The post is Maciej Ceglowski’s 2016 talk “Superintelligence: The Idea That Eats Smart People,” which argues that Nick Bostrom-style AI doom thinking borrows too much from religion, folklore, and elite techno-grandiosity. The talk’s core claim is that smart people get captured by a story about all-powerful machine minds, when the real harms from machine learning are much more ordinary and human: exploitation, bad incentives, surveillance, and institutions using automation to do ugly things at scale.
That basic framing still landed for a lot of readers, especially the part that now feels prophetic about tech elites using AI rhetoric to justify influence and infrastructure buildout. But the thread did not buy the essay as a technical takedown. The main verdict was that it is better at puncturing the culture around superintelligence than at disproving the possibility. Several of the essay’s examples, like cats in carriers, emus, or brilliant people failing at practical tasks, were seen as category errors. They mostly show that intelligence is constrained by embodiment, coordination, and incentives. They do not show that more capable systems cannot become dangerous.
The more grounded consensus was that the 2026 version of AI risk looks neither like the essay’s target nor like the clean sci-fi “
paperclip maximizer” caricature. Current models are not superintelligent. They are uneven tools with glaring weaknesses like hallucination, sycophancy, and reward hacking. Yet they are already economically and politically powerful because institutions can deploy them at scale. That shifted the center of gravity toward near-term harms such as propaganda, surveillance, labor control, biosecurity misuse, and capital-intensive data center buildout. Readers kept coming back to the same point: the dangerous actor is often not an autonomous god-machine but a company, state, or bureaucracy with access to lots of compute, data, and leverage.
From there, the argument split in a more useful way than the essay does. One camp said this makes superintelligence talk actively distracting. “Hard takeoff” and recursive self-improvement still look hand-wavy, bounded by data, hardware, physical limits, and messy interaction with the real world. Another camp said that none of those limits are comforting. Human intelligence is not obviously a ceiling. Narrow systems already outperform humans in some domains, self-play and automated research are real mechanisms for capability gains, and competition among labs makes slowing down individually hard. Even many skeptics of Bostrom-style doom still conceded that alignment remains unsolved at much smaller scales than “superintelligence,” which is a bad sign.
A smaller but strong thread widened the lens beyond technical capability altogether. It argued that “alignment” is too often framed as a machine problem when the real issue is political economy. The systems being built reflect the incentives of founders, investors, states, and managers. That is why readers talked more about labor organizing, local opposition to data centers, owning your own data, and community-controlled technology than about abstract utility functions. The most durable takeaway was blunt: the essay was right to attack AI as ideology and status theater, but wrong to think that puncturing the mythology settles the risk question. We are already living inside the more realistic version of the problem.