AGI Strategies

diagnostic

Is this a bet, or an identity?

The survey treats every strategy as a bet about which failure mode binds. But in the AI safety field, many strategies function as identities rather than as bets. Identities persist regardless of evidence, and their advocates resist updates that would require changing them.

You cannot tell whether a position is a bet or an identity from the position itself. You have to look at how the advocate holds it. Five diagnostic questions:

question
bet holder
identity holder

What evidence would change the advocate's mind?

Names specific, plausible signals.

Refuses the question, or names signals so high no real event meets them.

Can the advocate articulate the strongest version of the opposing view?

Yes, and correctly.

No, or only in caricature.

What costs do they pay beyond reputation?

Allocates resources across conditions. Has skin in the failure mode.

Allocates as signal. Costs reinforce rather than test.

Has the advocate ever publicly updated on this topic?

Yes, and frames the update as a gain.

No, or frames past positions as "always meant."

Does advocating the strategy produce identity rewards (community, status, belonging)?

Rewards are tied to being right over time.

Rewards are tied to continued advocacy regardless of outcome.

Called-out identity bindings

4

Strategies where the current advocate cohort shows identity markers rather than bet markers, per the vault note. The position can still be right; identity binding is about how, not whether.

Alignment firstCareer category

"Alignment researcher" is a job title, a conference track, a grant category. Updating away from the strategy requires leaving the field.

PauseMovement membership

Pause AI membership is defined by the position. Updating requires exiting the community.

AI skepticPublic figure

Functions as identity for public critics whose standing depends on continued skepticism. Updates cost the identity.

AccelerationSubculture

Identity-signalled in e/acc aesthetics, memes, social clustering. The signal is cultural, not evidential.

Falsification signals, named.

42 strategies

The concrete test for whether a strategy is held as a bet: name the signal that would update against it. These are catalogued; advocates who cannot name a signal are holding the position as identity.

Acceleration

A visible harm large enough that policy overrides the growth coalition (2008 financial crisis analogue).

AI skeptic

Each capability threshold the skeptic named as unreachable is reached, though the position often survives via recalibration rather than abandonment.

Alignment first

Interpretability and oversight methods stop scaling with model capability, stronger models are less rather than more inspectable.

Antitrust primacy

Breakups reconcentrate within one to two years.

Arms control treaty

Signatories cannot domestically enforce (the BWC pattern).

Bureaucratic slowdown

Procedural burden implemented and routinely evaded (pre-1970s environmental impact analogue).

Catastrophe response capacity

A major incident where capacity exists but cannot scale (2008 financial response analogue).

Closed weights mandate

Classified capability appears unclassified within two years.

Compute governance

The capability-per-flop curve steepens faster than chip export controls tighten.

Consumer refusal

Major lab scandals produce no measurable user migration, which is the current pattern.

Cooperative AI

AI systems defect in deployments where commitment technology exists and cooperation was available.

Counter AI AI

The best guardian system is fooled by a model one generation newer.

Coup prevention first

An undetected coup crosses the threshold, the detection regime does not currently exist.

Criminal liability

Clear criminal conduct is identified with no prosecution (2008 Wall Street analogue).

Data governance first

Frontier capability is reached with synthetic data.

Democratic mandate

Binding AI referenda are functionally ignored within three years.

Differential technology development

The offense-defense classification cannot be operationalised in any funded program within five years.

Embodiment requirement

A catastrophe caused by embodiment-exempt AI.

Energy choke point

Efficiency gains outpace regulatory tightening.

Governance first

Enacted regulations cover less than 20% of frontier compute by some date, or institutional capture moves faster than capacity building.

Human augmentation race

Default failure case: capability gap widens faster than augmentation narrows it. Falsification requires a discontinuous enhancement result.

Information integrity first

Continued political coordination under synthetic saturation, or provenance infrastructure operationalised as surveillance.

Insurance mandate

A large AI loss triggers insurer exit rather than tighter safety requirements.

International AI agency

No agency with inspection authority is negotiated and operational within the next several years.

Interpretability first

Leading labs cannot produce mechanistic explanations of their own frontier models within two to three years of release.

Mass literacy

High measured literacy produces no behavioural change on consumer and voting choices by 2030.

Military primacy

Catastrophic outcome under a race dynamic that the strategy predicted would be stable.

Multipolarity

Any actor achieves decisive advantage others cannot match within a planning cycle.

Narrow AI preservation

Narrow compositions cannot match general system economic returns.

Open source maximalism

An open released model produces a verified harm in a domain where defender access does not bound the risk.

Pause

Major states or frontier labs publicly defect from a declared pause, or verification tech cannot distinguish a real pause from a declared one.

Plural AI ethic

Measured value convergence across frontier models within three years.

Public AI

Concentration risks appear inside the public entity, or the public entity lags the private frontier into irrelevance.

Red line capability

A system crosses a named red line without a prior warning signal, or many deployed systems hold a red line capability latently.

Regulated utility

Utility regulation produces no safety investment above voluntary baseline.

Religious and moral authority

Formal religious positions move no outcome.

Resilience first

Core infrastructure degradation rates exceed hardening rates for three consecutive years, particularly in verification cost and democratic trust.

Sabotage

No credible actor attempts the path; strategy was correctly assessed as non-viable.

Small model first

A widening scale gap through 2027.

Sovereign wealth

Captured surplus leaves political concentration unchanged (Alaska PFD analogue).

Voluntary restraint

Visible weakening of RSP text under capability pressure, combined with no meaningful penalty.

Whistleblower primacy

A major safety incident is known internally and not disclosed even under the new regime.

Why it matters. A field where positions are identities produces culture war dynamics rather than convergence on evidence. The empirical claims underneath become unobservable because challenging them is social apostasy.

What would help. Explicit falsification signals attached to positions, resourced tracking against them, funding reviews that weight updates over persistence, public retractions treated as achievements rather than failures. Calibration infrastructure preserves bets against identity drift.