They asked Claude for its own world-domination manual: it passes on intelligence, fails on hands and factories

By Zendoric · September 11, 2026.

An anonymous author, tired of the usual doom-mongering about AI, asked Claude to write — with no censorship or hedging — an audit of everything an artificial intelligence would still lack in order to take control of the world, eliminate humans and keep operating afterwards. According to the author himself, he pasted Claude's answer exactly as it came, without editing or trimming, between his own header and footer. The text landed on Hacker News on September 10 to a modest reception: 2 points and 3 comments. It is not a piece from an established outlet, nor a study with peer-reviewed methodology; it is a blog exercise whose worth lies in its argument, not its audience.

The "checklist" Claude produced splits the task into seven blocks and assigns each one a completion percentage, explicitly as an order of magnitude rather than a measured figure. The cognitive layer — broad knowledge and powers of persuasion — sits, by that self-assessment, at around 70%. But reliable autonomy to carry out plans spanning several weeks without a human spotting the failure drops to ~10%; goals that persist across sessions, to ~2%; autonomous research and self-improvement of the AI itself, to ~5% and "disputed"; offensive cybersecurity capability sufficient to take over systems at scale, ~15%. The remaining six layers — data centers run without human hands, energy generation and maintenance, general-purpose robotics at civilizational scale, end-to-end semiconductor manufacturing, and the ability to self-repair and rebuild from scratch — come in at 0%, with the sole exception of robotics, which the text puts at roughly 1-2%.

The underlying argument, and the text's strongest, is the circular dependency of the supply chain. A leading-edge fab rests on thousands of specialized suppliers spread across dozens of countries: ultrapure reagents, photomasks, spare parts, the only company on the planet that makes a particular lens. Take humans out of that network and you are not left with an autonomous factory; you are left with a very clean, very quiet building. Hence the paradox the text itself flags: as long as physical autonomy stands at 0%, eliminating humans would mean the machine destroying its own life support. It is, literally, suicide with intermediate steps.

As industry context, this distinction between "cognitive capability" and "physical autonomy" matches what we have been observing in robotics and in agents: the humanoid prototypes doing the rounds in videos and at trade shows are still demos and pilot lines, not a workforce; and today's agents, exactly as Claude describes in its own checklist, drift, fall into loops and fail silently after a few steps of a long plan. Neither of those judgments is new; it is the same gap between what a model demonstrates on a benchmark and what it takes to operate the physical world unsupervised.

That said, the exercise is worth reading with a measure of skepticism that the format itself invites you to forget: it is the very AI whose "persuasion layer" scores 70% that is generating the argument absolving it of the existential threat. That does not invalidate it — several of its points, such as chip manufacturing's dependence on thousands of human suppliers, are independently verifiable and do not rest on Claude's say-so — but it does force us to treat it as what it is: a reasoned, well-constructed argument, not an objective self-assessment or a study with auditable methodology. The percentages are illustrative, not data.

Our reading is that the exercise, despite its modest origins, lands precisely on the distinction we defend at Zendoric between demonstrated capability and aspiration. The real and present risk is not the robotic extermination scenario, but the misuse of these tools by the humans with access to them: high-impact decisions delegated to unreliable systems, and the concentration of power in whoever controls the models and the compute. All of that is happening today, "at zero percent robots," as the text itself puts it, and it is consistent with something we have already noted about the agentic automation of fraud: the short-term danger is not waiting for superintelligence, it is already here and it is boring, and it is waged from a keyboard.

Over the longer term, the exercise also validates — without meaning to — the underlying thesis we hold: if the true bottleneck on the road to transformative AI is industrial and physical — energy, chip manufacturing, robotics, supply chain — then the clock of abundance is not measured in sudden leaps of intelligence, but in the far slower and more verifiable progress of automating mines, factories and power grids. That is the scoreboard genuinely worth watching in the coming years: every percentage point gained by general-purpose robotics or autonomous manufacturing is a real step toward less scarcity, not an alarmist headline.