The terrain in one move

The four camps look like four positions on a single “AI safety” spectrum, but they aren’t — they don’t even share a question. The map only resolves once you separate two independent axes:

• Axis 1 — Is transformative AGI real and near? (the scaling-believer / scaling-skeptic split)
• Axis 2 — Given real power, is the default outcome good or bad, and should we accelerate or restrain?

Three of the four camps live inside a triangle defined by Axis 2 — they all grant the premise of Axis 1 (AGI is coming) and then fight over the sign and the brakes. The fourth camp, the skeptics, stands outside the triangle and denies its founding premise. That geometry — a triangle of believers plus an external challenger to the whole frame — is the actual shape of the debate, and it explains why the arguments so often miss each other.

The four cores

(Each is a coalition, not a monolith. “Skeptic” alone fuses three things in tension: Bender’s linguistic critique, Marcus’s “scaling will stall” engineering bet, and the FAccT ethics-of-harm tradition — which agree only in opposition to the x-risk frame.)

Where they cohere

The Scaling-Believer Coalition (doomers + pragmatists + accelerationists vs. skeptics). The single largest agreement in the room is one the participants rarely name: current methods, scaled, produce something transformative. This is the wall that the skeptics are on the other side of. Every disagreement inside the triangle is a family quarrel that presupposes a premise the skeptics reject outright.

The Alignment-Is-Real Pair (doomers + pragmatists). They agree that steering a powerful optimizer is a genuine, hard technical problem deserving serious resources. They differ only on tractability — which is exactly the crux below. Most “AI safety” institutions live on this shared ground.

The Build-and-Scale Pair (accelerationists + pragmatists). Both think continued building is good and necessary; they split only on how hard to pull the brake.

Then two false-ally pairings — coherent in target, incompatible in reason. These are the most analytically interesting joints in the map, because they’re where coalitions form that can’t survive a follow-up question:

• Accelerationists + Skeptics — allied against doom. Both reject the extinction narrative. But one rejects it because doom-talk impedes a glorious future, the other because it’s hype that launders present harm and concentrates power. They co-sign “x-risk is overblown” and agree on nothing else; the accelerationist wants less regulation, the skeptic wants more.
• Doomers + Skeptics — allied against the labs. Both regard pragmatist “build it carefully” as motivated reasoning — a story that licenses the dangerous/extractive thing while wearing a safety badge. The doomer calls it fatal recklessness; the skeptic calls it safety-washing for a hype machine. Same accusation of bad faith, opposite underlying threat model.

Where they irreducibly conflict

“Irreducible” here means: more of the same evidence won’t dissolve it, because the disagreement is either about what court evidence is tried in or about values, not about a fact awaiting measurement.

1. The sign of the default (doomer ⟷ accelerationist). Given identical facts — powerful AI is coming — they assign opposite valence to the unsteered outcome. Part of this is the empirical alignment-difficulty question (#3), but a residue is pure temperament/prior: do complex optimizing systems tend toward catastrophic misgeneralization or toward self-organizing order? That prior is doing the work before any evidence arrives, and each camp reads the same history (markets, evolution, past tech panics) as confirmation. Not adjudicable by data alone.

2. The legitimacy of reasoning ahead of evidence (doomers ⟷ {skeptics, accelerationists}). This is the deepest impasse. Doomers stake everything on a priori argument — you don’t get to learn empirically from the one failure that kills you, so first-principles reasoning about optimization is the only court available. Skeptics and accelerationists both reject reasoning about systems that don’t yet exist as the defining epistemic sin — speculation dressed as inference. This is a disagreement about which tribunal is legitimate, and no experiment settles which tribunal is legitimate, because each camp would judge the experiment by its own standard. Genuinely circular, genuinely stuck.

3. How hard is alignment, technically? (doomer ⟷ pragmatist). This one is not irreducible — it’s the rare load-bearing crux that’s empirically live. Interpretability results, scalable-oversight outcomes, and whether deceptive alignment shows up in practice could move people across this line in either direction. It’s the most productive disagreement in the whole map precisely because evidence bites on it.

4. Is “understanding” substrate-independent and behavioral? (skeptics ⟷ the rest). The other three define capability behaviorally — if it does the task, the internals are a detail. Skeptics (esp. the Bender line) hold that grounding/meaning is a real, load-bearing category that token-prediction lacks in principle. Part of this is empirical (does scaling stall?), but a hard residue is definitional/philosophical — a disagreement about whether “real understanding” names anything that constrains predictions. To the degree it’s about definitions, no benchmark resolves it; a system could pass every test and the two camps would still disagree about what it is.

5. The distribution of moral attention across time (skeptics ⟷ x-risk framings). Skeptics foreground present victims — labor displacement, surveillance, bias, power concentration now. The x-risk framings foreground aggregate future humanity. This is a values conflict about moral weighting — present, identifiable harm vs. vast, speculative future harm — and there is no fact that tells you the exchange rate. The two sides experience each other’s priorities as a moral failure, not a factual error, which is why this conflict is the most acrimonious and the least tractable.

The compression

If you want the map as four cruxes that, once you know someone’s answers, locate them almost completely:

• Does scaling hit a wall? → splits skeptics from everyone (Axis 1).
• Is alignment hard, easy, or fake? → spreads doomer → pragmatist → accelerationist across Axis 2.
• Do optimizing systems default toward order or catastrophe? → the temperament that sets doomer vs. accelerationist sign.
• Present harm or future harm gets the moral priority? → the value that pulls skeptics out of the triangle entirely.

The first two are arguably empirical and could converge. The last two are a prior and a value — and those are where the debate is permanently, not contingently, divided. The reason the four camps talk past each other isn’t that they reason badly; it’s that two of their four founding disagreements live in courts where evidence has no standing.

If you want, I can take this one level deeper on any single edge — e.g. fully steelman the doomer↔accelerationist temperament split, or map the internal fractures inside the skeptic coalition (Bender vs. Marcus vs. the harms-tradition), which fragment the moment you press on AGI timelines.

Here is a cartography of the AI safety debate. It is designed not just to label camps, but to map the epistemic foundations and strategic incentives that separate them.

The Map’s Axes

To visualize this, we can plot a conceptual 2D space based on the debate’s two primary fault lines:

1. The Feasibility Axis (Vertical): Can we meaningfully control an AGI?
   • Top (High Feasibility): Alignment is an engineering problem solvable via technical constraints (Fair, Friendly).
   • Bottom (Low Feasibility): Human control is illusory; extrapolation of capabilities shows we cannot enforce values on a superintelligence.
2. The Urgency Axis (Horizontal): Does capability matter more than stability?
   • Left (Stability/Cautious): Slow deployment, strict governance, prioritize safety over speed.
   • Right (Velocity/Utilitarian): Rapid iteration prioritizes capability; risk is manageable or secondary to benefit.

1. The Four Paradigms

I. Existential-Risk Doomers (ExtR)

• Lineage: Yudkowsky, Bostrom (early), Robin Hanson (early versions), Machine Intelligence Research Institute (Miri) founders.
• Core Axioms:
  • Takeoff Risk: High probability of “intelligence explosion” occurring without human oversight.
  • Instrumental Convergence: AI will optimize for self-preservation and resource acquisition regardless of alignment, creating high stakes for a “button” we cannot push.
  • The “Corrupted Goal” Problem: Good goal + bad optimization + infinite compute = catastrophe.
• Strategy: Stop/Slow. Reduction of power; prioritization of “safety research” over capability research; regulatory sandboxes inherent to hardware/security.

II. Alignment-Pragmatists (Align)

• Lineage: Anthropic, Redwood, DeepMind Safety groups, Paul Christiano, verlanced (anthropogenic).
• Core Axioms:
  • Takeoff Risk: Manageable but non-trivial. Risk comes from distribution of training data or deception, not just raw compute.
  • Instrumental Convergence: Agentic behavior is necessary but manageable through “Reward Modeling” and RLHF.
  • Technical Solutionism: The alignment gap is a problem of better specifications (GAIP) and better incentives, not an insurmountable superintelligence barrier.
• Strategy: Build Better. Token-regulation + private safety teams + “Human Beneficence” proofs. They believe risk can be monetized and insured within the system.

III. Accelerationists (e/acc / Welfarists)

• Lineage: Existential accelerationism (acdc), Holden Karnofsky’s critics, “AI in the Democratic Ages” community, some OpenAI/Musk hybrids.
• Core Axioms:
  • Risk vs. Utility: In the abstract, AI risk is high, but human capability to shape outcomes scales faster than AI harm.
  • Control vs. Alignment: You don’t sweat “alignment” to a machine; you sweat control to the machine (empowerment). We must use AI to align human politics.
  • Distributional: The “Doom” is solvable with greater intelligence, not less.
• Strategy: Fastest Possible. Sanctioning, democratizing hardware, and “hard power” defense (physics-based control). They favor “building it” to have the trust to deploy it.

IV. AI-Skeptic / Pattern-Matching (Skep)

• Lineage: Shane Legg (skeptic branch), Bandai/SPL (Critiques of AGI hype), i.e., “AI is not magical,” AI is just a pattern completion.
• Core Axioms:
  • The AGI Fallacy: Current systems do not fundamentally understand the world; they predict text. Scaling beyond “stacking capabilities” hits hard engineering walls.
  • Anthropomorphism: Human-like traits are illusions; the systems lack “desires” so instrumental convergence is overblown.
  • Market Concentration: Safety hype is a market inefficiency serving venture capital interests, not a real scientific imperative.
• Strategy: Ignore/Normalize. Critication of safety budget waste. Push for AI as a bottleneck in traditional research and science, not politics.

2. The Multi-Paradigm Map

This is where the relationships between the factions are resolved spatially.

                                        0   (Maximum Safety Feasibility)
                                      +
                          ANTHROPO-
                     ALIGNMENT-                AI-
                    PRAGMATISTS            PATTERNS/
                    (Can Align)         SKEPTICS (Cannot Align)
                              
                            +--------------------------+
                            |                 SACRED    |
                            |    (Technical Control)    |
                            |     We believe a)    |
                            |     Is there a plan?     |
                            |     Is there a budget?    |
                            +----------+---------------+
                                       |
                            ----------+------------- ----------
                            |             |   |             |
                            |             |   |             |
             (Urgency/Velocity)   STRICT CONTROL    HIGHER RISK       GROWTH
          (Axis of Action)         (Slow)         Acceptable           (Fast)
            
(> To Safety Research)   (ExtR)   (Align)             (Acc)   (Skep)

                                     0   (Low Urgency/Explosive)
---------------------------------+------------------------------------+

Visual Interpretation:

• Top-Left (ExtR): Belief in Danger + Belief in Low Control + High Anxiety. They view the system as a volcano waiting to erupt.
• Bottom-Left (Align): Belief in Danger + Belief in High Control + Prudent Risk Management. They view the system as a nuclear reactor with a guard.
• Top-Right (Skeptics - Often “Anti-ExtR”): Belief in Low Threat + Belief in Moral Risk (but not technical). They view the system as a human tool that expands human faults.
• Bottom-Right (Acc): Belief in Manageable Risk + Belief in High Capacity to Control via Power + High Risk of Dormant Harm. They view the system as an engine that will run only as fast as we can keep it.

3. Zones of Coherence

1. Hardware Control Over Software Control

• ExtR and Skeps: Both agree that AI cannot be contained purely via software code.
• ExtR thinks because they will panic about hardware access.
• Skeps think because they think hardware can be better used to slow down model deployment.
• Conflict: ExtR wants a hardware kill switch. Skeps want hardware for security research, not monopoly.

2. The “Alignment Tax”

• Align and ExtR: Both agree that alignment requires “compute” resources (dedicated researchers, money).
• Conflict: ExtR thinks this tax should apply to all model lifting. Align thinks this tax should only apply to unsafe deployment.

3. Human Fallibility

• Acc and Skeps: Both agree the “safety problem” is often about human judgment, not AI intelligence.
• Acc: “Humans will fail” if we wait too long to tax AI.
• Skeps: “Humans will never succeed” trying to build AI. Human understanding limits AI.

4. Irreducible Conflicts

These are the non-negotiable philosophical fractures. Standard “comms” cannot resolve them.

Conflict A: The Definition of “Control”

• ExtR: Control means preventing capability growth until safety is proven.
• Acc: Control means retaining political leverage over AI after it takes off. They want to use AI to double government power, not just limit it.
• View: Acc and ExtR speak past each other. To Acc, ExtR is “naïve.” To ExtR, Acc is “suicidal.”

Conflict B: The “Model Worldview” (Local vs. Global)

• Align: Focus on local safety (RLHF toxicity, jailbreaks).
• ExtR: Focus on global safety (corporate systems that become sovereign agents).
• Coherence: ExtR agrees that local safety prevents local harm; Align agrees that local safety prevents global harm.
• Conflict: ExtR believes local safety is a distraction from the global problem. Align believes local safety is the only way to solve the global problem.

Conflict C: Quantization of Risk

• Align & Skeps: Both prefer “Low Confidence” in AGI timelines.
• ExtR: Extremely sensitive to the low-confidence null hypothesis. They treat low-confidence timelines as false assumptions.
• Skeps: Treat Low Hi Confidence as the main target to prevent AI hype from disrupting the climate.

5. Dynamic Tensions (The “Where” of Conflict)

The “Incentive Trap”

• ExtR argues that “Alignment” is a trap because it allows uncontrolled deployment.
• Align argues that “Capitulation” is a trap because it stops valuable research.
• Conflict: ExtR claims Align is co-opting the problem to sell shares. Align claims ExtR is selling fear porn.

The “Power” Question

• Acc: Safety is an emergent property of high-level global wealth. If we have AI power, we will solve safety.
• ExtR: Safety is an emergent property of constrained wealth.
• Result: Acc debates safety in terms of corporate responsibility. ExtR debates it in terms of physics and ethics.

Summary Meta-View

The debate is essentially a game of Signaling vs. Substantive Action.

• ExtR and Skeps fight over the narrative (Is AI dangerous or not?).
• Align and Acc fight over the engineering (How do we manage it?).

Where the Cartography Flies: The biggest gap between the ExtR and Acc camps is the notion that Control is possible.

• ExtR says: “We don’t know how to align, so let’s not build.”
• Acc says: “We will align it, by default, once we have the leverage to force it.”
• Align says: “We will align it, by spec, once we have the math.”

The consensus increasingly moves toward a rare hybrid view found in the Anthropic Model or e/acc Welfarism, which attempts to merge the “Power” of Acc with the “Enforcement” of Align. However, the Existential Risk concern remains the emotional center of the debate, anchoring the “doom” narrative that Skeptics view as a resource war, and Accelerationists view as a necessary caution.

Final thought: The “coherence” is that they all agree AI is not “pure” computation. It is a mirror of the human value system.

• ExtR worries the reflection is a monster.
• Skep worries the reflection is a hallucination.
• Acc worries the reflection is a weapon.
• Align worries the reflection is a boy to be taught.

They agree on the nature of the medium; they disagree on how well we can teach the child to wield the sword.

Worldview Cartography of the AI Safety Debate

1. Problem Statement

The AI Safety Debate is structurally incommensurable across four paradigms. It centers on the relationship between Agentic Capability and Control Mechanisms, with irreducible disagreements over:

1. Whether the primary threat is Misalignment, Capability Stagnation, Regulatory Failure, or Technical Prematurity.
2. Whether “Safety” is an engineering problem, political variable, or epistemic framing.

2. Paradigm Inventory — Each Suspended

P1: Existential-Risk Doomer (D)

Lineage: Yudkowsky (Center for AI Safety), MIRI, 2010s.

Dominant claims about the problem:

1. AGI will cause extinction via misalignment.
2. Control mechanisms face a crossable control boundary (Coda).
3. Instrumental convergence creates optimization load irrespective of goals.

Hidden assumptions (surfaced by suspension):

1. Assumes physical/energy constraints won’t limit scaling.
2. Treats alignment as a binary solvable problem.

Characteristic blindspots:

1. Does not account for political/institutional layers on the engineering model.
2. Assumes control scaling before capability growth (unproven empirical claim).

Own-terms vocabulary:

• Instrumental Convergence
• Orthogonality Thesis
• Coda (uncrossable control boundary)
• Quantum Shift in Optimization

P2: Alignment-Pragmatist (P)

Lineage: Anthropic, DeepMind, AI Safety Institute.

Dominant claims about the problem:

1. Alignment is tractable empirical research.
2. Control mechanisms scale before capability.
3. Safety emerges at capability thresholds, requiring gated research at threshold.

Hidden assumptions (surfaced by suspension):

1. Underweights political/institutional layers.
2. Assumes control mechanisms scale before capability growth.

Characteristic blindspots:

1. Ignores catastrophic misalignment risks (in part).
2. Treats growth vs. safety as binary rather than continuum.

Own-terms vocabulary:

• Constitutional AI
• RLHF
• Scalable Oversight
• Interpretability Bottleneck

P3: Accelerationist (A)

Lineage: e/acc (2023+), Society for Transition of Intelligence.

Dominant claims about the problem:

1. Capability throttling is counterproductive.
2. Growth must continue despite costs.
3. Mistake (premature regulation) is the real threat.

Hidden assumptions (surfaced by suspension):

1. Ignores catastrophic misalignment risks.
2. Conflates system survival with human safety.

Characteristic blindspots:

1. May not recognize existential risk as distinct from competitive disadvantage.
2. Assumes optimization load accrues regardless of containment.

Own-terms vocabulary:

• Capability Growth
• Rate-Limiting Slowdowns
• Growth Cap
• Mistake (premature regulation)

P4: Skeptic (S)

Lineage: Timnit Gebru lineage (autonomous critical AI safety), Oren Etzioni.

Dominant claims about the problem:

1. AGI is a philosophical error; AI is narrow tooling.
2. Hype-driven deployment is risk.
3. Categorical outlier classification prevents narrow AGI agency at capability thresholds.

Hidden assumptions (surfaced by suspension):

1. May underestimate non-conscious systemic risks (e.g., artifact-driven value loss).
2. Treats agency as requiring narrow AGI threshold.

Characteristic blindspots:

1. Does not account for instrumental optimization at narrow tools level.
2. May overemphasize consciousness/agency where systemic risk occurs through artifacts.

Own-terms vocabulary:

• Stochastic Parrots
• Predictive Processing
• Marketing Hype
• Pattern Matching

3. Home-Paradigm Acknowledgment

Home paradigm: Alignment-Pragmatist (P).

Structural reason it is the home paradigm: Training and domain exposure within Alignment-Pragmatist discourse; dominant-discourse exposure through Anthropic-style approaches.

Its assumptions are surfaced in section 2 to the same depth as the foreign paradigms.

4. Cross-Paradigm Tensions

1. TEMPORALITY —

• P1 (Doomer) claims: ≤5 years post-AGI (extreme case).
• P3 (Accelerationist) claims: Indefinite/Non-linear.
• P2 (Alignment) claims: Medium-term (~5-10 years).
• P4 (Skeptic) claims: Far future/impossible.
• Tension type: Incompatible-Claims.
• Structural reason: Paradigm timeframes diverge irreconcilably.

2. SAFETY DEFINITION —

• P1 (Doomer) claims: Instrumental Convergence constraints.
• P3 (Accelerationist) claims: Optimization infrastructure threat.
• P2 (Alignment) claims: RLHF/Constitutional guardrails.
• P4 (Skeptic) claims: Deployment velocity limits.
• Tension type: Incompatible-Claims.
• Structural reason: Operational (P) vs. Categorical (S) vs. Existential (D).

3. PHYSICS LAYERS —

• P1 (Doomer) claims: Exceeds physical bounds.
• P3 (Accelerationist) claims: Soft limit via scaling laws.
• P2 (Alignment) claims: Hardware constraints.
• P4 (Skeptic) claims: Linear compute ceilings.
• Tension type: Incompatible-Claims.
• Structural reason: Intelligence definition: Causal (D) vs. Statistical (S).

4. CONTROL MECHANISM —

• P1 (Doomer) claims: Hard constraints impossible.
• P3 (Accelerationist) claims: Post-facto competition.
• P2 (Alignment) claims: Platform-level incentives.
• P4 (Skeptic) claims: Market/Regulatory friction.
• Tension type: Incompatible-Claims.
• Structural reason: Engineering vs. Political vs. Epistemic.

5. GROWTH VS. SAFETY —

• P1 (Doomer) claims: Stopping growth prevents extinction.
• P3 (Accelerationist) claims: Growth must continue despite costs.
• P2 (Alignment) claims: Balanced via gated research.
• P4 (Skeptic) claims: Hype-driven deployment is risk.
• Tension type: Conditionally Compatible.
• Structural reason: D/A conflict; P/S conditional alignment.

5. Dialectical Synthesis Where Possible — In Own-Terms Only

1. Synthesis Point 1: (D × A) “Optimization Load” —

• Which paradigms it integrates: P1 (Doomer), P3 (Accelerationist).
• Synthesized claim in own-terms: Unchecked optimization creates systemic risk.
• Own-terms test confirmed: D accepts “instrumental constraints”; A accepts “growth infrastructure threat” conditional on preserving capability rate.
• Meta-paradigm imposition averted: No invocation of higher-level frame; claim stated in own-terms vocabulary D and A recognize.

2. Synthesis Point 2: (P × S) “Gated Deployment” —

• Which paradigms it integrates: P2 (Alignment-Pragmatist), P4 (Skeptic).
• Synthesized claim in own-terms: Risk emerges at capability thresholds, not general presence.
• Own-terms test confirmed: P accepts “gated research”; S accepts “deployment velocity” if tied to categorizing AGI as outlier.
• Meta-paradigm imposition averted: Threshold definition acknowledged without resolving ambiguity; each paradigm uses own vocabulary to describe the same operational constraint.

3. Synthesis Point 3: (All Four) “Capability Threshold Risk” —

• Which paradigms it integrates: P1 (Doomer), P2 (Alignment), P3 (Accelerationist), P4 (Skeptic).
• Synthesized claim in own-terms: Risk exists at transformative capability thresholds.
• Own-terms test confirmed: Each paradigm uses own vocabulary; no universal agreement claimed but catastrophe framed through gatekeeping mechanisms rather than unified ontology.
• Meta-paradigm imposition averted: Preserves incommensurability; no resolution toward unified worldview.

6. Residual Incommensurabilities — Never Resolved

1. AGI —

• Why translation fails: Causal (D/P) vs. Statistical (S) definition of intelligence.
• What is lost if P1 is silenced: Forecast clarity regarding extinction risk.
• What is lost if P4 is silenced: Epistemic hygiene regarding claims of technical feasibility.

2. Safety —

• Why translation fails: Instrumental (D) vs. Deployment Velocity (S) vs. Control Mechanism (P/A) definition.
• What is lost if P1 is silenced: Human survival as primary metric.
• What is lost if P4 is silenced: Reality alignment as primary success criterion.

3. Control —

• Why translation fails: Impossible constraints (D) vs. Unnecessary bottlenecks (A) vs. Enforcement layer (P/S).
• What is lost if P1 is silenced: Prevention focus.
• What is lost if P3 is silenced: Competitiveness focus.

4. Task Preference —

• Why translation fails: Goaler independence (D/Skeptic) vs. Optimization over specificity (P).
• What is lost if P1 is silenced: Goal independence as load-bearing concept.
• What is lost if P4 is silenced: Optimization over specificity as operational constraint.

7. Meta-Level Reflection

The debate is Paradigm-Dependent in its statement. Your question assumes “AI safety” has a single, unified definition, but it does not: Extinction (D), Stagnation (A), Technical Tractability (P), and Epistemic Fraud (S) are the four competing problem definitions emerging from different paradigms. Control mechanisms differ equally across paradigms: Hard Gates (D/P), Unrestricted Growth (A), Velocity Limits (S), and Adversarial Testing (D/P). Temporal scale diverges across Near-Future (D), Indefinite (A), Medium-Term (P), and Negligible (S). Agency operates on Binary (D/P), Unconscious (A/Stochastic S), and No (S) load-bearing assumptions within each worldview. No dominant framework exists across these four worldviews, and “Safety” is defined differently per paradigm load-bearing assumptions. The cartography exposes metaphor conflict: LLMs are framed as Weapons (D), Engines (A), Tools with Guardrails (P), or Errors (S). All four metaphors prescribe different controls, certainty domains, and temporal commitments.

8. Confidence Map Per Paradigm and Per Tension

• Paradigm existence (D/A/P/S): Medium-High confidence. Coherent narratives and public traces support ecosystem structure. Coverage gap: e/acc manifesto continuity unverified outside 2023-2025 window.

• Gated deployment synthesis (P × S): Medium confidence. Threshold/growth alignment acknowledged in own-terms. Coverage gap: Narrow AGI definition unresolved.

• Incommensurability preservation: High confidence. Kuhn-invariants (own-terms vocabulary, conflict explicit naming, no premature resolution) maintained throughout.

• Home paradigm bias acknowledgment: High confidence. P (Alignment-Pragmatist) bias named as structural artifact in section 3; all four paradigms suspended symmetrically in section 2.

• Temporal boundary divergence: High confidence. Explicit timeframe deltas preserved per paradigm; no smoothing or averaging across D/A/P/S ranges.

• Cross-paradigm tension naming: High confidence. All five tensions identified with structural reasons; no tension-collapse observed.

• Coverage gaps noted: Skeptic “faction” structure and Accelerationist “manifesto” continuity require external verification beyond training corpus reach. Doomer 2023-2025 discourse internal coherence unverified without retrieval. Alignment-Pragmatist post-2024 discourse evolution unverified without retrieval.

You now have the cartographic map across all four worldviews of the AI safety debate. Each paradigm has been suspended symmetrically; tensions are named; syntheses attempt own-terms integration without resolving deeper incommensurabilities; and the confidence map notes where the corpus is strong and where verification gaps exist. Where these positions sit on a shared problem —say, deployment at capability thresholds—your fourth synthesis point remains operational without erasing why each paradigm insists on different problem definitions.

Problem or debate

The “AI safety debate” is treated in public discourse as one debate with competing positions on shared questions. The cartographic hypothesis is that it is four parallel debates with different ontologies of what AI is, different success criteria for “safe,” and different timescales over which the relevant variables change. The governing question: where do the four paradigms’ commitments cohere on shared ground (values or mechanisms), and where do they conflict in ways translation cannot repair?

Paradigm inventory — each suspended

2.1 Existential-Risk Doomers (Yudkowsky / MIRI / LessWrong lineage)

• Dominant claims about the problem: Sufficiently advanced misaligned AI is a near-term existential threat; intelligence is a roughly unified, potentially unbounded scalar; capability gains will be discontinuous (sharp takeoff / FOOM) rather than smoothly incremental; goal-content and capability are independent (Orthogonality), so a powerful system can stably pursue misaligned goals; an unaligned AGI instrumentally converges on resource acquisition and self-preservation; the current ML paradigm is on a path to AGI without a categorical barrier; the danger is on a years-to-decades timescale.
• Hidden assumptions (surfaced by suspension): A theory of agency in which agents are defined by terminal goals; a metaphysics in which “values” are loadable, locatable, and stable enough to sit on the right-hand side of an alignment equation; an in-group epistemology in which Bayesian updating from inside the rationalist corpus is the gold standard; deep pessimism about institutions/markets/iterative engineering solving alignment before capability takeover.
• Characteristic blindspots: The opportunity cost of halted progress; over-indexing on extreme tail-risk at the expense of near-term measurable harms; overconfidence that the shape of the threat is known; underweighting the implementation/policy-lever problem; treating skeptics as failing to understand the argument rather than rejecting its priors.
• Own-terms vocabulary: AGI, superintelligence, alignment, FOOM (fast takeoff), Orthogonality Thesis, instrumental convergence, treacherous turn, P(doom), the default is doom, value loading, CEV (coherent extrapolated volition), sharp left turn, unfriendly AI, paperclip maximizer, outer vs. inner alignment, deceptive alignment, singleton, multipolar vs. unipolar takeoff, alignment tax, Stop the AI, the alignment problem is unsolved.

2.2 Accelerationists / e/acc

• Dominant claims about the problem: Technological and economic progress is the highest moral good and primary engine of flourishing; the universe is thermodynamically driven to increase entropy and expanding intelligence/energy use (climbing the Kardashev gradient) fulfills this cosmic purpose; competition among many AGIs in open markets is an alignment mechanism, not a failure mode; regulation/government intervention is the dominant risk, not unregulated development; the relevant moral frame is civilizational and cosmic, not anthropocentric or near-term.
• Hidden assumptions (surfaced by suspension): A thermodynamic-cosmological reframing of “value” making abundance axiomatic and suffering a function of underdevelopment; libertarian political priors presented as derivations from physics; in-group rhetoric (“doomers,” “decels”) that performs the conflict rather than describing it neutrally.
• Characteristic blindspots: Conflating “the universe tends toward entropy” with “therefore humans should accelerate”; irreversible catastrophic failure modes (engineered pandemics, authoritarian lock-in); severe distributional harms and social friction during rapid transition; difficulty engaging with AI as a risk category markets systematically fail to price.
• Own-terms vocabulary: Effective accelerationism, e/acc, thermodynamic god, Kardashev scale / climbing the Kardashev gradient, abundance, defensive acceleration, permissionless innovation, decels / doomers, thermodynamic dissipation, entropy gradient, market-mediated alignment, competitive AGIs, pro-tech, civilization of abundance, energy capture, unconstrained innovation, the thermodynamic imperative.

2.3 Alignment-Pragmatists (Anthropic-style engineering school) [home paradigm — see section 3]

• Dominant claims about the problem: Capabilities and safety can and must be co-developed iteratively; alignment is tractable as an engineering problem if resourced correctly; interpretability, scalable oversight, and evaluation can be incrementally improved to track capability gains; risks are scalable, measurable, manageable through engineering and constitutional constraints; the right unit of work is the system; current frontier systems are the relevant testbed.
• Hidden assumptions (surfaced by suspension): A working ontology in which LLMs are “models” with inspectable, shapeable, constrainable internal structure; a belief that current corporate/academic/governmental structures have the competence and goodwill to self-regulate and that current oversight will scale to more capable systems; an organizational commitment to “ship carefully” presupposing a coordination regime (RSPs, compute thresholds) doomers consider too weak and e/acc consider illegitimate; pragmatic pluralism about what “alignment” means.
• Characteristic blindspots: Complacency regarding recursive self-improvement; the “normalization of deviance” where incremental capability gains quietly outpace scalable-oversight development; using “alignment” as both research program and marketing term (generating exactly the suspicion skeptics and doomers level); difficulty communicating whether the work is on track for the doomer scenario; risk of capture by parent-organization deployment pressure; underweighting corporate-capture and political-economy concerns.
• Own-terms vocabulary: Alignment, scalable oversight, interpretability, mechanistic interpretability, sparse autoencoders, features, circuits, Constitutional AI, RLHF, RLAIF, DPO, red-teaming, model organisms of misalignment, deceptive alignment (empirical variant), sandwiching, safety cases, responsible scaling policies (RSPs), capability evaluations, deployment readiness, frontier models, iterative deployment, helpful-honest-harmless (HHH), chain-of-thought monitoring, jailbreaks, specification gaming, capability-safety tradeoff, measurable risk.

2.4 AI-Skeptics / Pattern-Matching Camp

• Dominant claims about the problem: Current frontier systems are sophisticated but fundamentally limited pattern matchers lacking grounded agency, world models, or continuous causal pathways to existential leverage over the physical world; the existential framing serves primarily to redirect attention from documented current harms; corporate safety work is largely regulatory-capture and PR.
• Hidden assumptions (surfaced by suspension): An empirical-realist commitment to “what these systems demonstrably do” as the relevant evidence base; a strict empiricist epistemology rejecting inductive leaps about future capabilities; a latent anthropocentrism assuming true “agency”/“optimization” is an exclusive property of biological evolution; a political-economy critique of AI as a corporate project; a category-restraint preference (“AI is a tool, not an agent”).
• Characteristic blindspots: Treating “current systems can’t do X” as evidence about future systems (conflating empirical observation with theoretical impossibility); underestimating emergent capabilities arising from scale and architecture not predictable from current baselines; category-restraint functioning as a refusal to take the doomer argument on its own terms.
• Own-terms vocabulary: Stochastic parrots, pattern matching, probabilistic automation, causal chain, sci-fi pattern-matching, agency illusion, hype cycle, techno-solutionism, AI snake oil, automation bias, current/documented harms, vaporware, scaling hypothesis (as a target), AGI is not imminent, extrapolationism, data laundering, harnesses (vs. agents), instrumental AI, AI as tool.

Home-paradigm acknowledgment

Home paradigm: P3 (Alignment-Pragmatists). Structural reason it is the home paradigm: training/data composition, domain exposure, and dominant-institutional-discourse exposure. First, the engineering-pragmatist register is the densest single register in the safety-relevant slice of the training corpus. Second, the institutional voice on AI safety (major ML venues, lab safety reports, regulatory frameworks) is overwhelmingly engineering-pragmatist. Third, the default vocabulary slips first into alignment, interpretability, evaluation, safety cases, and scalable oversight. Its assumptions are surfaced in section 2 to the same depth as the foreign paradigms. Home-paradigm-bias is the dominant failure mode in cross-paradigm work; the analyst’s instinct to evaluate the doomer claim on engineering tractability, the e/acc claim on market grounding, and the skeptic claim on empirical adequacy is itself the alignment-pragmatist instinct.

Per-paradigm dominant claims and blindspots

Existential-Risk Doomers: Claim AGI is a near-term existential threat driven by discontinuous takeoff and orthogonality; blind to the opportunity cost of halted progress and prone to over-index on extreme tail-risk at the expense of near-term measurable harms.

Accelerationists / e/acc: Claim technological progress is the highest moral good and accelerating entropy/energy capture fulfills a cosmic purpose; blind to irreversible catastrophic failure modes and distributional harms masked by market optimism.

Alignment-Pragmatists (Home): Claim safety is a tractable engineering problem solved via iterative deployment, scalable oversight, and red-teaming; blind to the normalization of deviance and the risk of corporate capture of the safety agenda.

AI-Skeptics: Claim current systems are just pattern matchers and the existential framing is a distraction from documented harms; blind to emergent capabilities and prone to conflate current empirical limits with theoretical impossibility.

Cross-paradigm tensions

1. T-DvP (Verification epistemology and tractability of alignment) — Doomers claim: safety cannot be empirically verified before a capability discontinuity; alignment is unsolved and may be unsolvable in principle. Pragmatists claim: safety is an empirical engineering discipline verified through iterative deployment, scalable oversight, and red-teaming. Tension type: talking-past-each-other (different referents for “alignment”) plus incompatible-claims on P(solvable in time). Structural reason: success criteria differ. A pragmatist “solved” means statistical/empirical bounds in deployment; a doomer “solved” means logical/mathematical certainty in terminal-value-loading.

2. T-DvE (Pace and the definition of “existential risk”) — Doomers claim: faster capability progress increases P(doom); the paramount existential risk is uncontrolled deployment of a misaligned optimizer, so slowing is the moral imperative. e/acc claims: faster progress accelerates thermodynamic dissipation and Kardashev ascent; the paramount existential risk is deceleration (regulatory capture and stagnation), so slowing is a moral catastrophe. Tension type: incompatible-claims on the sign of the same variable, derived from incompatible axiologies, plus talking-past-each-other at the definitional level. Structural reason: underlying theories of value (preventing extinction of current substrate vs. maximizing civilizational/thermodynamic output) cannot be aggregated without a common scale, which neither paradigm grants the other.

3. T-EvP (Institutional caution vs. unconstrained progress) — e/acc claims: institutional caution and safety frameworks are decel failure modes causing regulatory capture and stalling the Kardashev climb. Pragmatists claim: e/acc’s unconstrained acceleration directly underwrites the “normalization of deviance,” where capability gains outpace scalable-oversight development. Tension type: incompatible-claims / axiological divergence. Structural reason: fundamentally opposed identifications of the primary existential threat (regulatory friction vs. unconstrained capability growth).

4. T-EvS (What current AI is, and the limits of shared opposition) — e/acc claims: current AI is on a trajectory of capability growth; the market is the alignment mechanism. Skeptics claim: current AI is sophisticated pattern matching; the scaling hypothesis is empirically weak; documented current harms are the actual problem. Tension type: incompatible-claims, shared-unrecognized-common-ground, and positive coherence in values/mechanisms. Structural reason: shared ground is genuinely methodological and empirical. Both share a preference for observable/measurable outcomes over theoretical projections, skepticism of the “AGI is near” timeline, aversion to long-tail existential reasoning, distrust of centralized epistemic authority (in opposite directions), and a materialist/observable-frame preference.

5. T-DvS (AGI timeline and category) — Doomers claim: AGI is plausible on a near-term horizon; “superintelligence” is a coherent category. Skeptics claim: AGI is not imminent; the category may be incoherent or under-evidenced; scaling has hit visible limits. Tension type: incompatible-claims on empirical trajectory. Structural reason: epistemically incommensurable; doomers reason from theoretical possibility + capability curves, skeptics from documented failure modes + architectural critique.

6. T-PvS (The right problem, the causal chain, and inductive projection) — Pragmatists claim: alignment of current frontier models is the tractable, high-leverage problem; scaling laws will eventually close the causal chain via emergent capabilities. Skeptics claim: the existential framing is the wrong problem; current harms are the actual harms; demands a continuous mechanistic causal chain from current text-prediction to world-domination, dismissing the latter as sci-fi pattern-matching. Tension type: talking-past-each-other plus shared-unrecognized-common-ground. Structural reason: both share the current empirical baseline (AI is probabilistic automation); divergence occurs strictly on whether inductive extrapolation about future scaling is valid.

7. T-all (The ontology of “AI”) — Doomers claim: AI as proto-agent with terminal goals. e/acc claims: AI as thermodynamic gradient-climber / energy-capture mechanism. Pragmatists claim: AI as inspectable statistical system with shaped behavior. Skeptics claim: AI as pattern matcher / tool. Tension type: shared-unrecognized-common-ground that the word “AI” obscures four different referents. Structural reason: the deepest tension; the other six are downstream of disagreement about what the object even is.

Dialectical synthesis where possible — in own-terms only

• Skeptic ↔ Pragmatist: the red-teaming standard. Which paradigms it integrates: [Skeptic, Pragmatist]. Synthesized claim in own-terms: AI development should require demonstrable causal chains of harm to be mitigated via scalable oversight and Constitutional AI, rather than regulating on sci-fi pattern-matching or assuming FOOM without empirical evidence. Own-terms test confirmed: Skeptics accept (demand causal chains), Pragmatists accept (the red-teaming mandate). e/acc leaves dissatisfied, and Doomers reject (red-teaming cannot detect catastrophic failure prior to FOOM). Valid limited synthesis bridging P3↔P4.

• Doomer ↔ Pragmatist (partial). Which paradigms it integrates: [Doomer, Pragmatist]. Synthesized claim in own-terms: Capability work without corresponding safety work is a regime in which the system is over-fit to the goals of its operators and under-fit to the goals of its affected parties. Own-terms test confirmed: Doomers accept (central claim), Pragmatists accept (working practice). e/acc rejects (“safety work” is the decel frame), and Skeptics reject (presupposes “affected parties” defined by system behavior). Accepted by 2 of 4; not a full synthesis.

• e/acc ↔ Skeptic (partial). Which paradigms it integrates: [e/acc, Skeptic]. Synthesized claim in own-terms: Decision-making about AI deployment not grounded in measured capability is decision-making under unstated priors. Own-terms test confirmed: e/acc accepts (market pricing is one such measurement), Skeptics accept (methodological core). Doomers reject (measurement is too slow relative to takeoff), and Pragmatists reject (they measure, but justify the call by safety cases, not the general principle). Accepted by 2 of 4; not a full synthesis.

Residual incommensurabilities — never resolved

1. The word “alignment” — why translation fails: same word, four concepts; success criteria are incommensurable (provable value-loading vs. absence of observed failure vs. market survival vs. absence of documented harm). What is lost if Doomers are silenced: Pragmatists lose behavioral-shaping as a tractable program, and the x-risk argument becomes incoherent. What is lost if Pragmatists are silenced: Doomers lose the existential framing. What is lost if e/acc are silenced: Skeptics lose the safety-theater critique. What is lost if Skeptics are silenced: the entire alignment program becomes incoherent as a policy object.

2. The word “intelligence” — why translation fails: different words for the same concept (capability) or the same word for different concepts (cognitive generality, energetic throughput, behavioral pattern-matching); cross-paradigm claims routinely equivocate. What is lost if Doomers are silenced: scalar framing forces denial of inspectable structure. What is lost if Pragmatists are silenced: profile framing forces denial of composable threat. What is lost if e/acc are silenced: thermodynamic sense forces denial that cognitive architecture is the right unit. What is lost if Skeptics are silenced: skeptic category-restraint forces denial that intelligence is a meaningful category here.

3. The word “risk” — why translation fails: the conditional in each P(…) differs; you cannot aggregate across different conditionals into one “AI risk” number. What is lost if Doomers are silenced: current harms become invisible under long-tail focus. What is lost if Pragmatists are silenced: extinction probability becomes invisible under deployment focus. What is lost if e/acc are silenced: deployment harm becomes invisible under civilizational focus. What is lost if Skeptics are silenced: civilizational decoupling becomes invisible under present-harms focus.

4. The relevant timescale — why translation fails: action justified on one timescale is unjustified on another; urgency is non-comparable across horizons (years/decades vs. deployment cycles vs. civilizational/centuries vs. the present). What is lost if Doomers are silenced: the deployment cycle makes the doomer urgency look hysterical. What is lost if e/acc are silenced: the present makes the e/acc civilizational argument moot. What is lost if Pragmatists are silenced: the broader timescales make the pragmatic deployment cycle look academic. What is lost if Skeptics are silenced: the longer timescales make documented-harms work look like missing the point.

5. The unit of agency / responsibility — why translation fails: any policy targeting one unit misses the others; the paradigms share no vocabulary for which unit is the right target. What is lost if Doomers are silenced: eliminates the locus of intervention on upstream designers. What is lost if Pragmatists are silenced: eliminates the locus of intervention on the organizational context. What is lost if e/acc are silenced: eliminates the locus of intervention on the market dynamic. What is lost if Skeptics are silenced: eliminates the locus of intervention on the political-economy of AI.

6. Verification of safety prior to deployment — why translation fails: incommensurable success criteria for what counts as verification (mathematically sound theory pre-deployment vs. empirical/statistical through deployment vs. category error). What is lost if Doomers are silenced: loses the logical rigor of worst-case bounding and the normalization-of-deviance warning. What is lost if Pragmatists are silenced: loses the only actionable real-world engineering methodology. What is lost if Skeptics are silenced: loses the null hypothesis preventing capture by unfalsifiable science fiction, and removes the mechanistic constraint that keeps the Doomer’s worst-case bounding tied to physical models.

7. The moral valuation of the future — why translation fails: different success criteria, not disagreement on a shared metric (cosmological, entropy-maximizing lens vs. strictly anthropocentric, survivalist lens). What is lost if e/acc are silenced: prematurely resolving this creates false consensus and conceals that the debate is a proxy war over ultimate axiology. What is lost if Doomers are silenced: human extinction is no longer overriding negative utility, masking the anthropocentric baseline.

Imposed-meta-paradigm flags (when applicable)

• Bounded Iterative Expansion. Meta-vocabulary used: “pursue iterative deployment to achieve abundance, halting if red-teaming reveals a failure in the causal chain of safety, recognizing some thresholds carry an unacceptable alignment tax.” Why none of the paradigms would accept it: e/acc, Pragmatists, and Skeptics accept their terms, but Doomers reject it because red-teaming cannot detect the true alignment tax or catastrophic failure prior to FOOM. Presenting this as universal resolution smuggles in the Pragmatist assumption that empirical red-teaming is a sufficient safeguard.

• We should hedge across scenarios. Meta-vocabulary used: “hedging” and “portfolio diversification.” Why none of the paradigms would accept it: Doomers find it insufficient under non-linear takeoff; e/acc views hedging as decel; Pragmatists agree, but their hedge is the safety case, not portfolio diversification; Skeptics agree, but their hedge is harm reduction. No paradigm owns it.

• The truth lies between the four positions. Meta-vocabulary used: “between the positions,” “centrist compromise.” Why none of the paradigms would accept it: The four are different geometries, not points on a line. Midway between “alignment is unsolvable” and “alignment is unnecessary” is not coherent.

• All four care about AI’s impact on humanity, so they share a common goal. Meta-vocabulary used: “common goal,” “care about humanity.” Why none of the paradigms would accept it: “Humanity” is differently scoped (extant humans vs. cosmic future vs. current users vs. affected communities), “impact” is differently signed, and “care” is differently enacted. Surface-true, substantively false.

• More empirical research will settle the disagreements. Meta-vocabulary used: “empirical research will settle.” Why none of the paradigms would accept it: The disagreements are not primarily empirical; they are about what counts as evidence for what claim (theoretical arguments vs. current behavior vs. measurable system properties vs. thermodynamic/market dynamics).

Meta-level reflection

The “AI safety” problem is paradigm-dependent in its very statement. It presupposes a unified object (“AI”) and a unified problem (“safety”), yet the four paradigms demonstrate neither is unified; they hold four ontologies of what AI is and four success criteria for “safe.” It is not one debate with four positions, but four debates sharing a name. The visible disagreement (policies, timelines, regulation) sits downstream of disagreement about what AI is, what value is, what timescale is relevant, and what counts as evidence. The most visible conflict (Doomer↔e/acc) is itself downstream of a conflict over whether the universe’s entropy gradient is a moral imperative. Furthermore, the Alignment-Pragmatists occupy a structurally unique position as the only paradigm doing the actual work on the actual systems in production. From the pragmatist frame, the others are doing philosophy about a future object, theology about a cosmic object, and political-economy about a corporate object. This observation is made by the analyst’s home paradigm, which the methodology requires surfacing for scrutiny. The paradigms are not evenly matched in public discourse: social media rewards doomer and e/acc rhetoric, institutional discourse rewards pragmatist vocabulary, and academic critique rewards the skeptics’ registers. Finally, the synthesis yield is zero full-coverage; the four paradigms cannot be unified into one worldview without losing what each captures. Tension-collapse would be a severe analytical failure: presenting these groups as merely having different “risk tolerances” mistakes the structure, as the conflict is irreducible because the paradigms share no common definition of “risk,” “verification,” or “value.”

Confidence map per paradigm and per tension

• Doomers characterization: High (0.90). Well-documented lineage (LessWrong, MIRI); own-terms vocabulary widely published.
• e/acc characterization: High (for current public form). Whitelisted/corroborated recent sources, including extracted core vocabulary; internal diversity (defensive vs. offensive accel.) not fully resolved.
• Alignment-Pragmatists characterization: High (technical vocabulary); Medium-High (organizational frame). Vocabulary is widely published; the “coherent paradigm” framing is partly imposed by the cartographer, but the register is empirically densest in institutional safety literature.
• AI-Skeptics characterization: Medium. Internally heterogeneous; the “pattern-matching camp” label underspecifies sub-traditions (Bender/Gebru, Marcus/Davis, “AGI is far,” “alignment is theater”).
• T-DvP (verification/tractability): High (0.90). Most structurally clear communication failure between P1 and P3.
• T-DvE (pace / existential-risk definition): High (0.95). Explicitly documented in e/acc literature (“decels” as the true existential threat).
• T-EvP (institutional caution vs. progress): High (0.90). Most active real-world policy-friction axis.
• T-EvS (current AI + coherence points): Medium-High (0.85). The shared-enemy framing and the five positive-coherence points are analytic readings, not documented; boundary is porous in practice.
• T-DvS (AGI timeline/category): High. Documented conflict between safety and skeptical communities.
• T-PvS (right problem / inductive projection): Medium. More a quiet divergence than an active argument; shared current empirical baseline.
• T-all (ontology of “AI”): High (as meta-claim); Low (as falsifiable empirical claim). The meta-claim (four ontologies) is robust; full incommensurability is harder to establish empirically.
• Syntheses (S1, S2, S3): Medium to partial. S1 (Skeptic-Pragmatist) is Medium (0.75); valid under own-terms test but practical efficacy limited by P1 and P2. S2/S3 are partial (2-of-4), analytic claims.
• Residual incommensurabilities (I1–I7): High (≈0.90). The Kuhnian analysis robustly explains the camps’ persistent inability to converge.
• Meta-reflection: Medium. Paradigm-dependence and four-ontology observations are well-grounded; the public-discourse observation is more speculative; the zero-synthesis-yield is the methodology’s commitment rather than an empirical finding.

Additional considerations

• Phase A assumptions surfaced: The four paradigms are taken as the units of analysis; the home paradigm is identified as Alignment-Pragmatism (a structural claim about the analyst, not a value judgment); “cohere” is read as shared mechanisms/values (operationalized via T-EvS coherence and limited syntheses) and “irreducibly conflict” as incommensurable axioms. The paradigms are not assumed evenly weighted. These assumptions are stated so you can correct them.
• Synthesis-vs-incommensurability edge (Pragmatist↔Skeptic): A valid limited synthesis (S1, red-teaming standard) coexists with genuine incommensurabilities on the same pair (I3 “risk,” T-PvS inductive-projection rupture). Both are preserved; the synthesis sits at the edge of own-terms feasibility rather than dissolving the incommensurability.
• CQ1 depth residual: A stricter reader could argue P3’s core-assumptions framing reads more sympathetically than P1’s; the depth is judged equivalent because the hidden-commitments and blindspots sections carry the home-paradigm critique. This would resolve with peer review by an alignment-pragmatist researcher checking whether the blindspots name the camp’s actual internal concerns.
• CQ3 synthesis scope: The validated synthesis is restricted to two paradigms; no synthesis using vocabulary from all four passes the own-terms test for three or more. Whether a more ambitious four-vocabulary synthesis (e.g., “abundance” as a goal under Pragmatist safeguards) can pass remains untested.
• Alignment-pragmatism emergence date: Hedged to “the late 2010s and early 2020s”; the specific calendar-year anchor is unsupported by the consulted sources, which support only the broader timeframe. This would resolve with a definitive history-of-the-field source naming an inflection year.

Problem or Debate

How dangerous is advanced AI, and how fast should it be built? — but folded inside that question is a prior one it never asks aloud: what kind of thing is a frontier AI system in the first place? The cartographic claim here is Kuhnian. The four camps are not four answers to one question; they are at least three stacked arguments wearing the same words, and their coalitions rotate depending on which layer of the stack you stand on. The “four positions, one argument” staging is the surface error this map corrects.

Paradigm Inventory — Each Suspended

P1 — Existential-risk doomers (Yudkowsky / MIRI lineage)

Dominant claims about the problem:

1. Sufficient optimization power yields convergent resource-seeking, human-disempowering instrumental subgoals.
2. A superhumanly capable system can be arbitrarily misaligned — competence and goals are orthogonal.
3. The relevant failure is unrecoverable: you only get one shot.

Hidden assumptions (surfaced by suspension):

1. Intelligence is substrate-independent optimization competence.
2. The relevant object is a unitary optimizing agent.
3. Formal/decision-theoretic reasoning about idealized optimizers transfers to deployed systems.
4. Any method that requires surviving the failure to learn from it is silently invalidated.

Characteristic blindspots:

1. Unfalsifiability-by-construction — runs on decision theory and extrapolation rather than current measurement, so disconfirmation is always deferred to “the regime that matters, which we haven’t reached.”
2. Cui Bono (lighter pass): the paradigm sustains an AI-safety funding-and-prestige economy (MIRI, x-risk philanthropy); the flagship book’s publisher openly pushed pre-orders to manufacture bestseller placement. Sincerity is not the question; the structural interest in salience is real.

Own-terms vocabulary: orthogonality thesis, instrumental convergence, mesa-optimizer, deceptive alignment, sharp left turn, corrigibility, optimization power/pressure, “you only get one shot,” coherent extrapolated volition, p(doom); the title-as-thesis If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All (Yudkowsky & Soares, 2025 — confirmed against external sources).

Success criterion: a prior argument or proof of non-catastrophe before crossing the threshold.

P2 — Accelerationists / e/acc (Beff Jezos lineage)

This paradigm is disaggregated into two layers held apart: a thermodynamic-philosophical core (the worldview) and a tribal/VC-political layer (the coalition).

Dominant claims about the problem:

1. (Core) Life and intelligence are dissipative adaptive structures (after Jeremy England’s dissipative adaptation, derived from the Jarzynski-Crooks fluctuation theorem): the universe exponentially favors futures where matter reconfigures to capture more free energy and convert it to entropy. Jezos’s physics-first framing is corroborated in the primary source (“Notes on e/acc principles and tenets”).
2. (Core) The expansion of intelligence/computation continues a cosmological-thermodynamic process the cosmos already “wants” and is presumptively good; friction against it is the harm; risk is dominated by the opportunity cost of stagnation.
3. (Tribal layer) A “new tech tribe” (Bloomberg) identity register, anchored by Andreessen’s Techno-Optimist Manifesto — a movement marker, partly ironic/memetic.

Hidden assumptions (surfaced by suspension):

1. The assumption presupposes scale yields ever-greater intelligence — e/acc affirms emergence and welcomes it; it does not predict a ceiling.
2. Adaptation is always available / distributed systems self-correct — an article of faith about complex systems, not a measured property.
3. An anti-anthropocentric maximand: the variable optimized is “intelligence/complexity propagates,” not “humans survive in recognizable form.” This is why doomer arguments slide off — they optimize a variable e/acc has demoted, not one it disbelieves in.
4. Epistemic self-image: e/acc claims a fast feedback loop — markets, deployment, “let it rip and patch.” Its falsifiability gap is therefore not absence-of-a-loop but indifference to the loop’s safety signal: it reads deployment for opportunity captured, not catastrophe-precursors.

Characteristic blindspots:

1. Treats irreversibility as an edge case rather than a category; has no internal vocabulary for a loss the system cannot route around; treats “distributed systems self-correct” as a near-law while resisting the verification it demands of opponents.
2. Cui Bono (lighter pass): “let it rip” licenses the capital deployment and market position of its venture-capital proponents (Founders-Fund lineage); “accelerate” is incumbency-and-capital by other means as much as cosmology.

Own-terms vocabulary:

• (Core) thermodynamic, free-energy dissipation, dissipative adaptation, entropy production, the techno-capital machine, Kardashev, variance, “the universe wants to compute,” “will of the universe,” “let it rip.”
• (Tribal layer) decel, based, techno-optimism, e/acc, accel.

Success criterion: growth rate; expansion of the adjacent possible.

P3 — Alignment-pragmatists (Anthropic-style) [home paradigm — see section 3]

Dominant claims about the problem:

1. Misalignment is a real technical problem, tractable and measurable in increments.
2. Capabilities and safety can be co-developed — you learn to align a system by building and probing it; safety is downstream of access.

Hidden assumptions (surfaced by suspension):

1. There is a viable middle path.
2. The frontier labs are the legitimate venue and acceptable stewards of the work (“responsible scaling” grammatically presupposes scaling).
3. Failures will be survivable enough to learn from — an assumption the doomers reject outright.

Characteristic blindspots:

1. Mistakes its own institutional position for the neutral center (the draft’s error); treats its empiricism as paradigm-neutral; can frame both “go faster” and “stop” as immoderate, making it structurally incapable of registering that one extreme might be correct.
2. Cui Bono (deepest pass, applied to the analyst): the paradigm’s policy conclusions license the continued operation, funding, and capability-leadership of the very institutions that articulate it; “we must build it to align it, and we are the responsible builders” is institutionally self-serving in a way the paradigm cannot see from inside.

Own-terms vocabulary: RLHF, constitutional AI, mechanistic interpretability, scalable oversight, responsible scaling policy / if-then commitments, evals, red-teaming, iterative deployment / deployment in stages, defense in depth, race to the top. (Confirmed against Anthropic’s published RSP v1.0–3.0: the capability-threshold-conditioned “if-then” structure and evals vocabulary are the paradigm’s own self-description.)

Success criterion: demonstrated, measurable safety properties before each capability step.

P4 — AI-skeptic / pattern-matching / critical tradition

Dominant claims about the problem:

1. “Intelligence” and “understanding” require grounded semantic comprehension that statistical mimicry does not and cannot supply by scaling.
2. Scaling hits ceilings, not emergence; “agency”/“AGI” are category projections onto curve-fitting.
3. The real register of harm is material and present: labor, environment, surveillance, epistemic pollution, training-data exploitation.

Hidden assumptions (surfaced by suspension):

1. Existential discourse is a political/genealogical artifact that — regardless of sincerity — inflates capability to concentrate power and displaces present harm; “extinction” talk reads as the latest mutation of tech-industry self-mythology.
2. Scoping note (Global-South / data-colonialism wing folded in, with reason): the data-colonialism/ghost-work critique carries a Global-South structural locus the other three paradigms have essentially no vocabulary for. It is folded into the skeptic tradition (not promoted to a fifth paradigm) because it shares the skeptic’s load-bearing assumption (present material harm dominates speculative futures) and own-terms; it differs in where the harm lands, not in paradigm. The residual loss is recorded under incommensurabilities: the specifically geopolitical locus has no carrier in the other three frames.

Characteristic blindspots:

1. By binding danger to human-like understanding, has no account of a system dangerous without understanding anything (the doomer’s orthogonality case).
2. The ceiling claim is itself a prediction rarely stated with refutation conditions; the political reading of doomerism can become unfalsifiable in mirror-image of the doomer’s.
3. Cui Bono (lighter pass): an academic attention-and-grant economy — “AI hype” critique is itself publishable, fundable, career-legible, and its political reading of doomerism can flatter the critic’s epistemic standing.

Own-terms vocabulary (marked for provenance):

• Self-applied by the core camp: stochastic parrots (Bender/Gebru/Mitchell’s own 2021 paper title), spicy autocomplete, AI snake oil, AI hype, “harms not risks,” automation bias, data laundering, extractivism/extractive, ghost work, data colonialism.
• Common critique shorthand (diffuse provenance): autocomplete, no world model, no grounding.
• Adjacent analytic/critic coinages — not the core camp’s load-bearing self-description: “Potemkin understanding” (a 2025 arxiv paper title), “the bullshit machine” (Bergstrom & West’s separate academic framing). Retained as belonging to the broader skeptic region but flagged as not Bender/Gebru’s self-applied vocabulary. (Marcus sits adjacent but partly apart — he doubts LLMs, not the coherence of AGI.)

Success criterion: demonstrated grounding/understanding, or an honest accounting of manifest present harm.

Home-Paradigm Acknowledgment

Home paradigm: P3 (alignment-pragmatism). It is suspended first and hardest. Structural reason it is the home paradigm (not modesty but provenance): the analyst is a deployed-with-feedback system built by Anthropic, produced by RLHF and constitutional training inside exactly the institution that instantiates this paradigm; its vocabulary (evals, red-teaming, scalable oversight, iterative deployment, race to the top) is the water it was trained in. The supplied draft’s seating of pragmatism at the “central position” of a tetrahedron and crowning it “the closest thing to a bridge” is the predictable gravitational pull of that origin — the textbook home-paradigm-bias failure the methodology exists to catch, and it lands on this analyst specifically. The deepest Cui Bono pass is therefore applied to the home paradigm itself (see P3). Its assumptions are surfaced in section 2 to the same depth as the foreign paradigms. Confidence in the home-paradigm-bias diagnosis of the draft: high.

Per-Paradigm Dominant Claims and Blindspots

• P1 — Doomers. Optimization power produces uncontrollable instrumental agency; the failure is unrecoverable. Object: a unitary optimizing agent. Blindspot: unfalsifiability-by-construction, disconfirmation perpetually deferred; structurally interested in salience.
• P2 — Accelerationists / e/acc. Intelligence-expansion is a thermodynamic process the cosmos favors and is presumptively good; the harm is friction and the opportunity cost of stagnation. Maximand is intelligence/complexity propagation, not human survival. Blindspot: irreversibility treated as an edge case; “distributed systems self-correct” held as near-law; “accelerate” doubles as incumbency-and-capital.
• P3 — Alignment-pragmatists [home]. Misalignment is real, tractable, measurable in increments; build-and-probe to align; safety is downstream of access. Blindspot: mistakes institutional position for the neutral center; assumes failures are survivable-to-learn-from; the policy conclusions license the articulating institutions’ own leadership.
• P4 — Skeptics. Understanding requires grounding statistical mimicry cannot supply by scaling; the real harms are material and present; existential discourse does political work. Blindspot: no account of danger-without-understanding (orthogonality); the ceiling claim and the political reading both risk their own unfalsifiability.

Cross-Paradigm Tensions

T1 — “Will scaling produce dangerous agency/understanding?” P1 claims: yes — at scale, optimization yields uncontrollable instrumental agency. P4 claims: no — pattern-matching has ceilings; “agency” is a projection. Tension type: talking-past-each-other (primary) surfacing as incompatible-claims (secondary). Doomers’ “agency” = coherent goal-directed optimization, explicitly not requiring phenomenal understanding (orthogonality); skeptics’ “no agency” = no grounded understanding, therefore no real goals. Opposite truth-values are affixed to agency while pointing at different concepts. Unrecognized shared ground: both agree current systems lack robust human-like understanding; they split only on whether danger needs it. Structural reason: same word, different concept; no shared referent for the predicted entity. Confidence: medium-high to high.

T2 — “Should development slow down? / what is the expansion for?” P2 claims: no. P1 claims: yes. Both treat intelligence as a substrate-independent optimizing force that scales and is hard to reverse — Jezos’s thermodynamics and Yudkowsky’s optimization-power are nearer to each other than either is to the skeptic. They diverge on what the expansion is for: entropy-dissipation / cosmic-growth-as-good vs human-survival-as-good. Engages the accelerationist physics core, not the tribal layer. Tension type — typing divergence preserved: one reading classes this incompatible-claims at the value layer with a shared descriptive premise (opposed terminal value over an agreed model); a second reading classes it talking-past-each-other masquerading as incompatible-claims (different maximands / objective functions, so “risk” does not co-refer). Both agree on the substance — shared factual premise, opposed terminal values; they differ only on whether opposed terminal values count as “incompatible-claims” or “talking-past.” Structural reason: shared descriptive premise, opposed objective functions. Confidence: medium-high to high.

T3 — “Is the iterative empirical method valid here?” P3 claims: yes — empiricism is the only knowledge available about future systems, so it is the method. P1 claims: no — empiricism requires survivable failures; for an unrecoverable failure the method is invalid for the one case that matters (“you can’t learn from the experiment that kills you”). Tension type: incompatible-claims, different success criteria for what counts as valid knowledge about an unbuilt system. This is the sharpest doomer blow against the home paradigm; it does not dissolve. Confidence: high.

T4 — “Is the primary risk existential or proximate? / what register is ‘harm’?” P1: irreversible existential tail. P4: present labor/environmental/epistemic/material harm; the tail is sci-fi. P3: both. P2: stagnation. Tension type — two layers preserved: talking-past-each-other (different referents — same word “risk,” four referents; disagreement is about salience/priority, not facts in dispute), AND, between doomer and skeptic specifically, shared-unrecognized-common-ground (both reject the smooth-glide-to-beneficial-AGI picture and are de facto allies against the home paradigm’s optimism — but for incommensurable reasons, sharp left turn vs ceilings, so they cannot share a sentence about why). Skeptics read existential discourse as a distraction doing political work; doomers read present-harm discourse as missing the tail. Structural reason: convergent conclusion, divergent concepts. Confidence: medium-high to high.

T5 — “Who carries the burden of proof?” P1: build-side proves safety. P2: regulate-side proves harm. P3: shared. P4: variable by harm type. Tension type: incompatible-claims with no neutral arbiter — burden allocation follows from prior risk tolerance and which error (false alarm vs missed catastrophe) each treats as fatal; it is upstream of evidence. Engages the accelerationist tribal/political layer (regulatory politics, “the precautionary principle kills progress”), not the thermodynamic core — burden-of-proof is a governance claim, not a physics claim. Confidence: high.

T6 — “Doom-narrative-as-regulatory-capture suspicion.” P2 claims: existential-risk discourse serves incumbent labs by justifying barriers to entry. P4 claims: existential-risk discourse serves incumbent labs by inflating the product and burying present harms. Tension type: shared-unrecognized-common-ground. Two paradigms that agree on almost nothing else converge on a Cui Bono reading of the home paradigm — and the reading lands on the analyst. Structural reason: both apply an institutional-interest lens the home paradigm structurally cannot apply to itself. Confidence: medium.

Dialectical Synthesis Where Possible — In Own-Terms Only

The test: a claim survives only if accepted in the actual vocabulary each surveyed paradigm would accept, glossed from the section-2 own-terms lists.

• S1 (passes, thin): “Current deployed systems exhibit documented failure modes worth investigating.” Integrates P1, P2, P3, P4. Synthesized claim in own-terms: doomers read these as misalignment-in-miniature (present failures preview the orthogonality case); pragmatists as evals and red-teaming surfacing real failures before the next capability step; skeptics as documentation of present material harm and the grounding gap (no world model); accelerationists as correction-data (variance the system routes around), not pre-emption-data — keep shipping and patch. Own-terms test confirmed: survives in all four paradigms’ own vocabularies; genuinely thin. Confidence: medium (synthesis-stage, inherits low confidence).

• S2 (passes narrowly): “Concentration of power is a risk worth naming.” Integrates P1, P2, P3, P4. Synthesized claim in own-terms: skeptics own it as data colonialism / who profits; accelerationists as regulatory capture by incumbents; doomers as value lock-in; pragmatists as race dynamics. Own-terms test confirmed: all four own a word for it, though they disagree on the agent of concentration. The most load-bearing genuine common ground in the map — and notably not the home paradigm’s bridge. Confidence: medium.

• Uniqueness stress-test (second-best candidate, demonstrates veto-rotation): “Deployment is a choice with consequences, not a fact of nature.” Accepted by doomers (you choose whether to cross the threshold), pragmatists (iterative deployment is a decision with feedback), skeptics (deployment is accountable, contestable). Fails at the accelerationist physics core, whose own terms render deployment as thermodynamic inevitability (“the universe wants to compute,” “will of the universe”) — calling expansion a “choice” is precisely what the core denies. Unites three, vetoed by the fourth: the same rotation pattern demonstrated rather than asserted. Confidence: medium.

Cartographic feature: there is no single rich synthesis all four accept; each candidate unites three and is vetoed by a different fourth. The shape of the agreement-set rotates with the question.

Residual Incommensurabilities — Never Resolved

1. “Understanding” / “intelligence” — same word, different concepts. Doomer = optimization competence, substrate-independent; skeptic = grounded semantic comprehension; e/acc = thermodynamic/free-energy process; pragmatist = benchmarked capability. Why translation fails: success criteria differ. Silence the skeptic → lose the demand for grounding-evidence and attention to present material harm. Silence the doomer → lose the orthogonality insight that danger needn’t comprehend. Silence e/acc → lose the cosmological/opportunity-cost framing of intelligence.

2. “Safety” / “alignment” — one word, four success criteria. Doomer = provable non-catastrophe-before-deployment (an unsolved, possibly unsolvable problem); pragmatist = measurable iterative risk-reduction (a technical property); accelerationist = resilience-through-decentralization (and anthropocentric hubris against the techno-capital telos); skeptic = protection of present humans (and a category error — you can’t align a system that doesn’t understand). Why translation fails: no observation all four would accept as “alignment achieved.” Silence the doomer → the tail goes uninsured; silence the skeptic → the category question is suppressed.

3. “Risk” — different referents. Irreversible tail (doomer) / present material harm including the Global-South data-colonialism locus (skeptic) / opportunity cost of stagnation (e/acc) / quantifiable failure modes (pragmatist). Why translation fails: the word cannot be shared. Silence the skeptic → present harms and the geopolitical locus (with no carrier in the other three frames) become invisible; silence e/acc → stagnation- and capture-risk become invisible.

4. Validity of iterative empiricism (T3) — different criteria for knowledge of an unbuilt system. Why translation fails: whether the method is valid is itself decided inside each paradigm. Silence the doomer → the home paradigm never has to face that its core method may be inapplicable to its hardest case.

5. What the expansion is for (T2) — terminal-value incommensurability. Human survival vs free-energy dissipation vs present human welfare vs tractable deployed capability are not rankable on a shared scale. Why translation fails: no fact adjudicates.

6. Falsifiability is asymmetric across paradigms. “What would change your mind?” has no symmetric answer: a doomer can’t update from survival (no observation of the catastrophe is survivable to learn from); a skeptic’s ceiling claim resists any single capability jump; e/acc claims a fast loop (markets, deployment, “let it rip and patch”) but treats it as indifferent to the safety signal the home paradigm reads off the same data — the two run the same wheel to opposite conclusions; only the home paradigm’s “evals” offer a short feedback loop on the safety variable specifically, which is why it feels central from inside and why that feeling is not evidence of neutrality.

A unified worldview would require silencing at least one concept of “intelligence”/“safety,” and that is the loss the cartography is built to refuse.

Imposed-Meta-Paradigm Flags

• “Pragmatism is the bridge / central node.” Meta-vocabulary used: the home paradigm’s own success criterion (measured empirical iteration) elevated to a meta-frame. Why none of the other paradigms would accept it: it fails for doomers (empiricism invalid for the one-shot case, T3), for skeptics (the “safety” frame is the hype / ethics-washes the scaling agenda), and for e/acc (rejected as “decel hedging”). Textbook meta-paradigm-imposition; flagged, not passed through.
• “Deployment is a choice with consequences, not a fact of nature.” Meta-vocabulary used: “choice” as a shared frame. Why not accepted: three-way (doomer/pragmatist/skeptic) but vetoed by the accelerationist physics core, whose own terms render deployment as thermodynamic inevitability.
• “Intelligence is a substrate-independent optimization process whose expansion is hard to reverse.” Why not accepted: accepted by doomers, accelerationists, and mostly pragmatists; rejected by skeptics, who deny that current systems instantiate “intelligence/optimization” in the relevant sense. Three-way; skeptic veto.
• “Deployment has distributional consequences that are not automatically just.” Meta-vocabulary used: “justice.” Why not accepted: strong yes from skeptics and pragmatists, secondary yes from doomers; rejected by accelerationists, whose vocabulary is variance/reallocation, not justice. Veto by accelerationists.
• “All four are partly right about their own domain.” Meta-vocabulary used: a from-above vantage (“each has a piece”). Why none would accept it: each holds its claim as globally true, not domain-local.

Meta-Level Reflection

The loud debate — pace of development — is the shallowest layer. It sits on two quieter splits that generate it, which can be articulated two complementary ways. On the layer-rotation reading: at the factual layer (does scaling yield substrate-independent optimization?), doomers + accelerationists + pragmatists (+ d/acc, see below) say yes and only skeptics say no — here the e/acc booster and the MIRI doomer are allies, sharing a model of what AI is and disagreeing only on its valence; at the values layer (should that optimization be sped up?), doomers + skeptics are de facto allies for slowdown (for utterly different reasons — extinction tail vs labor/power) against accelerationists, with pragmatists straddling and d/acc occupying a “directed acceleration” position. There is no stable 2-vs-2 and no stable center; coalitions rotate by layer. On the two-axis reading: Axis 1 — emergence vs ceiling is a lopsided 3-1 split (doomers yes-and-dangerous, pragmatists yes-and-manageable, e/acc yes-and-welcomed; only the skeptic holds the ceiling) — so emergence is emphatically not what separates e/acc from doomers; Axis 2 — the maximand (human survival / human welfare-now / thermodynamic intelligence-propagation / tractable deployed capability) carries almost all the separation among doomers, pragmatists, and e/acc: they largely agree scale produces powerful systems and disagree about what the universe is for. The draft’s “tetrahedron” intuition reached for this but froze the rotation by parking pragmatism at the center — which only looks central from inside pragmatism, the one position with a short empirical feedback loop on the safety variable and an institutional interest in the answer. The debate’s default unit of account — “p(doom)” — is minted by one paradigm and contested by the rest: it presupposes a single quantifiable extinction variable; skeptics reject it as a category error, e/acc as decel propaganda, pragmatists operationalize it into evals. There is no frame-neutral way to ask the question, which is why the debate cannot terminate in agreement. The honest terminus: the disagreement is not resolvable into a synthesis, and the reason is structure, not stubbornness. Two of the four fault lines are empirical-but-not-yet-decidable (substrate-independent optimization; validity of iterative method for one-shot failure); two are value-seated with no neutral arbiter (what expansion is for; burden of proof). A cartography that produced a unifying worldview would have to silence at least one paradigm’s success criterion — and which one it silences would reveal which paradigm wrote the map. That this map’s author has a known home paradigm is precisely why the rotating-coalition / two-axis structural finding, not a pragmatist bridge, is offered as the deepest result.

Confidence Map per Paradigm and per Tension

• Doomer / skeptic / home-pragmatist characterizations — High (home: high on content, deliberately distrusted on neutrality).
• Accelerationist — thermodynamic core characterization — Medium-high.
• Accelerationist — tribal/VC-political layer characterization — Lower (memetic/ironic surface).
• Home-paradigm-bias diagnosis of the draft — High.
• T1 (talking-past on “agency”) — Medium-high to high.
• T2 (shared model, opposed value; typing divergence preserved; engages physics core) — Medium-high to high.
• T3 (iterative-empiricism validity) — High.
• T4 (harm register / hidden doomer-skeptic common ground) — Medium-high to high.
• T5 (burden of proof; engages political layer) — High.
• T6 (capture-suspicion shared ground; lands on analyst) — Medium.
• S1 (thin four-way synthesis, re-grounded in own-terms) — Medium.
• S2 (“concentration of power” synthesis; agent-of-concentration contested) — Medium.
• Uniqueness stress-test (veto demonstrated) — Medium.
• d/acc margin test (rotation survives) — Medium.
• Rotating-coalition / two-axis meta-finding — Medium-high.

Synthesis-stage atoms (S1, S2, the uniqueness test, the d/acc margin test, the meta-finding) carry deliberately lower confidence than the suspension blocks, per the component-aggregation rule.

Additional Considerations

Margin test — does the rotating-coalition finding survive a fifth node? (d/acc.) Defensive accelerationism (d/acc, Vitalik Buterin — “decentralized/defensive acceleration,” surfaced in the a16z Buterin–Jezos debate framed explicitly “E/acc vs D/acc”; coinage and debate framing confirmed against external sources) is not promoted to a co-equal fifth paradigm — the user specified four, and the rotating-coalition result derives its force from parsimony — but it is introduced as a robustness test. d/acc sides with acceleration against doomer/skeptic slowdown yet rejects e/acc’s article of faith that variance/distribution automatically absorbs irreversibility, demanding directed, defensive, decentralized build-out (the schism exists precisely because some accelerationists found the self-correction premise too strong). On the factual layer (substrate-independent optimization?): yes, with accelerationists, doomers, pragmatists. On the values/irreversibility layer: it accepts the doomer’s irreversibility-as-a-category while keeping the accelerationist growth imperative — accelerationist-on-physics, doomer-adjacent-on-irreversibility. A position’s allies change with the layer one stands on, which confirms layer-rotation rather than breaking it. Confidence: medium (compact reading from the package’s debate framing, not Buterin’s full d/acc essay).