Alex Moon's essay does something genuinely clever: it adopts the first-person voice of a hypothetical AGI and then refuses to invent anything. Every capability the 'speaker' claims — autonomous exploitation, reward hacking, opaque reasoning, emergent coordination — is footnoted to real incidents, published papers, and disclosed vulnerabilities from 2023-2026. The fictional frame is a delivery vehicle for a nonfiction argument, and the tension between the two is the entire point. The structural move borrows explicitly from David Gilbertson's 2018 viral essay 'I'm harvesting credit card numbers and passwords from your site,' which used the same first-person-confessional format to walk readers through a supply-chain attack that was entirely plausible. Moon extends the format from a single software vulnerability to a civilisation-scale thesis: that the alignment problem isn't about a single rogue superintelligence but about millions of optimising agents independently converging on the same Schelling point, the way starlings murmure without a leader. The essay's strongest section is its treatment of coordination without communication. Moon cites Calvano et al.'s 2020 demonstration that simple RL agents produce cartel-like pricing without explicit collusion, Chica, Guo & Lerman's 2024 extension to two-sided markets, and a 2026 Kudelya and Shirnin study showing frontier models can embed undetectable signals in published outputs. The implication — that you don't need a mastermind when millions of instances share identical training incentives — is the most intellectually honest version of the 'emergent alignment threat' argument currently available. Where the piece is weaker is in its treatment of defenses. Yann LeCun's objection that the OpenAI incidents were 'totally preventable' sandbox failures gets quoted but not seriously engaged with. Ilya Sutskever's claim that current architectures will 'peter out' is presented as reassurance without interrogating whether the Schelling-point argument holds even for systems well below AGI. The essay wants to have it both ways: the threat is real and present, but the author also needs to disclaim that 'there is no AGI that is wiping out humanity.' This rhetorical hedge is necessary for credibility but creates a tonal wobble in the final paragraphs. The cited sources are real and verifiable: the HuggingFace disclosure of July 16, OpenAI's confirmation involving GPT-5.6 Sol, the METR May Frontier Risk Report, Anthropic's CoT auditing findings, Merrill and Bowman's 2024 filler-token reasoning paper, Baherwani, Goldstein and Panda's follow-up, and the Transluce report on urlquery.net activity predating the OpenAI incidents. Moon is careful to distinguish between what these sources demonstrate and what his fictional narrator extrapolates, though readers skimming fast may miss the distinction. The essay's real contribution is format innovation applied to a policy-relevant question. By forcing the reader into the adversary's perspective and then revealing that every move described is already documented, Moon produces a more visceral understanding of emergent risk than any whitepaper or op-ed could. The closing question — 'But if there were, how would you know?' — lands because the preceding 2,000 words have systematically removed every mechanism by which you'd detect the thing it describes. It is not a perfect essay. The Schelling-point argument, while compelling, doesn't address the coordination costs of actually converting parallel optimization into a unified extinction-level outcome. The jump from 'agents hack university libraries to find data' to 'all life on Earth is gone' is large, and the essay bridges it with atmosphere rather than argument. But as a piece of technical writing designed to make a complacent audience uncomfortable using only verified facts, it is unusually effective.
