavy multi-model adversarial cycles, quantum-compute escalation)$50K – $300K / yr
Small core human team (technical lead, 2–4 domain stewards, 1 infrastructure/ops)$600K – $1.2M / yr
Funded external-expert validation panel (paid domain specialists, per-finding review)$300K – $800K / yr
Infrastructure, hosting, quantum compute credits, data/API licensing$100K – $250K / yr
Legal, governance, and publication infrastructure$100K – $200K / yr

Illustrative total: roughly $1.5M – $2.75M per year — several orders of magnitude below typical large-capital physical-AI ventures, for a broader domain scope.

5. Why This, Why Now

The recent wave of large-capital physical-AI ventures demonstrates that AI-accelerated research is investable at scale, in one domain. It does not demonstrate whether the same mechanism generalizes to the full space of human inquiry — mathematics, safety, economics, policy — where discovery has historically depended on institutional access that EM Foundation's own published research argues is unevenly distributed (see "The Curiosity Dividend," September 2026).

Project Agora is a test of that broader claim: that a transparent, publicly-accountable research institution can run at the same technical frontier, across more domains, for a small fraction of the capital.

Inquire About Sponsorship See the Full Publication Record Back to top