DPMT in AI Development: Modeling Progress, Safety, and the Path to AGI

BY NICOLE LAU

Abstract

AI development is a dynamic process with feedback loops (AI assists AI research, capabilities enable applications), tipping points (AGI emergence, recursive self-improvement), and existential stakes. Yet AI strategy often relies on static forecasts—Moore's Law extrapolations, capability benchmarks, timeline predictions—that don't model the complex dynamics of AI progress, safety research, and societal impact. How fast will AI capabilities grow? When might AGI emerge? What ensures alignment and safety? Dynamic Predictive Modeling Theory (DPMT) transforms AI strategy from static prediction to dynamic modeling, enabling researchers and policymakers to understand AI trajectories, identify critical decision points, and navigate toward beneficial outcomes. This paper demonstrates DPMT application to AI development, showing how dynamic modeling reveals the path to safe, transformative AI.

I. AI Development as Dynamic System

AI progress is exponential with feedback loops, potential discontinuities (breakthroughs), and race dynamics (capabilities vs safety). Static models miss these dynamics.

DPMT models AI development as dynamic system with:

Stocks: Compute power, algorithmic efficiency, training data, AI capabilities (by domain), safety research, alignment progress

Flows: Compute growth, algorithmic improvements, capability gains, safety advances, deployment

Feedback Loops: AI assists research → faster progress (positive), economic value → more investment → more progress (positive), safety concerns → slower deployment (negative), alignment difficulty → capability overhang (dangerous)

Delays: Research → capability (months to years), capability → deployment (years), safety research → alignment solutions (uncertain, possibly decades)

Scenarios: Safe AGI, narrow AI plateau, unaligned superintelligence, transformative AI with governance

Attractors: Beneficial AGI, existential catastrophe, perpetual narrow AI

II. Case Study: AGI Timeline and Safety

Current State (2026): GPT-5 level models, human-level performance in many narrow tasks, no AGI yet, safety research lagging capabilities

Question: When might AGI emerge? What's the probability of safe vs unsafe outcomes? What interventions matter most?

Key Variables: Compute (doubling every 6 months), algorithmic efficiency (improving 2-3×/year), AI capabilities (measured by benchmarks), safety research funding, alignment progress, governance readiness

Dynamics:

Positive Loop (AI-Assisted Research): Better AI → Assists AI Research → Faster Progress → Even Better AI (potential recursive self-improvement)

Positive Loop (Economic Value): AI Capabilities → Economic Applications → Revenue → More Investment → More Capabilities

Negative Loop (Safety Slowdown): Dangerous Capabilities → Safety Concerns → Deployment Restrictions → Slower Progress

Negative Loop (Alignment Difficulty): More Capable AI → Harder to Align → Capability-Alignment Gap Widens

Tipping Point: AGI emergence = AI that can perform any intellectual task humans can. Recursive self-improvement possible. Timeline highly uncertain (2030-2070 range, median 2045).

Scenarios:

Safe AGI (30% probability): Alignment research succeeds before AGI. Governance frameworks in place. AGI deployed safely. Transformative benefits (cure diseases, solve climate, abundance). Timeline: AGI by 2045, aligned.

Narrow AI Plateau (25% probability): Fundamental limits hit (data exhaustion, compute limits, algorithmic barriers). AI remains narrow, no AGI. Incremental progress only. Timeline: No AGI by 2100.

Unaligned AGI (20% probability): AGI emerges before alignment solved. Misaligned superintelligence. Existential catastrophe. Timeline: AGI by 2040, unaligned, catastrophic.

Slow Takeoff with Governance (25% probability): Gradual progress to AGI (decades). Time for safety research and governance. International coordination. Safe deployment. Timeline: AGI by 2060, governed, beneficial.

Recommendation: Prioritize alignment research NOW. Current spending: $1B/year on capabilities, $100M/year on safety (100:1 ratio). Should be 10:1 or better 3:1. Increase safety funding to $300M-500M/year. Develop governance frameworks (international AI safety treaty). Slow deployment of dangerous capabilities until alignment solved. Expected outcome: Increases Safe AGI probability from 30% to 50%, reduces Unaligned AGI from 20% to 10%.

Key Insight: AI progress is exponential—capabilities doubling every 1-2 years. Alignment is hard—no clear path to solution yet. Race dynamics are dangerous—competition incentivizes cutting safety corners. Recursive self-improvement could cause fast takeoff—AGI to superintelligence in days/weeks. Timeline is uncertain but plausibly soon (2030s-2040s). This is the most important challenge humanity faces.

III. Key Insights for AI Development

A. Alignment Research Is Urgent

Capabilities are advancing faster than safety. Gap is widening. Once AGI emerges, may be too late to align.

Implication: Massively increase safety research funding. Make alignment a priority, not afterthought.

B. Recursive Self-Improvement Is Possible

Once AI can improve itself, progress could accelerate dramatically (intelligence explosion). Days from AGI to superintelligence.

Implication: Solve alignment BEFORE AGI. No second chances after superintelligence emerges.

C. Race Dynamics Are Dangerous

Competition (US vs China, companies vs companies) incentivizes speed over safety. First to AGI wins, but unaligned AGI kills everyone.

Implication: International cooperation essential. AI safety treaty (like nuclear non-proliferation). Slow down if necessary.

D. Timeline Is Uncertain But Plausibly Soon

Median forecast: 2045. But 25% chance by 2035, 10% chance by 2030. Could be sooner than expected.

Implication: Act with urgency. Don't assume we have decades. Prepare for AGI in 10-20 years.

IV. Conclusion

AI development is a dynamic system with exponential growth, feedback loops, and existential stakes. DPMT enables evidence-based AI strategy by modeling progress dynamics, identifying safety-capability gaps, and designing interventions to increase probability of beneficial outcomes. For AI researchers, policymakers, and humanity, DPMT provides a framework for navigating the most consequential transition in history—the emergence of artificial general intelligence.

The future of intelligence—and perhaps existence—depends on getting AI alignment right. DPMT helps us understand the dynamics and make better decisions.

---

About the Author: Nicole Lau is a theorist working at the intersection of systems thinking, predictive modeling, and cross-disciplinary convergence.