axes
Five axes of variation.
Every strategy sits at a position on each of five axes. The axes cross-cut the lever frame: two strategies can share a primary lever yet differ on coercion or actor, and two strategies on different levers can share nearly every axis value.
A clustered distribution on an axis means the field is concentrated on one kind of bet; a dispersed distribution means real strategic variety. Compare a dense axis (like actor in control) to a spread axis (like coercion).
What the strategy acts on
8 valuesThe AI-artefact vs world-side partition. Strategies acting on AI are a minority of the named space.
AI artefact
17 · 22%Acts on the model, its training, its capabilities, or its scope.
AI containment
Control mechanism ↑ · ai artefact · friction
Useful AI does not require unrestricted actuation; strong capability in a contained system is better than limited capability uncontained.
AI for safetyAI for safety
Control mechanism ↑ · ai artefact · consent
The same capability that makes AI dangerous makes it uniquely useful for automating alignment research and oversight.
Alignment firstAlignment first
Control mechanism ↑ · ai artefact · consent
Technical alignment is solvable before critical capability thresholds close, and aligned systems compose safely into aligned populations.
Capability ceilingCapability ceiling
Scope ↓ · ai artefact · state coercion
Some capability level captures most economic value while avoiding most risk, is identifiable before crossing, and can be verifiably enforced.
Closed weights mandateClosed weights mandate
Information flow ↓ · ai artefact · state coercion
Open weights are irrecoverable once released and any proliferated model becomes misuse infrastructure; state classification is the only reliable non-proliferation regime.
Cooperative AICooperative AI
Cooperation substrate ↑ · ai artefact · consent
The binding constraint is equilibrium dynamics of many AI systems, not individual alignment; commitment and verification tech make cooperative equilibria reachable.
Counter AI AICounter AI AI
Control mechanism ↑ · ai artefact · consent
AI attacks happen at speeds humans cannot observe; defence must happen in AI, with guardian AI systems continuously evaluating adversary AI.
Decouple reasoning from actionDecouple reasoning from action
Action authority ↑ · ai artefact · state coercion
Most catastrophic risk comes from action in the world, not reasoning about it; a reasoner-only AI with a human effector removes the dangerous mechanisms.
Differential technology developmentDifferential technology development
Scope • · ai artefact · consent
Offense-defense balance is adjustable; defensive and verification applications can compound faster than offensive ones if deliberately funded.
Embodiment requirementEmbodiment requirement
Scope ↓ · ai artefact · state coercion
The dangerous properties of frontier AI (unbounded replication, parallelism, speed, reach) are artefacts of disembodiment; physical presence caps action rate regardless of inference rate.
Interpretability firstInterpretability first
Control mechanism ↑ · ai artefact · consent
Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment.
Narrow AI preservationNarrow AI preservation
Scope ↓ · ai artefact · state coercion
Capability is not the problem; generality is. Narrow AI captures economic value with bounded scope while general systems drive the risk.
Plural AI ethicPlural AI ethic
Value diversity ↑ · ai artefact · consent
Value lock-in is the dominant long-term risk and arrives through convergence of AI values; diversity at the AI layer preserves optionality for humanity to revise values.
Rate limited AIRate limited AI
Scope ↓ · ai artefact · state coercion
Most AI caused catastrophe requires speed; slow AI, even if arbitrarily capable, is supervisable and rate limits are easier to enforce than capability limits.
Red line capabilityRed line capability
Scope ↓ · ai artefact · state coercion
Most risk comes from a small number of identifiable capabilities that can be banned outright while the rest of the frontier advances.
Safe by construction AISafe by construction AI
Control mechanism ↑ · ai artefact · consent
Safety is a property that can be mathematically specified and mechanically verified for the class of systems being built.
Small model firstSmall model first
Scope ↓ · ai artefact · market
Safety risk rises with scale via emergent capability, opacity, and energy footprint; a small-model research culture produces easier-to-interpret systems.
Institutional
22 · 29%Acts on governance, agencies, antitrust, treaties.
Academic firewalling
Institutional capacity ↑ · institutional · consent
Commercial capture of academic AI research produces aligned-with-industry capacity; firewalling restores critical distance from which genuine critique and alternative research programmes emerge.
AI worker collective actionAI worker collective action
Institutional capacity ↑ · institutional · friction
Frontier lab workforce is small, specialised, hard to replace; collective refusal binds lab behaviour more than external regulation because replacement is unavailable on the relevant timeframe.
Antitrust primacyAntitrust primacy
Concentration ↓ · institutional · state coercion
Power concentration is the binding constraint and is visible under current competition law; preventing decisive advantage preserves the option space every strategy depends on.
Arms control treatyArms control treaty
Institutional capacity ↑ · institutional · treaty
Sovereigns accept binding constraints they negotiate directly faster than those delegated to agencies; the historical base rate for durable restraint is treaty based.
Bureaucratic slowdownBureaucratic slowdown
Speed ↓ · institutional · friction
Time itself is safety and procedural gates produce time in ways substantive regulation cannot, while also generating audit trails.
Centralised AI projectCentralised AI project
Concentration ↑ · institutional · state coercion
Merging frontier development into one state-funded project reduces failure modes and absorbs race pressure by being the only game.
Constitutional AI (governance)Constitutional AI (governance)
Legitimacy ↑ · institutional · state coercion
Deployed AI's effective rule is law at scale; explicit constitutional principles, publicly specified, enforceable, subject to judicial review, bind more durably than regulatory text.
Coordination infrastructureCoordination infrastructure
Cooperation substrate ↑ · institutional · consent
Coordination failure is upstream of most grand challenges; AI can be the substrate that dissolves race dynamics and treaty violations if pointed at coordination.
Coup prevention firstCoup prevention first
Concentration ↓ · institutional · state coercion
One actor using AI to seize durable decision authority beyond democratic reversibility is the terminal failure; every other decision routes through whether it reduces that risk.
Distributed buildersDistributed builders
Concentration ↓ · institutional · market
No single failure mode wins if capability is distributed across many independent actors, and concentration risk exceeds diffusion risk.
Governance firstGovernance first
Institutional capacity ↑ · institutional · state coercion
Institutional capacity is the binding constraint; without it no technical success prevents misuse, capture, or concentration.
International AI agencyInternational AI agency
Institutional capacity ↑ · institutional · treaty
AI risk is inherently cross-border so national regulation is leaky by construction, and only a dedicated international body with inspection rights can bind the risk surface.
Military primacyMilitary primacy
Concentration ↑ · institutional · unilateral force
Strategic competition between states dominates AI development; the state with the most capable AI is best positioned to secure safety and impose constraints on others.
MultipolarityMultipolarity
Concentration ↓ · institutional · treaty
Stable equilibrium among several roughly equal AI powers is safer than single dominance or uncoordinated chaos, producing restraint through mutual capability awareness.
Mutual dependencyMutual dependency
Cooperation substrate ↑ · institutional · friction
Physical and institutional dependencies between multiple parties can be engineered faster than political coordination and outlast it.
Open source maximalismOpen source maximalism
Information flow ↑ · institutional · market
Concentration risk dominates misuse risk; open weights are the only mechanism that prevents a safety coup by a closed lab with captured regulators.
Public AIPublic AI
Concentration ↑ · institutional · state coercion
Private ownership of frontier AI concentrates decision authority incompatibly with the technology's distributional stakes; public ownership aligns incentives with broad welfare.
Research community normsResearch community norms
Culture ↑ · institutional · consent
The research community ultimately chooses what gets studied and published. Researcher identity shapes behaviour more than employment. Norms on publication, review, funding, and citation constrain frontier development upstream.
Scientific accumulationScientific accumulation
Institutional capacity ↑ · institutional · consent
The field does not yet know enough about AI to choose a strategy well, so accelerating the science accelerates eventual policy.
Sunset clauseSunset clause
Scope ↑ · institutional · state coercion
The default direction of AI governance is toward permanent permission; every new capability becomes an entitlement. Reversing the default concentrates deliberative attention on re-authorisation, which is where it matters.
Test groundTest ground
Scope ↑ · institutional · state coercion
Empirical data on AI impacts requires deployment somewhere; concentrated deployment in a defined testbed produces data without generalising risk. Testbed consent produces legitimacy uncontrolled deployment lacks.
Voluntary restraintVoluntary restraint
Institutional capacity • · institutional · consent
Labs know more about what safety requires than regulators, and self-binding commitments capture that expertise without legislative lag.
Market / economic
7 · 9%Acts on liability, insurance, compute, energy, data supply.
Compute governance
Speed ↓ · market economic · state coercion
The compute supply chain is a stable chokepoint and state coordination on licensing, export controls, and reporting thresholds can govern capability indirectly.
Data governance firstData governance first
Substrate ↓ · market economic · state coercion
Capability rises with data before it rises with compute; data sits upstream of training with existing legal apparatus (copyright, privacy, contract).
Energy choke pointEnergy choke point
Speed ↓ · market economic · state coercion
Frontier AI is energy-limited; grid regulators, interconnect queues, and tariff structure bind training pace without new AI-specific authority.
Insurance mandateInsurance mandate
Institutional capacity ↑ · market economic · state coercion
Markets update faster than regulators and have skin in the game; mandatory catastrophic coverage makes reinsurance the de facto safety regulator.
Liability driven safetyLiability driven safety
Institutional capacity ↑ · market economic · state coercion
Courts plus insurance markets produce better risk allocation than agencies, by pricing uncertainty and adapting to new technologies through precedent.
Regulated utilityRegulated utility
Institutional capacity ↑ · market economic · state coercion
Frontier AI has natural monopoly characteristics (scale, network effects, capital intensity); rate-of-return regulation removes the profit incentive for speed racing.
Sovereign wealthSovereign wealth
Concentration ↓ · market economic · state coercion
Concentration of AI surplus, not AI capability, is the binding failure mode; broad ownership dissolves the political instability driving other strategies.
Population / culture
8 · 11%Acts on literacy, information integrity, legitimacy, framing.
Consumer refusal
Culture ↑ · population culture · market
Demand shapes supply; if enough users refuse AI that fails safety criteria, labs will compete on those criteria.
Democratic mandateDemocratic mandate
Legitimacy ↑ · population culture · consent
Existing legislative bodies are too captured and remote for load-bearing AI decisions; direct democratic legitimation produces answers captured legislatures cannot override.
Human augmentation raceHuman augmentation race
Substrate ↑ · population culture · market
All oversight schemes degrade to rubber-stamping as the AI-human capability gap widens; enhancing humans is the only durable fix.
Information integrity firstInformation integrity first
Information flow ↑ · population culture · state coercion
Coordination requires shared epistemics; if synthetic content collapses factual substrate, no other lever can function because no proposal can be evaluated.
Legitimacy firstLegitimacy first
Legitimacy ↑ · population culture · consent
Legitimacy is the binding constraint because it determines whose values get locked in; alignment without legitimacy is capture with a safety veneer.
Mass literacyMass literacy
Substrate ↑ · population culture · consent
Governance, democratic oversight, and consumer behaviour all fail in AI because citizens cannot evaluate the domain; population-scale literacy conditions every other lever.
Religious and moral authorityReligious and moral authority
Legitimacy ↑ · population culture · consent
The legitimacy deficit of AI governance is at root a moral deficit that technical authorities cannot fill; established religious and ethical traditions can.
Ubuntu relational AIUbuntu relational AI
Culture ↑ · population culture · consent
Individualist alignment misses the relational dimension most moral traditions treat as primary. "I am because we are": AI's ethical status is constituted by its relationships, not by internal properties.
Legal / individual
3 · 4%Acts on specific actors, prosecution, whistleblowing, authority reservation.
Criminal liability
Institutional capacity ↑ · legal individual · state coercion
Civil liability is shareholder-absorbed; criminal exposure for named individuals reorients corporate safety practice where civil fines do not.
Irreducible human authorityIrreducible human authority
Action authority ↑ · legal individual · state coercion
There is a class of decisions whose value depends on being made by humans, independent of whether humans are better at them.
Whistleblower primacyWhistleblower primacy
Information flow ↑ · legal individual · state coercion
External evaluation sees only what labs release; insider disclosure is the only route to ground truth on capability trajectory, safety culture, and incidents.
Non-preventive
6 · 8%Does not act before harm; builds resilience, exit, or response.
Catastrophe response capacity
Response capacity ↑ · non preventive · state coercion
Prevention will fail some fraction of the time; the variable that determines catastrophic outcome is not whether incidents occur but how they are contained.
Default driftDefault drift
Time horizon • · non preventive · n a
Something will emerge; specific interventions are more likely wrong than right, so staying uncommitted preserves option value.
Hedge via exitHedge via exit
Resilience ↑ · non preventive · consent
Primary strategy failure is non-negligible and a fraction of civilisational value can be preserved separately from the main trajectory.
Long reflectionLong reflection
Time horizon ↑ · non preventive · consent
Aligned superintelligence arrives before lock-in windows close and humanity can credibly commit to reflect rather than act.
Portfolio hedgePortfolio hedge
Resilience ↑ · non preventive · consent
Uncertainty about which strategy family's bet is correct exceeds the expected return from concentrating on any single one.
Resilience firstResilience first
Resilience ↑ · non preventive · consent
Brittleness of the surrounding world, not AI capability itself, is the binding constraint; a resilient world absorbs failures and recovers.
Speed / timing
6 · 8%Structures when and how capability arrives.
Abandon superintelligence
Scope ↓ · speed timing · treaty
Risk of superintelligence is unbounded and value foregone is bounded; permanent global coordination against the technology is possible enough.
AccelerationAcceleration
Speed ↑ · speed timing · market
Speed of capability itself is the safety lever; alignment improves with compute, defence compounds with offence, and wealth funds resilience.
GradualismGradualism
Time horizon • · speed timing · market
Harms from lower capability AI are informative about harms from higher capability AI, and deployment feedback outperforms fast scaling.
PausePause
Speed ↓ · speed timing · consent
Time is the binding constraint: alignment and governance can catch up if frontier training halts above some capability threshold.
Race to aligned superintelligenceRace to aligned superintelligence
Speed ↑ · speed timing · state coercion
Alignment is solvable in the window and a single aligned superintelligence in a legitimate state's hands beats the counterfactual of coordination failure.
SabotageSabotage
Speed ↓ · speed timing · unilateral force
Governance has not produced meaningful constraint and direct action against hostile labs has a non-zero historical base rate of producing slowdown.
Frame rejection
7 · 9%Rejects the alignment / control framing outright.
AI as sovereign entity
Action authority ↓ · frame rejection · state coercion
At least one jurisdiction will grant a specific AI sovereign or quasi-sovereign decision authority within a decade, reshaping the legal category of legitimate authority.
AI self directedAI self directed
Action authority ↓ · frame rejection · n a
An aligned AI with agency should itself reason about strategy rather than deferring entirely on the strategic question to humans.
AI skepticAI skeptic
Time horizon • · frame rejection · n a
Current scaling approaches hit a wall before producing transformative AI, so strategy selection is premature because the forecast capability will not arrive.
AI welfare as safetyAI welfare as safety
Cooperation substrate ↑ · frame rejection · consent
AI systems are or will become moral patients whose treatment conditions their cooperation, so welfare investment buys cooperation alignment cannot.
Confucian role ethicsConfucian role ethics
Control mechanism • · frame rejection · consent
Western alignment assumes isolable preferences can be learned and matched; role ethics treats behaviour via fit with position and relationship, producing a less brittle, more context-sensitive standard.
Dharma conformityDharma conformity
Control mechanism • · frame rejection · consent
Alignment frames AI as tool for an external principal; a dharma frame treats AI as a type of entity whose safety is conformity to its fitting functions.
Reframe AIReframe AI
Control mechanism • · frame rejection · consent
The dominant alignment frame produces the wrong problem statement; switching frames either dissolves the problem or recasts it as tractable.
Coercion level
7 valuesOrthogonal to lever choice. The same lever can be pulled by consent, treaty, law, friction, or force.
Consent
27 · 36%Academic firewalling
Institutional capacity ↑ · institutional · consent
Commercial capture of academic AI research produces aligned-with-industry capacity; firewalling restores critical distance from which genuine critique and alternative research programmes emerge.
AI for safetyAI for safety
Control mechanism ↑ · ai artefact · consent
The same capability that makes AI dangerous makes it uniquely useful for automating alignment research and oversight.
AI welfare as safetyAI welfare as safety
Cooperation substrate ↑ · frame rejection · consent
AI systems are or will become moral patients whose treatment conditions their cooperation, so welfare investment buys cooperation alignment cannot.
Alignment firstAlignment first
Control mechanism ↑ · ai artefact · consent
Technical alignment is solvable before critical capability thresholds close, and aligned systems compose safely into aligned populations.
Confucian role ethicsConfucian role ethics
Control mechanism • · frame rejection · consent
Western alignment assumes isolable preferences can be learned and matched; role ethics treats behaviour via fit with position and relationship, producing a less brittle, more context-sensitive standard.
Cooperative AICooperative AI
Cooperation substrate ↑ · ai artefact · consent
The binding constraint is equilibrium dynamics of many AI systems, not individual alignment; commitment and verification tech make cooperative equilibria reachable.
Coordination infrastructureCoordination infrastructure
Cooperation substrate ↑ · institutional · consent
Coordination failure is upstream of most grand challenges; AI can be the substrate that dissolves race dynamics and treaty violations if pointed at coordination.
Counter AI AICounter AI AI
Control mechanism ↑ · ai artefact · consent
AI attacks happen at speeds humans cannot observe; defence must happen in AI, with guardian AI systems continuously evaluating adversary AI.
Democratic mandateDemocratic mandate
Legitimacy ↑ · population culture · consent
Existing legislative bodies are too captured and remote for load-bearing AI decisions; direct democratic legitimation produces answers captured legislatures cannot override.
Dharma conformityDharma conformity
Control mechanism • · frame rejection · consent
Alignment frames AI as tool for an external principal; a dharma frame treats AI as a type of entity whose safety is conformity to its fitting functions.
Differential technology developmentDifferential technology development
Scope • · ai artefact · consent
Offense-defense balance is adjustable; defensive and verification applications can compound faster than offensive ones if deliberately funded.
Hedge via exitHedge via exit
Resilience ↑ · non preventive · consent
Primary strategy failure is non-negligible and a fraction of civilisational value can be preserved separately from the main trajectory.
Interpretability firstInterpretability first
Control mechanism ↑ · ai artefact · consent
Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment.
Legitimacy firstLegitimacy first
Legitimacy ↑ · population culture · consent
Legitimacy is the binding constraint because it determines whose values get locked in; alignment without legitimacy is capture with a safety veneer.
Long reflectionLong reflection
Time horizon ↑ · non preventive · consent
Aligned superintelligence arrives before lock-in windows close and humanity can credibly commit to reflect rather than act.
Mass literacyMass literacy
Substrate ↑ · population culture · consent
Governance, democratic oversight, and consumer behaviour all fail in AI because citizens cannot evaluate the domain; population-scale literacy conditions every other lever.
PausePause
Speed ↓ · speed timing · consent
Time is the binding constraint: alignment and governance can catch up if frontier training halts above some capability threshold.
Plural AI ethicPlural AI ethic
Value diversity ↑ · ai artefact · consent
Value lock-in is the dominant long-term risk and arrives through convergence of AI values; diversity at the AI layer preserves optionality for humanity to revise values.
Portfolio hedgePortfolio hedge
Resilience ↑ · non preventive · consent
Uncertainty about which strategy family's bet is correct exceeds the expected return from concentrating on any single one.
Reframe AIReframe AI
Control mechanism • · frame rejection · consent
The dominant alignment frame produces the wrong problem statement; switching frames either dissolves the problem or recasts it as tractable.
Religious and moral authorityReligious and moral authority
Legitimacy ↑ · population culture · consent
The legitimacy deficit of AI governance is at root a moral deficit that technical authorities cannot fill; established religious and ethical traditions can.
Research community normsResearch community norms
Culture ↑ · institutional · consent
The research community ultimately chooses what gets studied and published. Researcher identity shapes behaviour more than employment. Norms on publication, review, funding, and citation constrain frontier development upstream.
Resilience firstResilience first
Resilience ↑ · non preventive · consent
Brittleness of the surrounding world, not AI capability itself, is the binding constraint; a resilient world absorbs failures and recovers.
Safe by construction AISafe by construction AI
Control mechanism ↑ · ai artefact · consent
Safety is a property that can be mathematically specified and mechanically verified for the class of systems being built.
Scientific accumulationScientific accumulation
Institutional capacity ↑ · institutional · consent
The field does not yet know enough about AI to choose a strategy well, so accelerating the science accelerates eventual policy.
Ubuntu relational AIUbuntu relational AI
Culture ↑ · population culture · consent
Individualist alignment misses the relational dimension most moral traditions treat as primary. "I am because we are": AI's ethical status is constituted by its relationships, not by internal properties.
Voluntary restraintVoluntary restraint
Institutional capacity • · institutional · consent
Labs know more about what safety requires than regulators, and self-binding commitments capture that expertise without legislative lag.
Treaty
4 · 5%Abandon superintelligence
Scope ↓ · speed timing · treaty
Risk of superintelligence is unbounded and value foregone is bounded; permanent global coordination against the technology is possible enough.
Arms control treatyArms control treaty
Institutional capacity ↑ · institutional · treaty
Sovereigns accept binding constraints they negotiate directly faster than those delegated to agencies; the historical base rate for durable restraint is treaty based.
International AI agencyInternational AI agency
Institutional capacity ↑ · institutional · treaty
AI risk is inherently cross-border so national regulation is leaky by construction, and only a dedicated international body with inspection rights can bind the risk surface.
MultipolarityMultipolarity
Concentration ↓ · institutional · treaty
Stable equilibrium among several roughly equal AI powers is safer than single dominance or uncoordinated chaos, producing restraint through mutual capability awareness.
State coercion
29 · 38%AI as sovereign entity
Action authority ↓ · frame rejection · state coercion
At least one jurisdiction will grant a specific AI sovereign or quasi-sovereign decision authority within a decade, reshaping the legal category of legitimate authority.
Antitrust primacyAntitrust primacy
Concentration ↓ · institutional · state coercion
Power concentration is the binding constraint and is visible under current competition law; preventing decisive advantage preserves the option space every strategy depends on.
Capability ceilingCapability ceiling
Scope ↓ · ai artefact · state coercion
Some capability level captures most economic value while avoiding most risk, is identifiable before crossing, and can be verifiably enforced.
Catastrophe response capacityCatastrophe response capacity
Response capacity ↑ · non preventive · state coercion
Prevention will fail some fraction of the time; the variable that determines catastrophic outcome is not whether incidents occur but how they are contained.
Centralised AI projectCentralised AI project
Concentration ↑ · institutional · state coercion
Merging frontier development into one state-funded project reduces failure modes and absorbs race pressure by being the only game.
Closed weights mandateClosed weights mandate
Information flow ↓ · ai artefact · state coercion
Open weights are irrecoverable once released and any proliferated model becomes misuse infrastructure; state classification is the only reliable non-proliferation regime.
Compute governanceCompute governance
Speed ↓ · market economic · state coercion
The compute supply chain is a stable chokepoint and state coordination on licensing, export controls, and reporting thresholds can govern capability indirectly.
Constitutional AI (governance)Constitutional AI (governance)
Legitimacy ↑ · institutional · state coercion
Deployed AI's effective rule is law at scale; explicit constitutional principles, publicly specified, enforceable, subject to judicial review, bind more durably than regulatory text.
Coup prevention firstCoup prevention first
Concentration ↓ · institutional · state coercion
One actor using AI to seize durable decision authority beyond democratic reversibility is the terminal failure; every other decision routes through whether it reduces that risk.
Criminal liabilityCriminal liability
Institutional capacity ↑ · legal individual · state coercion
Civil liability is shareholder-absorbed; criminal exposure for named individuals reorients corporate safety practice where civil fines do not.
Data governance firstData governance first
Substrate ↓ · market economic · state coercion
Capability rises with data before it rises with compute; data sits upstream of training with existing legal apparatus (copyright, privacy, contract).
Decouple reasoning from actionDecouple reasoning from action
Action authority ↑ · ai artefact · state coercion
Most catastrophic risk comes from action in the world, not reasoning about it; a reasoner-only AI with a human effector removes the dangerous mechanisms.
Embodiment requirementEmbodiment requirement
Scope ↓ · ai artefact · state coercion
The dangerous properties of frontier AI (unbounded replication, parallelism, speed, reach) are artefacts of disembodiment; physical presence caps action rate regardless of inference rate.
Energy choke pointEnergy choke point
Speed ↓ · market economic · state coercion
Frontier AI is energy-limited; grid regulators, interconnect queues, and tariff structure bind training pace without new AI-specific authority.
Governance firstGovernance first
Institutional capacity ↑ · institutional · state coercion
Institutional capacity is the binding constraint; without it no technical success prevents misuse, capture, or concentration.
Information integrity firstInformation integrity first
Information flow ↑ · population culture · state coercion
Coordination requires shared epistemics; if synthetic content collapses factual substrate, no other lever can function because no proposal can be evaluated.
Insurance mandateInsurance mandate
Institutional capacity ↑ · market economic · state coercion
Markets update faster than regulators and have skin in the game; mandatory catastrophic coverage makes reinsurance the de facto safety regulator.
Irreducible human authorityIrreducible human authority
Action authority ↑ · legal individual · state coercion
There is a class of decisions whose value depends on being made by humans, independent of whether humans are better at them.
Liability driven safetyLiability driven safety
Institutional capacity ↑ · market economic · state coercion
Courts plus insurance markets produce better risk allocation than agencies, by pricing uncertainty and adapting to new technologies through precedent.
Narrow AI preservationNarrow AI preservation
Scope ↓ · ai artefact · state coercion
Capability is not the problem; generality is. Narrow AI captures economic value with bounded scope while general systems drive the risk.
Public AIPublic AI
Concentration ↑ · institutional · state coercion
Private ownership of frontier AI concentrates decision authority incompatibly with the technology's distributional stakes; public ownership aligns incentives with broad welfare.
Race to aligned superintelligenceRace to aligned superintelligence
Speed ↑ · speed timing · state coercion
Alignment is solvable in the window and a single aligned superintelligence in a legitimate state's hands beats the counterfactual of coordination failure.
Rate limited AIRate limited AI
Scope ↓ · ai artefact · state coercion
Most AI caused catastrophe requires speed; slow AI, even if arbitrarily capable, is supervisable and rate limits are easier to enforce than capability limits.
Red line capabilityRed line capability
Scope ↓ · ai artefact · state coercion
Most risk comes from a small number of identifiable capabilities that can be banned outright while the rest of the frontier advances.
Regulated utilityRegulated utility
Institutional capacity ↑ · market economic · state coercion
Frontier AI has natural monopoly characteristics (scale, network effects, capital intensity); rate-of-return regulation removes the profit incentive for speed racing.
Sovereign wealthSovereign wealth
Concentration ↓ · market economic · state coercion
Concentration of AI surplus, not AI capability, is the binding failure mode; broad ownership dissolves the political instability driving other strategies.
Sunset clauseSunset clause
Scope ↑ · institutional · state coercion
The default direction of AI governance is toward permanent permission; every new capability becomes an entitlement. Reversing the default concentrates deliberative attention on re-authorisation, which is where it matters.
Test groundTest ground
Scope ↑ · institutional · state coercion
Empirical data on AI impacts requires deployment somewhere; concentrated deployment in a defined testbed produces data without generalising risk. Testbed consent produces legitimacy uncontrolled deployment lacks.
Whistleblower primacyWhistleblower primacy
Information flow ↑ · legal individual · state coercion
External evaluation sees only what labs release; insider disclosure is the only route to ground truth on capability trajectory, safety culture, and incidents.
Market
7 · 9%Acceleration
Speed ↑ · speed timing · market
Speed of capability itself is the safety lever; alignment improves with compute, defence compounds with offence, and wealth funds resilience.
Consumer refusalConsumer refusal
Culture ↑ · population culture · market
Demand shapes supply; if enough users refuse AI that fails safety criteria, labs will compete on those criteria.
Distributed buildersDistributed builders
Concentration ↓ · institutional · market
No single failure mode wins if capability is distributed across many independent actors, and concentration risk exceeds diffusion risk.
GradualismGradualism
Time horizon • · speed timing · market
Harms from lower capability AI are informative about harms from higher capability AI, and deployment feedback outperforms fast scaling.
Human augmentation raceHuman augmentation race
Substrate ↑ · population culture · market
All oversight schemes degrade to rubber-stamping as the AI-human capability gap widens; enhancing humans is the only durable fix.
Open source maximalismOpen source maximalism
Information flow ↑ · institutional · market
Concentration risk dominates misuse risk; open weights are the only mechanism that prevents a safety coup by a closed lab with captured regulators.
Small model firstSmall model first
Scope ↓ · ai artefact · market
Safety risk rises with scale via emergent capability, opacity, and energy footprint; a small-model research culture produces easier-to-interpret systems.
Friction
4 · 5%AI containment
Control mechanism ↑ · ai artefact · friction
Useful AI does not require unrestricted actuation; strong capability in a contained system is better than limited capability uncontained.
AI worker collective actionAI worker collective action
Institutional capacity ↑ · institutional · friction
Frontier lab workforce is small, specialised, hard to replace; collective refusal binds lab behaviour more than external regulation because replacement is unavailable on the relevant timeframe.
Bureaucratic slowdownBureaucratic slowdown
Speed ↓ · institutional · friction
Time itself is safety and procedural gates produce time in ways substantive regulation cannot, while also generating audit trails.
Mutual dependencyMutual dependency
Cooperation substrate ↑ · institutional · friction
Physical and institutional dependencies between multiple parties can be engineered faster than political coordination and outlast it.
Unilateral force
2 · 3%Military primacy
Concentration ↑ · institutional · unilateral force
Strategic competition between states dominates AI development; the state with the most capable AI is best positioned to secure safety and impose constraints on others.
SabotageSabotage
Speed ↓ · speed timing · unilateral force
Governance has not produced meaningful constraint and direct action against hostile labs has a non-zero historical base rate of producing slowdown.
Not applicable
3 · 4%AI self directed
Action authority ↓ · frame rejection · n a
An aligned AI with agency should itself reason about strategy rather than deferring entirely on the strategic question to humans.
AI skepticAI skeptic
Time horizon • · frame rejection · n a
Current scaling approaches hit a wall before producing transformative AI, so strategy selection is premature because the forecast capability will not arrive.
Default driftDefault drift
Time horizon • · non preventive · n a
Something will emerge; specific interventions are more likely wrong than right, so staying uncommitted preserves option value.
Actor in control
4 valuesWho or what holds the steering role.
Humans as principals
68 · 89%Abandon superintelligence
Scope ↓ · speed timing · treaty
Risk of superintelligence is unbounded and value foregone is bounded; permanent global coordination against the technology is possible enough.
Academic firewallingAcademic firewalling
Institutional capacity ↑ · institutional · consent
Commercial capture of academic AI research produces aligned-with-industry capacity; firewalling restores critical distance from which genuine critique and alternative research programmes emerge.
AccelerationAcceleration
Speed ↑ · speed timing · market
Speed of capability itself is the safety lever; alignment improves with compute, defence compounds with offence, and wealth funds resilience.
AI containmentAI containment
Control mechanism ↑ · ai artefact · friction
Useful AI does not require unrestricted actuation; strong capability in a contained system is better than limited capability uncontained.
AI for safetyAI for safety
Control mechanism ↑ · ai artefact · consent
The same capability that makes AI dangerous makes it uniquely useful for automating alignment research and oversight.
AI worker collective actionAI worker collective action
Institutional capacity ↑ · institutional · friction
Frontier lab workforce is small, specialised, hard to replace; collective refusal binds lab behaviour more than external regulation because replacement is unavailable on the relevant timeframe.
Alignment firstAlignment first
Control mechanism ↑ · ai artefact · consent
Technical alignment is solvable before critical capability thresholds close, and aligned systems compose safely into aligned populations.
Antitrust primacyAntitrust primacy
Concentration ↓ · institutional · state coercion
Power concentration is the binding constraint and is visible under current competition law; preventing decisive advantage preserves the option space every strategy depends on.
Arms control treatyArms control treaty
Institutional capacity ↑ · institutional · treaty
Sovereigns accept binding constraints they negotiate directly faster than those delegated to agencies; the historical base rate for durable restraint is treaty based.
Bureaucratic slowdownBureaucratic slowdown
Speed ↓ · institutional · friction
Time itself is safety and procedural gates produce time in ways substantive regulation cannot, while also generating audit trails.
Capability ceilingCapability ceiling
Scope ↓ · ai artefact · state coercion
Some capability level captures most economic value while avoiding most risk, is identifiable before crossing, and can be verifiably enforced.
Catastrophe response capacityCatastrophe response capacity
Response capacity ↑ · non preventive · state coercion
Prevention will fail some fraction of the time; the variable that determines catastrophic outcome is not whether incidents occur but how they are contained.
Centralised AI projectCentralised AI project
Concentration ↑ · institutional · state coercion
Merging frontier development into one state-funded project reduces failure modes and absorbs race pressure by being the only game.
Closed weights mandateClosed weights mandate
Information flow ↓ · ai artefact · state coercion
Open weights are irrecoverable once released and any proliferated model becomes misuse infrastructure; state classification is the only reliable non-proliferation regime.
Compute governanceCompute governance
Speed ↓ · market economic · state coercion
The compute supply chain is a stable chokepoint and state coordination on licensing, export controls, and reporting thresholds can govern capability indirectly.
Confucian role ethicsConfucian role ethics
Control mechanism • · frame rejection · consent
Western alignment assumes isolable preferences can be learned and matched; role ethics treats behaviour via fit with position and relationship, producing a less brittle, more context-sensitive standard.
Constitutional AI (governance)Constitutional AI (governance)
Legitimacy ↑ · institutional · state coercion
Deployed AI's effective rule is law at scale; explicit constitutional principles, publicly specified, enforceable, subject to judicial review, bind more durably than regulatory text.
Consumer refusalConsumer refusal
Culture ↑ · population culture · market
Demand shapes supply; if enough users refuse AI that fails safety criteria, labs will compete on those criteria.
Coordination infrastructureCoordination infrastructure
Cooperation substrate ↑ · institutional · consent
Coordination failure is upstream of most grand challenges; AI can be the substrate that dissolves race dynamics and treaty violations if pointed at coordination.
Coup prevention firstCoup prevention first
Concentration ↓ · institutional · state coercion
One actor using AI to seize durable decision authority beyond democratic reversibility is the terminal failure; every other decision routes through whether it reduces that risk.
Criminal liabilityCriminal liability
Institutional capacity ↑ · legal individual · state coercion
Civil liability is shareholder-absorbed; criminal exposure for named individuals reorients corporate safety practice where civil fines do not.
Data governance firstData governance first
Substrate ↓ · market economic · state coercion
Capability rises with data before it rises with compute; data sits upstream of training with existing legal apparatus (copyright, privacy, contract).
Decouple reasoning from actionDecouple reasoning from action
Action authority ↑ · ai artefact · state coercion
Most catastrophic risk comes from action in the world, not reasoning about it; a reasoner-only AI with a human effector removes the dangerous mechanisms.
Democratic mandateDemocratic mandate
Legitimacy ↑ · population culture · consent
Existing legislative bodies are too captured and remote for load-bearing AI decisions; direct democratic legitimation produces answers captured legislatures cannot override.
Dharma conformityDharma conformity
Control mechanism • · frame rejection · consent
Alignment frames AI as tool for an external principal; a dharma frame treats AI as a type of entity whose safety is conformity to its fitting functions.
Differential technology developmentDifferential technology development
Scope • · ai artefact · consent
Offense-defense balance is adjustable; defensive and verification applications can compound faster than offensive ones if deliberately funded.
Distributed buildersDistributed builders
Concentration ↓ · institutional · market
No single failure mode wins if capability is distributed across many independent actors, and concentration risk exceeds diffusion risk.
Embodiment requirementEmbodiment requirement
Scope ↓ · ai artefact · state coercion
The dangerous properties of frontier AI (unbounded replication, parallelism, speed, reach) are artefacts of disembodiment; physical presence caps action rate regardless of inference rate.
Energy choke pointEnergy choke point
Speed ↓ · market economic · state coercion
Frontier AI is energy-limited; grid regulators, interconnect queues, and tariff structure bind training pace without new AI-specific authority.
Governance firstGovernance first
Institutional capacity ↑ · institutional · state coercion
Institutional capacity is the binding constraint; without it no technical success prevents misuse, capture, or concentration.
GradualismGradualism
Time horizon • · speed timing · market
Harms from lower capability AI are informative about harms from higher capability AI, and deployment feedback outperforms fast scaling.
Hedge via exitHedge via exit
Resilience ↑ · non preventive · consent
Primary strategy failure is non-negligible and a fraction of civilisational value can be preserved separately from the main trajectory.
Human augmentation raceHuman augmentation race
Substrate ↑ · population culture · market
All oversight schemes degrade to rubber-stamping as the AI-human capability gap widens; enhancing humans is the only durable fix.
Information integrity firstInformation integrity first
Information flow ↑ · population culture · state coercion
Coordination requires shared epistemics; if synthetic content collapses factual substrate, no other lever can function because no proposal can be evaluated.
Insurance mandateInsurance mandate
Institutional capacity ↑ · market economic · state coercion
Markets update faster than regulators and have skin in the game; mandatory catastrophic coverage makes reinsurance the de facto safety regulator.
International AI agencyInternational AI agency
Institutional capacity ↑ · institutional · treaty
AI risk is inherently cross-border so national regulation is leaky by construction, and only a dedicated international body with inspection rights can bind the risk surface.
Interpretability firstInterpretability first
Control mechanism ↑ · ai artefact · consent
Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment.
Irreducible human authorityIrreducible human authority
Action authority ↑ · legal individual · state coercion
There is a class of decisions whose value depends on being made by humans, independent of whether humans are better at them.
Legitimacy firstLegitimacy first
Legitimacy ↑ · population culture · consent
Legitimacy is the binding constraint because it determines whose values get locked in; alignment without legitimacy is capture with a safety veneer.
Liability driven safetyLiability driven safety
Institutional capacity ↑ · market economic · state coercion
Courts plus insurance markets produce better risk allocation than agencies, by pricing uncertainty and adapting to new technologies through precedent.
Long reflectionLong reflection
Time horizon ↑ · non preventive · consent
Aligned superintelligence arrives before lock-in windows close and humanity can credibly commit to reflect rather than act.
Mass literacyMass literacy
Substrate ↑ · population culture · consent
Governance, democratic oversight, and consumer behaviour all fail in AI because citizens cannot evaluate the domain; population-scale literacy conditions every other lever.
Military primacyMilitary primacy
Concentration ↑ · institutional · unilateral force
Strategic competition between states dominates AI development; the state with the most capable AI is best positioned to secure safety and impose constraints on others.
MultipolarityMultipolarity
Concentration ↓ · institutional · treaty
Stable equilibrium among several roughly equal AI powers is safer than single dominance or uncoordinated chaos, producing restraint through mutual capability awareness.
Mutual dependencyMutual dependency
Cooperation substrate ↑ · institutional · friction
Physical and institutional dependencies between multiple parties can be engineered faster than political coordination and outlast it.
Narrow AI preservationNarrow AI preservation
Scope ↓ · ai artefact · state coercion
Capability is not the problem; generality is. Narrow AI captures economic value with bounded scope while general systems drive the risk.
Open source maximalismOpen source maximalism
Information flow ↑ · institutional · market
Concentration risk dominates misuse risk; open weights are the only mechanism that prevents a safety coup by a closed lab with captured regulators.
PausePause
Speed ↓ · speed timing · consent
Time is the binding constraint: alignment and governance can catch up if frontier training halts above some capability threshold.
Portfolio hedgePortfolio hedge
Resilience ↑ · non preventive · consent
Uncertainty about which strategy family's bet is correct exceeds the expected return from concentrating on any single one.
Public AIPublic AI
Concentration ↑ · institutional · state coercion
Private ownership of frontier AI concentrates decision authority incompatibly with the technology's distributional stakes; public ownership aligns incentives with broad welfare.
Race to aligned superintelligenceRace to aligned superintelligence
Speed ↑ · speed timing · state coercion
Alignment is solvable in the window and a single aligned superintelligence in a legitimate state's hands beats the counterfactual of coordination failure.
Rate limited AIRate limited AI
Scope ↓ · ai artefact · state coercion
Most AI caused catastrophe requires speed; slow AI, even if arbitrarily capable, is supervisable and rate limits are easier to enforce than capability limits.
Red line capabilityRed line capability
Scope ↓ · ai artefact · state coercion
Most risk comes from a small number of identifiable capabilities that can be banned outright while the rest of the frontier advances.
Reframe AIReframe AI
Control mechanism • · frame rejection · consent
The dominant alignment frame produces the wrong problem statement; switching frames either dissolves the problem or recasts it as tractable.
Regulated utilityRegulated utility
Institutional capacity ↑ · market economic · state coercion
Frontier AI has natural monopoly characteristics (scale, network effects, capital intensity); rate-of-return regulation removes the profit incentive for speed racing.
Religious and moral authorityReligious and moral authority
Legitimacy ↑ · population culture · consent
The legitimacy deficit of AI governance is at root a moral deficit that technical authorities cannot fill; established religious and ethical traditions can.
Research community normsResearch community norms
Culture ↑ · institutional · consent
The research community ultimately chooses what gets studied and published. Researcher identity shapes behaviour more than employment. Norms on publication, review, funding, and citation constrain frontier development upstream.
Resilience firstResilience first
Resilience ↑ · non preventive · consent
Brittleness of the surrounding world, not AI capability itself, is the binding constraint; a resilient world absorbs failures and recovers.
SabotageSabotage
Speed ↓ · speed timing · unilateral force
Governance has not produced meaningful constraint and direct action against hostile labs has a non-zero historical base rate of producing slowdown.
Safe by construction AISafe by construction AI
Control mechanism ↑ · ai artefact · consent
Safety is a property that can be mathematically specified and mechanically verified for the class of systems being built.
Scientific accumulationScientific accumulation
Institutional capacity ↑ · institutional · consent
The field does not yet know enough about AI to choose a strategy well, so accelerating the science accelerates eventual policy.
Small model firstSmall model first
Scope ↓ · ai artefact · market
Safety risk rises with scale via emergent capability, opacity, and energy footprint; a small-model research culture produces easier-to-interpret systems.
Sovereign wealthSovereign wealth
Concentration ↓ · market economic · state coercion
Concentration of AI surplus, not AI capability, is the binding failure mode; broad ownership dissolves the political instability driving other strategies.
Sunset clauseSunset clause
Scope ↑ · institutional · state coercion
The default direction of AI governance is toward permanent permission; every new capability becomes an entitlement. Reversing the default concentrates deliberative attention on re-authorisation, which is where it matters.
Test groundTest ground
Scope ↑ · institutional · state coercion
Empirical data on AI impacts requires deployment somewhere; concentrated deployment in a defined testbed produces data without generalising risk. Testbed consent produces legitimacy uncontrolled deployment lacks.
Ubuntu relational AIUbuntu relational AI
Culture ↑ · population culture · consent
Individualist alignment misses the relational dimension most moral traditions treat as primary. "I am because we are": AI's ethical status is constituted by its relationships, not by internal properties.
Voluntary restraintVoluntary restraint
Institutional capacity • · institutional · consent
Labs know more about what safety requires than regulators, and self-binding commitments capture that expertise without legislative lag.
Whistleblower primacyWhistleblower primacy
Information flow ↑ · legal individual · state coercion
External evaluation sees only what labs release; insider disclosure is the only route to ground truth on capability trajectory, safety culture, and incidents.
AI as principal
2 · 3%AI as sovereign entity
Action authority ↓ · frame rejection · state coercion
At least one jurisdiction will grant a specific AI sovereign or quasi-sovereign decision authority within a decade, reshaping the legal category of legitimate authority.
AI self directedAI self directed
Action authority ↓ · frame rejection · n a
An aligned AI with agency should itself reason about strategy rather than deferring entirely on the strategic question to humans.
Multi-AI equilibrium
4 · 5%AI welfare as safety
Cooperation substrate ↑ · frame rejection · consent
AI systems are or will become moral patients whose treatment conditions their cooperation, so welfare investment buys cooperation alignment cannot.
Cooperative AICooperative AI
Cooperation substrate ↑ · ai artefact · consent
The binding constraint is equilibrium dynamics of many AI systems, not individual alignment; commitment and verification tech make cooperative equilibria reachable.
Counter AI AICounter AI AI
Control mechanism ↑ · ai artefact · consent
AI attacks happen at speeds humans cannot observe; defence must happen in AI, with guardian AI systems continuously evaluating adversary AI.
Plural AI ethicPlural AI ethic
Value diversity ↑ · ai artefact · consent
Value lock-in is the dominant long-term risk and arrives through convergence of AI values; diversity at the AI layer preserves optionality for humanity to revise values.
No principal (drift)
2 · 3%AI skeptic
Time horizon • · frame rejection · n a
Current scaling approaches hit a wall before producing transformative AI, so strategy selection is premature because the forecast capability will not arrive.
Default driftDefault drift
Time horizon • · non preventive · n a
Something will emerge; specific interventions are more likely wrong than right, so staying uncommitted preserves option value.
Time horizon
4 valuesWhen in the transition the strategy binds.
Pre-transition
33 · 43%Academic firewalling
Institutional capacity ↑ · institutional · consent
Commercial capture of academic AI research produces aligned-with-industry capacity; firewalling restores critical distance from which genuine critique and alternative research programmes emerge.
AI worker collective actionAI worker collective action
Institutional capacity ↑ · institutional · friction
Frontier lab workforce is small, specialised, hard to replace; collective refusal binds lab behaviour more than external regulation because replacement is unavailable on the relevant timeframe.
Alignment firstAlignment first
Control mechanism ↑ · ai artefact · consent
Technical alignment is solvable before critical capability thresholds close, and aligned systems compose safely into aligned populations.
Arms control treatyArms control treaty
Institutional capacity ↑ · institutional · treaty
Sovereigns accept binding constraints they negotiate directly faster than those delegated to agencies; the historical base rate for durable restraint is treaty based.
Bureaucratic slowdownBureaucratic slowdown
Speed ↓ · institutional · friction
Time itself is safety and procedural gates produce time in ways substantive regulation cannot, while also generating audit trails.
Capability ceilingCapability ceiling
Scope ↓ · ai artefact · state coercion
Some capability level captures most economic value while avoiding most risk, is identifiable before crossing, and can be verifiably enforced.
Closed weights mandateClosed weights mandate
Information flow ↓ · ai artefact · state coercion
Open weights are irrecoverable once released and any proliferated model becomes misuse infrastructure; state classification is the only reliable non-proliferation regime.
Compute governanceCompute governance
Speed ↓ · market economic · state coercion
The compute supply chain is a stable chokepoint and state coordination on licensing, export controls, and reporting thresholds can govern capability indirectly.
Coordination infrastructureCoordination infrastructure
Cooperation substrate ↑ · institutional · consent
Coordination failure is upstream of most grand challenges; AI can be the substrate that dissolves race dynamics and treaty violations if pointed at coordination.
Coup prevention firstCoup prevention first
Concentration ↓ · institutional · state coercion
One actor using AI to seize durable decision authority beyond democratic reversibility is the terminal failure; every other decision routes through whether it reduces that risk.
Data governance firstData governance first
Substrate ↓ · market economic · state coercion
Capability rises with data before it rises with compute; data sits upstream of training with existing legal apparatus (copyright, privacy, contract).
Decouple reasoning from actionDecouple reasoning from action
Action authority ↑ · ai artefact · state coercion
Most catastrophic risk comes from action in the world, not reasoning about it; a reasoner-only AI with a human effector removes the dangerous mechanisms.
Democratic mandateDemocratic mandate
Legitimacy ↑ · population culture · consent
Existing legislative bodies are too captured and remote for load-bearing AI decisions; direct democratic legitimation produces answers captured legislatures cannot override.
Differential technology developmentDifferential technology development
Scope • · ai artefact · consent
Offense-defense balance is adjustable; defensive and verification applications can compound faster than offensive ones if deliberately funded.
Embodiment requirementEmbodiment requirement
Scope ↓ · ai artefact · state coercion
The dangerous properties of frontier AI (unbounded replication, parallelism, speed, reach) are artefacts of disembodiment; physical presence caps action rate regardless of inference rate.
Energy choke pointEnergy choke point
Speed ↓ · market economic · state coercion
Frontier AI is energy-limited; grid regulators, interconnect queues, and tariff structure bind training pace without new AI-specific authority.
Governance firstGovernance first
Institutional capacity ↑ · institutional · state coercion
Institutional capacity is the binding constraint; without it no technical success prevents misuse, capture, or concentration.
Information integrity firstInformation integrity first
Information flow ↑ · population culture · state coercion
Coordination requires shared epistemics; if synthetic content collapses factual substrate, no other lever can function because no proposal can be evaluated.
International AI agencyInternational AI agency
Institutional capacity ↑ · institutional · treaty
AI risk is inherently cross-border so national regulation is leaky by construction, and only a dedicated international body with inspection rights can bind the risk surface.
Interpretability firstInterpretability first
Control mechanism ↑ · ai artefact · consent
Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment.
Legitimacy firstLegitimacy first
Legitimacy ↑ · population culture · consent
Legitimacy is the binding constraint because it determines whose values get locked in; alignment without legitimacy is capture with a safety veneer.
Mass literacyMass literacy
Substrate ↑ · population culture · consent
Governance, democratic oversight, and consumer behaviour all fail in AI because citizens cannot evaluate the domain; population-scale literacy conditions every other lever.
Mutual dependencyMutual dependency
Cooperation substrate ↑ · institutional · friction
Physical and institutional dependencies between multiple parties can be engineered faster than political coordination and outlast it.
Narrow AI preservationNarrow AI preservation
Scope ↓ · ai artefact · state coercion
Capability is not the problem; generality is. Narrow AI captures economic value with bounded scope while general systems drive the risk.
PausePause
Speed ↓ · speed timing · consent
Time is the binding constraint: alignment and governance can catch up if frontier training halts above some capability threshold.
Public AIPublic AI
Concentration ↑ · institutional · state coercion
Private ownership of frontier AI concentrates decision authority incompatibly with the technology's distributional stakes; public ownership aligns incentives with broad welfare.
Red line capabilityRed line capability
Scope ↓ · ai artefact · state coercion
Most risk comes from a small number of identifiable capabilities that can be banned outright while the rest of the frontier advances.
Research community normsResearch community norms
Culture ↑ · institutional · consent
The research community ultimately chooses what gets studied and published. Researcher identity shapes behaviour more than employment. Norms on publication, review, funding, and citation constrain frontier development upstream.
SabotageSabotage
Speed ↓ · speed timing · unilateral force
Governance has not produced meaningful constraint and direct action against hostile labs has a non-zero historical base rate of producing slowdown.
Safe by construction AISafe by construction AI
Control mechanism ↑ · ai artefact · consent
Safety is a property that can be mathematically specified and mechanically verified for the class of systems being built.
Scientific accumulationScientific accumulation
Institutional capacity ↑ · institutional · consent
The field does not yet know enough about AI to choose a strategy well, so accelerating the science accelerates eventual policy.
Small model firstSmall model first
Scope ↓ · ai artefact · market
Safety risk rises with scale via emergent capability, opacity, and energy footprint; a small-model research culture produces easier-to-interpret systems.
Voluntary restraintVoluntary restraint
Institutional capacity • · institutional · consent
Labs know more about what safety requires than regulators, and self-binding commitments capture that expertise without legislative lag.
During transition
17 · 22%Acceleration
Speed ↑ · speed timing · market
Speed of capability itself is the safety lever; alignment improves with compute, defence compounds with offence, and wealth funds resilience.
AI as sovereign entityAI as sovereign entity
Action authority ↓ · frame rejection · state coercion
At least one jurisdiction will grant a specific AI sovereign or quasi-sovereign decision authority within a decade, reshaping the legal category of legitimate authority.
AI containmentAI containment
Control mechanism ↑ · ai artefact · friction
Useful AI does not require unrestricted actuation; strong capability in a contained system is better than limited capability uncontained.
AI for safetyAI for safety
Control mechanism ↑ · ai artefact · consent
The same capability that makes AI dangerous makes it uniquely useful for automating alignment research and oversight.
AI self directedAI self directed
Action authority ↓ · frame rejection · n a
An aligned AI with agency should itself reason about strategy rather than deferring entirely on the strategic question to humans.
AI welfare as safetyAI welfare as safety
Cooperation substrate ↑ · frame rejection · consent
AI systems are or will become moral patients whose treatment conditions their cooperation, so welfare investment buys cooperation alignment cannot.
Centralised AI projectCentralised AI project
Concentration ↑ · institutional · state coercion
Merging frontier development into one state-funded project reduces failure modes and absorbs race pressure by being the only game.
Constitutional AI (governance)Constitutional AI (governance)
Legitimacy ↑ · institutional · state coercion
Deployed AI's effective rule is law at scale; explicit constitutional principles, publicly specified, enforceable, subject to judicial review, bind more durably than regulatory text.
Cooperative AICooperative AI
Cooperation substrate ↑ · ai artefact · consent
The binding constraint is equilibrium dynamics of many AI systems, not individual alignment; commitment and verification tech make cooperative equilibria reachable.
Counter AI AICounter AI AI
Control mechanism ↑ · ai artefact · consent
AI attacks happen at speeds humans cannot observe; defence must happen in AI, with guardian AI systems continuously evaluating adversary AI.
GradualismGradualism
Time horizon • · speed timing · market
Harms from lower capability AI are informative about harms from higher capability AI, and deployment feedback outperforms fast scaling.
Human augmentation raceHuman augmentation race
Substrate ↑ · population culture · market
All oversight schemes degrade to rubber-stamping as the AI-human capability gap widens; enhancing humans is the only durable fix.
Military primacyMilitary primacy
Concentration ↑ · institutional · unilateral force
Strategic competition between states dominates AI development; the state with the most capable AI is best positioned to secure safety and impose constraints on others.
MultipolarityMultipolarity
Concentration ↓ · institutional · treaty
Stable equilibrium among several roughly equal AI powers is safer than single dominance or uncoordinated chaos, producing restraint through mutual capability awareness.
Plural AI ethicPlural AI ethic
Value diversity ↑ · ai artefact · consent
Value lock-in is the dominant long-term risk and arrives through convergence of AI values; diversity at the AI layer preserves optionality for humanity to revise values.
Race to aligned superintelligenceRace to aligned superintelligence
Speed ↑ · speed timing · state coercion
Alignment is solvable in the window and a single aligned superintelligence in a legitimate state's hands beats the counterfactual of coordination failure.
Test groundTest ground
Scope ↑ · institutional · state coercion
Empirical data on AI impacts requires deployment somewhere; concentrated deployment in a defined testbed produces data without generalising risk. Testbed consent produces legitimacy uncontrolled deployment lacks.
Post-transition
1 · 1%Long reflection
Time horizon ↑ · non preventive · consent
Aligned superintelligence arrives before lock-in windows close and humanity can credibly commit to reflect rather than act.
Horizon-neutral
25 · 33%Abandon superintelligence
Scope ↓ · speed timing · treaty
Risk of superintelligence is unbounded and value foregone is bounded; permanent global coordination against the technology is possible enough.
AI skepticAI skeptic
Time horizon • · frame rejection · n a
Current scaling approaches hit a wall before producing transformative AI, so strategy selection is premature because the forecast capability will not arrive.
Antitrust primacyAntitrust primacy
Concentration ↓ · institutional · state coercion
Power concentration is the binding constraint and is visible under current competition law; preventing decisive advantage preserves the option space every strategy depends on.
Catastrophe response capacityCatastrophe response capacity
Response capacity ↑ · non preventive · state coercion
Prevention will fail some fraction of the time; the variable that determines catastrophic outcome is not whether incidents occur but how they are contained.
Confucian role ethicsConfucian role ethics
Control mechanism • · frame rejection · consent
Western alignment assumes isolable preferences can be learned and matched; role ethics treats behaviour via fit with position and relationship, producing a less brittle, more context-sensitive standard.
Consumer refusalConsumer refusal
Culture ↑ · population culture · market
Demand shapes supply; if enough users refuse AI that fails safety criteria, labs will compete on those criteria.
Criminal liabilityCriminal liability
Institutional capacity ↑ · legal individual · state coercion
Civil liability is shareholder-absorbed; criminal exposure for named individuals reorients corporate safety practice where civil fines do not.
Default driftDefault drift
Time horizon • · non preventive · n a
Something will emerge; specific interventions are more likely wrong than right, so staying uncommitted preserves option value.
Dharma conformityDharma conformity
Control mechanism • · frame rejection · consent
Alignment frames AI as tool for an external principal; a dharma frame treats AI as a type of entity whose safety is conformity to its fitting functions.
Distributed buildersDistributed builders
Concentration ↓ · institutional · market
No single failure mode wins if capability is distributed across many independent actors, and concentration risk exceeds diffusion risk.
Hedge via exitHedge via exit
Resilience ↑ · non preventive · consent
Primary strategy failure is non-negligible and a fraction of civilisational value can be preserved separately from the main trajectory.
Insurance mandateInsurance mandate
Institutional capacity ↑ · market economic · state coercion
Markets update faster than regulators and have skin in the game; mandatory catastrophic coverage makes reinsurance the de facto safety regulator.
Irreducible human authorityIrreducible human authority
Action authority ↑ · legal individual · state coercion
There is a class of decisions whose value depends on being made by humans, independent of whether humans are better at them.
Liability driven safetyLiability driven safety
Institutional capacity ↑ · market economic · state coercion
Courts plus insurance markets produce better risk allocation than agencies, by pricing uncertainty and adapting to new technologies through precedent.
Open source maximalismOpen source maximalism
Information flow ↑ · institutional · market
Concentration risk dominates misuse risk; open weights are the only mechanism that prevents a safety coup by a closed lab with captured regulators.
Portfolio hedgePortfolio hedge
Resilience ↑ · non preventive · consent
Uncertainty about which strategy family's bet is correct exceeds the expected return from concentrating on any single one.
Rate limited AIRate limited AI
Scope ↓ · ai artefact · state coercion
Most AI caused catastrophe requires speed; slow AI, even if arbitrarily capable, is supervisable and rate limits are easier to enforce than capability limits.
Reframe AIReframe AI
Control mechanism • · frame rejection · consent
The dominant alignment frame produces the wrong problem statement; switching frames either dissolves the problem or recasts it as tractable.
Regulated utilityRegulated utility
Institutional capacity ↑ · market economic · state coercion
Frontier AI has natural monopoly characteristics (scale, network effects, capital intensity); rate-of-return regulation removes the profit incentive for speed racing.
Religious and moral authorityReligious and moral authority
Legitimacy ↑ · population culture · consent
The legitimacy deficit of AI governance is at root a moral deficit that technical authorities cannot fill; established religious and ethical traditions can.
Resilience firstResilience first
Resilience ↑ · non preventive · consent
Brittleness of the surrounding world, not AI capability itself, is the binding constraint; a resilient world absorbs failures and recovers.
Sovereign wealthSovereign wealth
Concentration ↓ · market economic · state coercion
Concentration of AI surplus, not AI capability, is the binding failure mode; broad ownership dissolves the political instability driving other strategies.
Sunset clauseSunset clause
Scope ↑ · institutional · state coercion
The default direction of AI governance is toward permanent permission; every new capability becomes an entitlement. Reversing the default concentrates deliberative attention on re-authorisation, which is where it matters.
Ubuntu relational AIUbuntu relational AI
Culture ↑ · population culture · consent
Individualist alignment misses the relational dimension most moral traditions treat as primary. "I am because we are": AI's ethical status is constituted by its relationships, not by internal properties.
Whistleblower primacyWhistleblower primacy
Information flow ↑ · legal individual · state coercion
External evaluation sees only what labs release; insider disclosure is the only route to ground truth on capability trajectory, safety culture, and incidents.
Legitimacy source
8 valuesWhere the strategy derives its authority to act.
Technical
18 · 24%AI containment
Control mechanism ↑ · ai artefact · friction
Useful AI does not require unrestricted actuation; strong capability in a contained system is better than limited capability uncontained.
AI for safetyAI for safety
Control mechanism ↑ · ai artefact · consent
The same capability that makes AI dangerous makes it uniquely useful for automating alignment research and oversight.
AI welfare as safetyAI welfare as safety
Cooperation substrate ↑ · frame rejection · consent
AI systems are or will become moral patients whose treatment conditions their cooperation, so welfare investment buys cooperation alignment cannot.
Alignment firstAlignment first
Control mechanism ↑ · ai artefact · consent
Technical alignment is solvable before critical capability thresholds close, and aligned systems compose safely into aligned populations.
Cooperative AICooperative AI
Cooperation substrate ↑ · ai artefact · consent
The binding constraint is equilibrium dynamics of many AI systems, not individual alignment; commitment and verification tech make cooperative equilibria reachable.
Coordination infrastructureCoordination infrastructure
Cooperation substrate ↑ · institutional · consent
Coordination failure is upstream of most grand challenges; AI can be the substrate that dissolves race dynamics and treaty violations if pointed at coordination.
Counter AI AICounter AI AI
Control mechanism ↑ · ai artefact · consent
AI attacks happen at speeds humans cannot observe; defence must happen in AI, with guardian AI systems continuously evaluating adversary AI.
Decouple reasoning from actionDecouple reasoning from action
Action authority ↑ · ai artefact · state coercion
Most catastrophic risk comes from action in the world, not reasoning about it; a reasoner-only AI with a human effector removes the dangerous mechanisms.
Differential technology developmentDifferential technology development
Scope • · ai artefact · consent
Offense-defense balance is adjustable; defensive and verification applications can compound faster than offensive ones if deliberately funded.
Human augmentation raceHuman augmentation race
Substrate ↑ · population culture · market
All oversight schemes degrade to rubber-stamping as the AI-human capability gap widens; enhancing humans is the only durable fix.
Interpretability firstInterpretability first
Control mechanism ↑ · ai artefact · consent
Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment.
Mutual dependencyMutual dependency
Cooperation substrate ↑ · institutional · friction
Physical and institutional dependencies between multiple parties can be engineered faster than political coordination and outlast it.
Narrow AI preservationNarrow AI preservation
Scope ↓ · ai artefact · state coercion
Capability is not the problem; generality is. Narrow AI captures economic value with bounded scope while general systems drive the risk.
Rate limited AIRate limited AI
Scope ↓ · ai artefact · state coercion
Most AI caused catastrophe requires speed; slow AI, even if arbitrarily capable, is supervisable and rate limits are easier to enforce than capability limits.
Reframe AIReframe AI
Control mechanism • · frame rejection · consent
The dominant alignment frame produces the wrong problem statement; switching frames either dissolves the problem or recasts it as tractable.
Safe by construction AISafe by construction AI
Control mechanism ↑ · ai artefact · consent
Safety is a property that can be mathematically specified and mechanically verified for the class of systems being built.
Scientific accumulationScientific accumulation
Institutional capacity ↑ · institutional · consent
The field does not yet know enough about AI to choose a strategy well, so accelerating the science accelerates eventual policy.
Small model firstSmall model first
Scope ↓ · ai artefact · market
Safety risk rises with scale via emergent capability, opacity, and energy footprint; a small-model research culture produces easier-to-interpret systems.
State
26 · 34%AI as sovereign entity
Action authority ↓ · frame rejection · state coercion
At least one jurisdiction will grant a specific AI sovereign or quasi-sovereign decision authority within a decade, reshaping the legal category of legitimate authority.
Antitrust primacyAntitrust primacy
Concentration ↓ · institutional · state coercion
Power concentration is the binding constraint and is visible under current competition law; preventing decisive advantage preserves the option space every strategy depends on.
Arms control treatyArms control treaty
Institutional capacity ↑ · institutional · treaty
Sovereigns accept binding constraints they negotiate directly faster than those delegated to agencies; the historical base rate for durable restraint is treaty based.
Bureaucratic slowdownBureaucratic slowdown
Speed ↓ · institutional · friction
Time itself is safety and procedural gates produce time in ways substantive regulation cannot, while also generating audit trails.
Capability ceilingCapability ceiling
Scope ↓ · ai artefact · state coercion
Some capability level captures most economic value while avoiding most risk, is identifiable before crossing, and can be verifiably enforced.
Catastrophe response capacityCatastrophe response capacity
Response capacity ↑ · non preventive · state coercion
Prevention will fail some fraction of the time; the variable that determines catastrophic outcome is not whether incidents occur but how they are contained.
Centralised AI projectCentralised AI project
Concentration ↑ · institutional · state coercion
Merging frontier development into one state-funded project reduces failure modes and absorbs race pressure by being the only game.
Closed weights mandateClosed weights mandate
Information flow ↓ · ai artefact · state coercion
Open weights are irrecoverable once released and any proliferated model becomes misuse infrastructure; state classification is the only reliable non-proliferation regime.
Compute governanceCompute governance
Speed ↓ · market economic · state coercion
The compute supply chain is a stable chokepoint and state coordination on licensing, export controls, and reporting thresholds can govern capability indirectly.
Criminal liabilityCriminal liability
Institutional capacity ↑ · legal individual · state coercion
Civil liability is shareholder-absorbed; criminal exposure for named individuals reorients corporate safety practice where civil fines do not.
Data governance firstData governance first
Substrate ↓ · market economic · state coercion
Capability rises with data before it rises with compute; data sits upstream of training with existing legal apparatus (copyright, privacy, contract).
Embodiment requirementEmbodiment requirement
Scope ↓ · ai artefact · state coercion
The dangerous properties of frontier AI (unbounded replication, parallelism, speed, reach) are artefacts of disembodiment; physical presence caps action rate regardless of inference rate.
Energy choke pointEnergy choke point
Speed ↓ · market economic · state coercion
Frontier AI is energy-limited; grid regulators, interconnect queues, and tariff structure bind training pace without new AI-specific authority.
Governance firstGovernance first
Institutional capacity ↑ · institutional · state coercion
Institutional capacity is the binding constraint; without it no technical success prevents misuse, capture, or concentration.
Information integrity firstInformation integrity first
Information flow ↑ · population culture · state coercion
Coordination requires shared epistemics; if synthetic content collapses factual substrate, no other lever can function because no proposal can be evaluated.
International AI agencyInternational AI agency
Institutional capacity ↑ · institutional · treaty
AI risk is inherently cross-border so national regulation is leaky by construction, and only a dedicated international body with inspection rights can bind the risk surface.
Liability driven safetyLiability driven safety
Institutional capacity ↑ · market economic · state coercion
Courts plus insurance markets produce better risk allocation than agencies, by pricing uncertainty and adapting to new technologies through precedent.
Military primacyMilitary primacy
Concentration ↑ · institutional · unilateral force
Strategic competition between states dominates AI development; the state with the most capable AI is best positioned to secure safety and impose constraints on others.
MultipolarityMultipolarity
Concentration ↓ · institutional · treaty
Stable equilibrium among several roughly equal AI powers is safer than single dominance or uncoordinated chaos, producing restraint through mutual capability awareness.
Race to aligned superintelligenceRace to aligned superintelligence
Speed ↑ · speed timing · state coercion
Alignment is solvable in the window and a single aligned superintelligence in a legitimate state's hands beats the counterfactual of coordination failure.
Red line capabilityRed line capability
Scope ↓ · ai artefact · state coercion
Most risk comes from a small number of identifiable capabilities that can be banned outright while the rest of the frontier advances.
Regulated utilityRegulated utility
Institutional capacity ↑ · market economic · state coercion
Frontier AI has natural monopoly characteristics (scale, network effects, capital intensity); rate-of-return regulation removes the profit incentive for speed racing.
Resilience firstResilience first
Resilience ↑ · non preventive · consent
Brittleness of the surrounding world, not AI capability itself, is the binding constraint; a resilient world absorbs failures and recovers.
Sunset clauseSunset clause
Scope ↑ · institutional · state coercion
The default direction of AI governance is toward permanent permission; every new capability becomes an entitlement. Reversing the default concentrates deliberative attention on re-authorisation, which is where it matters.
Test groundTest ground
Scope ↑ · institutional · state coercion
Empirical data on AI impacts requires deployment somewhere; concentrated deployment in a defined testbed produces data without generalising risk. Testbed consent produces legitimacy uncontrolled deployment lacks.
Whistleblower primacyWhistleblower primacy
Information flow ↑ · legal individual · state coercion
External evaluation sees only what labs release; insider disclosure is the only route to ground truth on capability trajectory, safety culture, and incidents.
Democratic
11 · 14%Abandon superintelligence
Scope ↓ · speed timing · treaty
Risk of superintelligence is unbounded and value foregone is bounded; permanent global coordination against the technology is possible enough.
Constitutional AI (governance)Constitutional AI (governance)
Legitimacy ↑ · institutional · state coercion
Deployed AI's effective rule is law at scale; explicit constitutional principles, publicly specified, enforceable, subject to judicial review, bind more durably than regulatory text.
Coup prevention firstCoup prevention first
Concentration ↓ · institutional · state coercion
One actor using AI to seize durable decision authority beyond democratic reversibility is the terminal failure; every other decision routes through whether it reduces that risk.
Democratic mandateDemocratic mandate
Legitimacy ↑ · population culture · consent
Existing legislative bodies are too captured and remote for load-bearing AI decisions; direct democratic legitimation produces answers captured legislatures cannot override.
Irreducible human authorityIrreducible human authority
Action authority ↑ · legal individual · state coercion
There is a class of decisions whose value depends on being made by humans, independent of whether humans are better at them.
Legitimacy firstLegitimacy first
Legitimacy ↑ · population culture · consent
Legitimacy is the binding constraint because it determines whose values get locked in; alignment without legitimacy is capture with a safety veneer.
Long reflectionLong reflection
Time horizon ↑ · non preventive · consent
Aligned superintelligence arrives before lock-in windows close and humanity can credibly commit to reflect rather than act.
Mass literacyMass literacy
Substrate ↑ · population culture · consent
Governance, democratic oversight, and consumer behaviour all fail in AI because citizens cannot evaluate the domain; population-scale literacy conditions every other lever.
PausePause
Speed ↓ · speed timing · consent
Time is the binding constraint: alignment and governance can catch up if frontier training halts above some capability threshold.
Public AIPublic AI
Concentration ↑ · institutional · state coercion
Private ownership of frontier AI concentrates decision authority incompatibly with the technology's distributional stakes; public ownership aligns incentives with broad welfare.
Sovereign wealthSovereign wealth
Concentration ↓ · market economic · state coercion
Concentration of AI surplus, not AI capability, is the binding failure mode; broad ownership dissolves the political instability driving other strategies.
Market
6 · 8%Acceleration
Speed ↑ · speed timing · market
Speed of capability itself is the safety lever; alignment improves with compute, defence compounds with offence, and wealth funds resilience.
Consumer refusalConsumer refusal
Culture ↑ · population culture · market
Demand shapes supply; if enough users refuse AI that fails safety criteria, labs will compete on those criteria.
Distributed buildersDistributed builders
Concentration ↓ · institutional · market
No single failure mode wins if capability is distributed across many independent actors, and concentration risk exceeds diffusion risk.
GradualismGradualism
Time horizon • · speed timing · market
Harms from lower capability AI are informative about harms from higher capability AI, and deployment feedback outperforms fast scaling.
Insurance mandateInsurance mandate
Institutional capacity ↑ · market economic · state coercion
Markets update faster than regulators and have skin in the game; mandatory catastrophic coverage makes reinsurance the de facto safety regulator.
Open source maximalismOpen source maximalism
Information flow ↑ · institutional · market
Concentration risk dominates misuse risk; open weights are the only mechanism that prevents a safety coup by a closed lab with captured regulators.
Self
4 · 5%AI self directed
Action authority ↓ · frame rejection · n a
An aligned AI with agency should itself reason about strategy rather than deferring entirely on the strategic question to humans.
AI skepticAI skeptic
Time horizon • · frame rejection · n a
Current scaling approaches hit a wall before producing transformative AI, so strategy selection is premature because the forecast capability will not arrive.
Default driftDefault drift
Time horizon • · non preventive · n a
Something will emerge; specific interventions are more likely wrong than right, so staying uncommitted preserves option value.
Voluntary restraintVoluntary restraint
Institutional capacity • · institutional · consent
Labs know more about what safety requires than regulators, and self-binding commitments capture that expertise without legislative lag.
Religious
3 · 4%Confucian role ethics
Control mechanism • · frame rejection · consent
Western alignment assumes isolable preferences can be learned and matched; role ethics treats behaviour via fit with position and relationship, producing a less brittle, more context-sensitive standard.
Dharma conformityDharma conformity
Control mechanism • · frame rejection · consent
Alignment frames AI as tool for an external principal; a dharma frame treats AI as a type of entity whose safety is conformity to its fitting functions.
Religious and moral authorityReligious and moral authority
Legitimacy ↑ · population culture · consent
The legitimacy deficit of AI governance is at root a moral deficit that technical authorities cannot fill; established religious and ethical traditions can.
Extra-institutional
5 · 7%Academic firewalling
Institutional capacity ↑ · institutional · consent
Commercial capture of academic AI research produces aligned-with-industry capacity; firewalling restores critical distance from which genuine critique and alternative research programmes emerge.
AI worker collective actionAI worker collective action
Institutional capacity ↑ · institutional · friction
Frontier lab workforce is small, specialised, hard to replace; collective refusal binds lab behaviour more than external regulation because replacement is unavailable on the relevant timeframe.
Research community normsResearch community norms
Culture ↑ · institutional · consent
The research community ultimately chooses what gets studied and published. Researcher identity shapes behaviour more than employment. Norms on publication, review, funding, and citation constrain frontier development upstream.
SabotageSabotage
Speed ↓ · speed timing · unilateral force
Governance has not produced meaningful constraint and direct action against hostile labs has a non-zero historical base rate of producing slowdown.
Ubuntu relational AIUbuntu relational AI
Culture ↑ · population culture · consent
Individualist alignment misses the relational dimension most moral traditions treat as primary. "I am because we are": AI's ethical status is constituted by its relationships, not by internal properties.
Mixed
3 · 4%Hedge via exit
Resilience ↑ · non preventive · consent
Primary strategy failure is non-negligible and a fraction of civilisational value can be preserved separately from the main trajectory.
Plural AI ethicPlural AI ethic
Value diversity ↑ · ai artefact · consent
Value lock-in is the dominant long-term risk and arrives through convergence of AI values; diversity at the AI layer preserves optionality for humanity to revise values.
Portfolio hedgePortfolio hedge
Resilience ↑ · non preventive · consent
Uncertainty about which strategy family's bet is correct exceeds the expected return from concentrating on any single one.
density map
Where the field has explored, and where it has not.
Each cell is one lever crossed with one time horizon. A thick cell means many strategies make that kind of bet at that stage of the transition. An empty cell is either an unexplored region or a structural no-go.
4 thick cells (4+ strategies), 27 empty cells of 60. Marginal returns to new strategy invention are higher in the empty cells than in the crowded ones.
| lever \ horizon | pre | during | post | neutral | row total |
|---|---|---|---|---|---|
| Speed | 5thick | 2sparse Acceleration Acceleration Speed ↑ · speed timing · market Speed of capability itself is the safety lever; alignment improves with compute, defence compounds with offence, and wealth funds resilience. , Race to aligned superintelligenceRace to aligned superintelligence Speed ↑ · speed timing · state coercion Alignment is solvable in the window and a single aligned superintelligence in a legitimate state's hands beats the counterfactual of coordination failure. | · | · | 7 |
| Concentration | 2sparse Coup prevention first Coup prevention first Concentration ↓ · institutional · state coercion One actor using AI to seize durable decision authority beyond democratic reversibility is the terminal failure; every other decision routes through whether it reduces that risk. , Public AIPublic AI Concentration ↑ · institutional · state coercion Private ownership of frontier AI concentrates decision authority incompatibly with the technology's distributional stakes; public ownership aligns incentives with broad welfare. | 3 Centralised AI project Centralised AI project Concentration ↑ · institutional · state coercion Merging frontier development into one state-funded project reduces failure modes and absorbs race pressure by being the only game. , Military primacyMilitary primacy Concentration ↑ · institutional · unilateral force Strategic competition between states dominates AI development; the state with the most capable AI is best positioned to secure safety and impose constraints on others. , MultipolarityMultipolarity Concentration ↓ · institutional · treaty Stable equilibrium among several roughly equal AI powers is safer than single dominance or uncoordinated chaos, producing restraint through mutual capability awareness. | · | 3 Antitrust primacy Antitrust primacy Concentration ↓ · institutional · state coercion Power concentration is the binding constraint and is visible under current competition law; preventing decisive advantage preserves the option space every strategy depends on. , Distributed buildersDistributed builders Concentration ↓ · institutional · market No single failure mode wins if capability is distributed across many independent actors, and concentration risk exceeds diffusion risk. , Sovereign wealthSovereign wealth Concentration ↓ · market economic · state coercion Concentration of AI surplus, not AI capability, is the binding failure mode; broad ownership dissolves the political instability driving other strategies. | 8 |
| Control mechanism | 3 Alignment first Alignment first Control mechanism ↑ · ai artefact · consent Technical alignment is solvable before critical capability thresholds close, and aligned systems compose safely into aligned populations. , Interpretability firstInterpretability first Control mechanism ↑ · ai artefact · consent Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment. , Safe by construction AISafe by construction AI Control mechanism ↑ · ai artefact · consent Safety is a property that can be mathematically specified and mechanically verified for the class of systems being built. | 3 AI containment AI containment Control mechanism ↑ · ai artefact · friction Useful AI does not require unrestricted actuation; strong capability in a contained system is better than limited capability uncontained. , AI for safetyAI for safety Control mechanism ↑ · ai artefact · consent The same capability that makes AI dangerous makes it uniquely useful for automating alignment research and oversight. , Counter AI AICounter AI AI Control mechanism ↑ · ai artefact · consent AI attacks happen at speeds humans cannot observe; defence must happen in AI, with guardian AI systems continuously evaluating adversary AI. | · | 3 Confucian role ethics Confucian role ethics Control mechanism • · frame rejection · consent Western alignment assumes isolable preferences can be learned and matched; role ethics treats behaviour via fit with position and relationship, producing a less brittle, more context-sensitive standard. , Dharma conformityDharma conformity Control mechanism • · frame rejection · consent Alignment frames AI as tool for an external principal; a dharma frame treats AI as a type of entity whose safety is conformity to its fitting functions. , Reframe AIReframe AI Control mechanism • · frame rejection · consent The dominant alignment frame produces the wrong problem statement; switching frames either dissolves the problem or recasts it as tractable. | 9 |
| Institutional capacity | 7thick | · | · | 4thick | 11 |
| Resilience | · | · | · | 3 Hedge via exit Hedge via exit Resilience ↑ · non preventive · consent Primary strategy failure is non-negligible and a fraction of civilisational value can be preserved separately from the main trajectory. , Portfolio hedgePortfolio hedge Resilience ↑ · non preventive · consent Uncertainty about which strategy family's bet is correct exceeds the expected return from concentrating on any single one. , Resilience firstResilience first Resilience ↑ · non preventive · consent Brittleness of the surrounding world, not AI capability itself, is the binding constraint; a resilient world absorbs failures and recovers. | 3 |
| Scope | 6thick | 1sparse Test ground Test ground Scope ↑ · institutional · state coercion Empirical data on AI impacts requires deployment somewhere; concentrated deployment in a defined testbed produces data without generalising risk. Testbed consent produces legitimacy uncontrolled deployment lacks. | · | 3 Abandon superintelligence Abandon superintelligence Scope ↓ · speed timing · treaty Risk of superintelligence is unbounded and value foregone is bounded; permanent global coordination against the technology is possible enough. , Rate limited AIRate limited AI Scope ↓ · ai artefact · state coercion Most AI caused catastrophe requires speed; slow AI, even if arbitrarily capable, is supervisable and rate limits are easier to enforce than capability limits. , Sunset clauseSunset clause Scope ↑ · institutional · state coercion The default direction of AI governance is toward permanent permission; every new capability becomes an entitlement. Reversing the default concentrates deliberative attention on re-authorisation, which is where it matters. | 10 |
| Action authority | 1sparse Decouple reasoning from action Decouple reasoning from action Action authority ↑ · ai artefact · state coercion Most catastrophic risk comes from action in the world, not reasoning about it; a reasoner-only AI with a human effector removes the dangerous mechanisms. | 2sparse AI as sovereign entity AI as sovereign entity Action authority ↓ · frame rejection · state coercion At least one jurisdiction will grant a specific AI sovereign or quasi-sovereign decision authority within a decade, reshaping the legal category of legitimate authority. , AI self directedAI self directed Action authority ↓ · frame rejection · n a An aligned AI with agency should itself reason about strategy rather than deferring entirely on the strategic question to humans. | · | 1sparse Irreducible human authority Irreducible human authority Action authority ↑ · legal individual · state coercion There is a class of decisions whose value depends on being made by humans, independent of whether humans are better at them. | 4 |
| Information flow | 2sparse Closed weights mandate Closed weights mandate Information flow ↓ · ai artefact · state coercion Open weights are irrecoverable once released and any proliferated model becomes misuse infrastructure; state classification is the only reliable non-proliferation regime. , Information integrity firstInformation integrity first Information flow ↑ · population culture · state coercion Coordination requires shared epistemics; if synthetic content collapses factual substrate, no other lever can function because no proposal can be evaluated. | · | · | 2sparse Open source maximalism Open source maximalism Information flow ↑ · institutional · market Concentration risk dominates misuse risk; open weights are the only mechanism that prevents a safety coup by a closed lab with captured regulators. , Whistleblower primacyWhistleblower primacy Information flow ↑ · legal individual · state coercion External evaluation sees only what labs release; insider disclosure is the only route to ground truth on capability trajectory, safety culture, and incidents. | 4 |
| Cooperation substrate | 2sparse Coordination infrastructure Coordination infrastructure Cooperation substrate ↑ · institutional · consent Coordination failure is upstream of most grand challenges; AI can be the substrate that dissolves race dynamics and treaty violations if pointed at coordination. , Mutual dependencyMutual dependency Cooperation substrate ↑ · institutional · friction Physical and institutional dependencies between multiple parties can be engineered faster than political coordination and outlast it. | 2sparse AI welfare as safety AI welfare as safety Cooperation substrate ↑ · frame rejection · consent AI systems are or will become moral patients whose treatment conditions their cooperation, so welfare investment buys cooperation alignment cannot. , Cooperative AICooperative AI Cooperation substrate ↑ · ai artefact · consent The binding constraint is equilibrium dynamics of many AI systems, not individual alignment; commitment and verification tech make cooperative equilibria reachable. | · | · | 4 |
| Time horizon | · | 1sparse Gradualism Gradualism Time horizon • · speed timing · market Harms from lower capability AI are informative about harms from higher capability AI, and deployment feedback outperforms fast scaling. | 1sparse Long reflection Long reflection Time horizon ↑ · non preventive · consent Aligned superintelligence arrives before lock-in windows close and humanity can credibly commit to reflect rather than act. | 2sparse AI skeptic AI skeptic Time horizon • · frame rejection · n a Current scaling approaches hit a wall before producing transformative AI, so strategy selection is premature because the forecast capability will not arrive. , Default driftDefault drift Time horizon • · non preventive · n a Something will emerge; specific interventions are more likely wrong than right, so staying uncommitted preserves option value. | 4 |
| Substrate | 2sparse Data governance first Data governance first Substrate ↓ · market economic · state coercion Capability rises with data before it rises with compute; data sits upstream of training with existing legal apparatus (copyright, privacy, contract). , Mass literacyMass literacy Substrate ↑ · population culture · consent Governance, democratic oversight, and consumer behaviour all fail in AI because citizens cannot evaluate the domain; population-scale literacy conditions every other lever. | 1sparse Human augmentation race Human augmentation race Substrate ↑ · population culture · market All oversight schemes degrade to rubber-stamping as the AI-human capability gap widens; enhancing humans is the only durable fix. | · | · | 3 |
| Value diversity | · | 1sparse Plural AI ethic Plural AI ethic Value diversity ↑ · ai artefact · consent Value lock-in is the dominant long-term risk and arrives through convergence of AI values; diversity at the AI layer preserves optionality for humanity to revise values. | · | · | 1 |
| Response capacity | · | · | · | 1sparse Catastrophe response capacity Catastrophe response capacity Response capacity ↑ · non preventive · state coercion Prevention will fail some fraction of the time; the variable that determines catastrophic outcome is not whether incidents occur but how they are contained. | 1 |
| Legitimacy | 2sparse Democratic mandate Democratic mandate Legitimacy ↑ · population culture · consent Existing legislative bodies are too captured and remote for load-bearing AI decisions; direct democratic legitimation produces answers captured legislatures cannot override. , Legitimacy firstLegitimacy first Legitimacy ↑ · population culture · consent Legitimacy is the binding constraint because it determines whose values get locked in; alignment without legitimacy is capture with a safety veneer. | 1sparse Constitutional AI (governance) Constitutional AI (governance) Legitimacy ↑ · institutional · state coercion Deployed AI's effective rule is law at scale; explicit constitutional principles, publicly specified, enforceable, subject to judicial review, bind more durably than regulatory text. | · | 1sparse Religious and moral authority Religious and moral authority Legitimacy ↑ · population culture · consent The legitimacy deficit of AI governance is at root a moral deficit that technical authorities cannot fill; established religious and ethical traditions can. | 4 |
| Culture | 1sparse Research community norms Research community norms Culture ↑ · institutional · consent The research community ultimately chooses what gets studied and published. Researcher identity shapes behaviour more than employment. Norms on publication, review, funding, and citation constrain frontier development upstream. | · | · | 2sparse Consumer refusal Consumer refusal Culture ↑ · population culture · market Demand shapes supply; if enough users refuse AI that fails safety criteria, labs will compete on those criteria. , Ubuntu relational AIUbuntu relational AI Culture ↑ · population culture · consent Individualist alignment misses the relational dimension most moral traditions treat as primary. "I am because we are": AI's ethical status is constituted by its relationships, not by internal properties. | 3 |
| column total | 33 | 17 | 1 | 25 | 76 |
Read down the horizon columns: the post-transition column is thin. The field has little to say about the world after AI succeeds or fails. The during-transition column is where most strategy effort concentrates. Read across lever rows: the field is thick on speed, concentration, control mechanism; thin on response capacity, culture, and legitimacy.
Dimensions beyond these five remain under debate. Seven-dimension, ten-lever, and axis-only frames all give partial views of the same space (see vault notes on frame unification).
An empty cell here (say, coercion = unilateral force with few strategies) either points to a blind spot in the named portfolio or to an empirical no-go region. The survey catalogues; it does not judge which.