How does compute become capability?
Quantify how effective-compute growth translates into genuinely new capabilities.
Section 7.1 · Research agenda
Plenty that needs to be done
Open research questions the report leaves for the field, grouped by theme.
Scaling & capabilities 2
Better benchmark extrapolation
Use benchmark stitching to forecast capabilities across heterogeneous models.
Paradigms & algorithms 2
Is the neural paradigm enough?
Map where today's approach might hit natural limits, and what could replace it.
Predicting paradigm shifts
Find early indicators of breakthroughs that scaling alone won't produce.
Recursive improvement 2
When do feedback loops kick in?
Track quantitative indicators of AI research automation and recursive improvement.
Exponential or hyperbolic?
Distinguish S-shaped from singular growth as frictions and feedback compete.
Multi-agent & forecasting 3
Emergence in agent collectives
Understand how group agency emerges in complex multi-agent systems.
Forecasting AI progress
Build models that produce uncertainty estimates and ensemble forecasts.
Avoid solipsistic superintelligence
Design training and evaluation so cooperative, non-solipsistic intelligence is favoured.