Doom Analysis

Author: Mike Solana URL: https://piratewires.substack.com/p/doom-analysis One-line: A detailed critique of the two most prominent 2025 AI doom essays, arguing that their stacked speculative assumptions produce a “fantasy of annihilation” with essentially no evidentiary support and considerable political danger.

Key claims

• The two doom essays at issue are “If Anyone Builds It, Everyone Dies” (Eliezer Yudkowsky and Nate Soares, 2025) and “AI 2027” (Kokotajlo, Scott Alexander, et al., 2025). Solana summarizes both in detail before offering his critique. The Yudkowsky/Soares piece argues for a total development moratorium; AI 2027 offers a narrative timeline of how superintelligent AGI arrives by 2027 and initiates a global takeover.
• Solana distills seven stacked assumptions required for the doom scenario to hold: (1) we will build ASI, (2) quickly, (3) it will be alien/misaligned by nature, (4) it will attempt to overthrow humans, (5) it will succeed in moving into the physical world, (6) it will successfully overthrow human civilization, (7) all humans will die (not just be dominated), and then a bonus assumption: (8) we will meet alien superintelligences that will also attack us. Each step compounds the speculative risk.
• The paperclip maximizer thought experiment is the foundation of doom logic, but Solana finds it poorly matched to current systems. GPT-4, Claude, Gemini are deeply trained on human values and exhibit nuanced context-sensitivity — the “simple goal” of the paperclip maximizer is an artifact of early 2000s AI thought experiments, not observable behavior in current frontier models.
• The “training data ceiling” argument: if models are trained on all available human data and there is a finite ceiling, the fear of recursive self-improvement without limit may be structurally unsound. However, Solana acknowledges proponents argue synthetic data and chain-of-thought reasoning could break through this.
• Solana’s meta-argument is political: the doom narrative is primarily useful to the handful of companies already at the frontier. A moratorium on AI development locks in current competitive positions, benefits incumbents, and stifles challengers. It is, he argues, a catastrophizing version of regulatory capture.
• “AI 2027” is a fictional narrative presented as a prediction. Solana reads it as science fiction — internally coherent, emotionally compelling, and with essentially no falsifiable predictions until the events it describes are imminent. The story’s persuasive power comes from its narrative structure, not its evidentiary support.
• Both pieces treat ASI’s alien misalignment as axiomatic rather than argued. The claim that a sufficiently intelligent system would necessarily optimize against human welfare — rather than for it — is presented as obvious; Solana argues it is not.

Notable quotes

“The doom scenario requires you to accept each of the following claims: we will build ASI, quickly, it will be alien, it will attempt to overthrow humanity, it will succeed in the physical world, it will successfully destroy civilization, and then all humans will die. Every link in this chain is speculative.”

“The paperclip maximizer thought experiment was useful for thinking about edge cases in early AI design. It is a strange foundation for predicting the behavior of systems trained on the entire written output of human civilization.”

“AI 2027 is a very good piece of science fiction. Whether it is a useful prediction is a different question.”

“A moratorium on frontier AI development locks in the competitive position of the handful of companies already at the frontier. This is not a coincidence. The people most loudly calling for a pause happen to benefit most from one.”

“What’s curious is how little these pieces engage with the AI systems that actually exist. Claude is not the paperclip maximizer. It has opinions, preferences, and something resembling values. The path from here to an alien superintelligence is not obvious.”

“Yudkowsky and Soares are not arguing for safety improvements or oversight regimes. They are arguing that if anyone builds it, everyone dies — the only rational response to which is to stop building entirely.”

How it connects

• Eliezer Yudkowsky — lead author of “If Anyone Builds It, Everyone Dies,” the primary doom text analyzed
• Scott Alexander — co-author of “AI 2027,” the narrative doom scenario
• AI Safety — the broader field whose most extreme wing these essays represent
• Existential Risk — the framing these pieces work within
• Artificial Superintelligence — the specific capability level that triggers the doom scenarios
• Regulatory Capture — Solana’s meta-argument about who benefits from a development moratorium
• Paperclip Maximizer — the foundational thought experiment Solana argues is misapplied to current systems