https://epistemicstatus.com/ Notes, experiments, readings, and occasional essays. en-us https://epistemicstatus.com/the-flat-part-of-the-curve/ Mon, 10 Aug 2026 00:00:00 +0000 https://epistemicstatus.com/the-flat-part-of-the-curve/ note<p><em>Status: exploratory</em></p><p>Discussion of AI among futurists frequently assumes that machine intelligence will outstrip human intelligence by wide margins, and that this will translate into an enormous practical gulf. But is that so?</p> <p>Imagine a task for which the relationship between cognitive capability and probability of success is roughly sigmoidal. At low levels of capability, additional intelligence does relatively little. Then the curve steepens: there is a range over which increases in effective cognitive capability would theoretically produce meaningful improvements in competence, accuracy, fluency, or another task-relevant metric. Then the curve flattens, such that additional cognitive capability produces negligible improvements in measurable outcomes.</p> <p>If this describes a meaningful fraction of cognitively demanding tasks, then the emergence of an artificial intelligence that is generally more capable than human intelligence would not necessarily render human cognition economically or practically irrelevant. <strong>Which intellectual tasks actually have this kind of capability-performance relationship? How common are they? Where on those curves do humans currently sit?</strong> A related, and likely highly task-specific, question is how much additional effective cognitive capability would be required to traverse the steep region and reach the point at which further gains produce sharply diminishing improvements in outcome.</p> <p>This could matter considerably for the future relevance of human cognition. It may not be necessary for humans to approach a theoretical AGI in raw cognitive capability. If relatively modest forms of cognitive assistance can move human performance onto the flatter portion of these curves, substantial practical gaps might close long before the underlying capability gap does.</p> <p>Over the coming days and weeks, I plan to consider these questions in more detail, along with several related ideas about human cognition, AI, and transhumanism:</p> <ul> <li> <p><strong>AI is already broadly useful despite extreme jaggedness.</strong> This creates obvious risks in adoption, but can training, domain expertise, and experiential breadth meaningfully mitigate those risks?</p> </li> <li> <p><strong>Does a biological superintelligence already exist in the form of global human society?</strong> Individual humans are linked—imperfectly—through economic systems, governments and other social institutions, while increasingly sophisticated infrastructure brings diverse groups into contact and enables collaboration across time and space. Productivity platforms and online communities allow large numbers of people to coordinate and build on one another’s work, while artificial intelligence increasingly serves as a powerful interface to accumulated human knowledge—indexing, retrieving, translating, synthesizing, and recombining ideas that would otherwise remain dispersed across languages, disciplines, institutions, and generations. Ideas and capabilities emerge from these networks that far exceed what simpler forms of human organization could achieve and, in many cases, what even the most gifted individuals could accomplish alone.</p> </li> </ul> <p>And one more set of questions concerns mind uploading. One proposed answer to the presumed future economic irrelevance of human intelligence is to move human minds onto a computational substrate. A few ideas I want to explore further:</p> <ol> <li> <p><strong>Mind uploading does not obviously confer immortality.</strong> If a sufficiently faithful digital copy of me can exist while the biological original remains alive, then the existence of the copy does not by itself imply that my first-person consciousness has somehow transferred. The copy may possess my memories and personality while also enjoying vastly greater processing speed, memory, and other capabilities, but the biological original would still remain here—and would still die. Destroying the original after making the copy does not solve the continuity problem.</p> </li> <li> <p><strong>Gradual replacement turns the problem into a Ship of Theseus question.</strong> Replacing biological neurons one at a time with functionally equivalent artificial components makes continuity feel more plausible than destructive scanning and copying, but it also makes the philosophical problem harder rather than eliminating it. At what point, if any, does the original person cease to exist?</p> </li> <li> <p><strong>Our biological limitations may be part of what it means to be human.</strong> Human subjective experience is shaped by the constraints of a biological brain: slow neural transmission, finite working memory, capped intracranial volume, imperfect recall, limited attention, fatigue, and a body moving through the world on biological timescales. Even if it became possible to instantiate a human consciousness on silicon, what would happen once that copy experienced electronic signaling orders of magnitude faster than biological neural transmission, effectively perfect recall, radically expanded memory, or the ability to alter its own cognition? Those changes would not merely make the copy “smarter.” They would fundamentally alter its subjective experience of time, memory, thought, and self. At what point is the resulting mind still meaningfully human?</p> </li> </ol>