AGI Takeoff Speed Simulator

An interactive model that lets users tune self-improvement parameters and visualize domain-specific capability growth curves over time

Difficulty: weekend | Stack: Python, Streamlit, NumPy, Matplotlib, Pandas

Who this is for

AI safety researchers, students, and curious technologists who want to viscerally explore how assumptions about compounding rates, domain bottlenecks, and experiment cycle times change AGI timeline projections

Build steps

1. Define a domain taxonomy (e.g., hallucination reduction, math reasoning, coding, low-resource language) each with a configurable ‘data ceiling’ and ‘experiment cycle time’ parameter
2. Implement a discrete-time simulation where each tick represents one training run; capability per domain grows as a function of current level, available data, and a user-tunable self-improvement coefficient
3. Add a ‘rate-limiting step’ mechanic: domains with long real-world feedback loops (e.g., robotics, low-resource NLP) progress at a fraction of the speed of domains with automated eval pipelines
4. Build a Streamlit dashboard with sliders for key parameters (takeoff coefficient, domain data ceilings, experiment parallelism) and render per-domain capability timelines side-by-side
5. Add a ‘fast takeoff vs slow climb’ comparison toggle that overlays the classic exponential curve against the domain-bottlenecked simulation output

Risks

• The simulation may feel toy-like without careful parameter calibration — spend time grounding default values in published scaling-law research so outputs feel plausible rather than arbitrary
• Streamlit’s reactivity can lag with large NumPy simulations; if sliders feel sluggish, pre-compute curve families and cache them on slider change
• Framing risk: users may treat the simulation as a predictive tool rather than an intuition pump — add prominent disclaimers and explain model assumptions in the UI