Methodology
Transparent calculation of the Frontier AI Risk Index
The Skynet Barometer calculates a composite risk score (0-100) by aggregating multiple data sources into "Swords" (risk-accelerating factors) and "Shields" (risk-mitigating factors). The methodology is fully transparent and open-source, allowing for community review and alternative weighting schemes.
Risk Accelerators (Swords)
Risk Mitigators (Shields)
Sigmoid Transformation
The sigmoid function (1 / (1 + e-x)) maps the weighted sum to a 0-1 range, which is then scaled to 0-100. This creates smooth transitions and prevents extreme values while maintaining sensitivity to changes in the underlying components.
⚔️Swords (Risk Accelerators)
Capability Momentum (C)
Tracks compute growth, parameter scaling, and benchmark performance improvements. Normalized using z-scores across key metrics from Epoch AI.
Agentic/Autonomy Signals (A)
Measures tool use, multi-step planning, and delegation capabilities from AISI Inspect evaluations and similar frameworks.
Open Access & Diffusion (D)
Availability of powerful models, decreasing costs per token, and open-weights releases with high capabilities.
Real-world Incidents (I)
Frequency and severity of documented AI incidents from AIID, weighted by impact across bio, cyber, and information security domains.
Market Odds (M)
Aggregated probabilities from prediction markets (Polymarket, Kalshi, Manifold) for AGI timelines and capability milestones.
🛡️Shields (Risk Mitigators)
Governance Strength (G)
Strength of safety frameworks from frontier labs (OpenAI Preparedness, DeepMind FSF), independent evaluation access, and regulatory oversight. Includes evaluation gates, red teaming requirements, and safety commitments.
Weight Presets
Swords (Risk Accelerators)
Shields (Risk Mitigators)
| Component | Source | Description | Update Freq. | Reliability |
|---|---|---|---|---|
| Capability Momentum | Epoch AI | Compute trends, parameter scaling, benchmark results | Monthly | High |
| Agentic Signals | UK AISI Inspect | Standardized evaluations for dangerous capabilities | Quarterly | High |
| Open Access & Diffusion | Multiple Sources | Model releases, API pricing, open-weights tracking | Weekly | Medium |
| Real-world Incidents | AI Incident Database | Documented AI-related harms and failures | Continuous | Medium |
| Market Odds | Prediction Markets | Polymarket, Kalshi, Manifold, Metaculus aggregation | Daily | Medium |
| Governance Strength | Policy Tracking | Safety frameworks, regulatory developments | Monthly | Low |
Not Investment Advice
This index is for research and educational purposes only. It should not be used for investment decisions or policy recommendations without additional analysis.
Methodological Limitations
- AGI definitions remain contested and operationalization is challenging
- Data sources may have reporting biases or incomplete coverage
- Weight assignments involve subjective judgments despite transparency
- Prediction markets may not reflect true probabilities due to liquidity constraints
- Incident reporting may be inconsistent across different domains and regions
Interpretation Guidelines
- Focus on trends and relative changes rather than absolute values
- Consider confidence intervals and data quality indicators
- Review methodology updates and version changes regularly
- Supplement with domain-specific expertise and additional sources