You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
The Theory of Recursive Displacement (TRD) is a falsifiable framework of 33 named mechanisms describing how AI capability deployment erodes the human oversight capacity that alignment solutions depend on. Each mechanism has explicit falsification conditions — testable statements of what would prove it wrong.
Two mechanisms are directly relevant to AI safety:
- Competence Insolvency: As AI absorbs cognitive work, human operators lose the domain expertise needed to detect misalignment, evaluate AI outputs, and intervene during failures. Published December 2025, independently corroborated by Andrej Karpathy three months later.
- Compute Feudalism: Concentration of AI compute and capability in few actors creates governance bottlenecks where alignment failures become systemic rather than local.
The entire framework — API, MCP tool, essays, video series, dashboard — is built, public, and producing outputs. This was developed with zero external funding by one AI-augmented researcher. This grant sustains and extends the work.
Website: https://recursive.institute
Framework: https://recursive.institute/theory
Goals:
1. Deepen the 8 mechanisms most directly relevant to AI safety oversight (Competence Insolvency, Compute Feudalism, Recursive Dependency Spiral, Oversight Decay, Capability Concentration, Skill Atrophy Cascade, Governance Lag, Alignment Assumption Failure) with quantitative indicators and thresholds.
2. Continue to build the model mapping mechanisms and their interactions — which mechanisms accelerate or inhibit others under different AI deployment scenarios.
3. Increase service access to the public API so that other safety researchers can use TRD in their own work.
How: I run an adversarial multi-agent AI pipeline that stress-tests mechanism candidates against falsification conditions. Most candidates fail — that's the quality filter working. The infrastructure exists and is operational. What's needed is runway to run it full-time for 6 months instead of when I can justify the compute costs.
$25,000 — Researcher stipend (6 months full-time, South San Francisco cost of living, supporting family — well below market rate but sustainable)
$10,000 — Additional compute for running the adversarial multi-agent research pipeline. Each mechanism candidate requires significant compute to stress-test against falsification conditions. As the framework matures, most candidates correctly fail validation, raising compute cost per accepted mechanism.
$5,000 — Server infrastructure, hosting costs, LLM API costs for the research pipeline, domain and CDN.
Total: $40,000. No institutional overhead. No administrative staff. No travel. Every dollar funds research output.
Solo researcher: Tyler Maddox, based in South San Francisco.
What I have:
1. A producing research program. 33 mechanisms with falsification conditions, public API, MCP tool server (any LLM can query the framework), 30+ published essays, video series, research dashboard — built and shipped, not proposed.
2. A validated prediction. Competence Insolvency (published December 2025) was independently corroborated by Andrej Karpathy (March 2026). The framework generates predictions that hold up against independent observation.
3. Extreme efficiency. I operate as a one-person, AI-augmented research institute. recursive.institute produces output that would require a small team at a traditional institution.
4. I have health constraints that prevent traditional employment. I built this framework because the problem matters. Zero external funding has come from outside sources to date.
All work is public and documented at recursive.institute. The framework is designed to be extensible — if something happens to me, anyone can pick it up from the API docs and published methods.
Most likely failure modes:
1. Regrantors judge the safety connection as too indirect. Outcome: No funding, research continues at reduced pace on personal time. The framework does not stall — it slows.
2. Health constraints reduce output below deliverable targets. Outcome: Partial delivery. The 6-month timeline already accounts for variable capacity, and deliverables are scoped conservatively. All partial work would still be published.
3. Mechanisms fail validation at higher rate than expected. Outcome: Fewer new mechanisms but stronger framework integrity. Failed candidates are the falsification conditions doing their job — the framework becomes more trustworthy, not less.
4. No one uses the tools. Outcome: The API, MCP tool, and essays remain public infrastructure. Usage may lag publication. This is normal for foundational work.
The framework itself cannot "fail" in a catastrophic sense because every mechanism carries explicit falsification conditions. If a mechanism is wrong, that's a successful test of the framework's epistemic standards, not a research failure.
$0. Zero external funding to date.
Current applications:
- EA Long-Term Future Fund: $55K application under evaluation (submitted April 11, 2026)
- Emergent Ventures: $50K — rejected
- BlueDot Impact Rapid Grant: $10K — rejected
All work to date has been self-funded from personal savings.
There are no bids on this project.