The 100k whys of AI

People largely bought that framing. The strongest recurring point was that LLMs and image generators are very good at polished average. A single output can feel competent or even clever, but a batch of 50 or 500 exposes the template underneath. Several people said this matches direct experience with blog posts, podcast summaries, AI music, and junk YouTube genres. Once you consume enough of it, you can predict the cadence, the fake tension, the resolution, and even the character names. That also explains why many people can now spot AI writing by smell rather than by any one tell. The giveaway is not em dashes or stock phrases by themselves. It is the total rhetorical shape. The most useful refinement was that this sameness is partly by design. These models are trained to be helpful, safe, and on-distribution. For code and other utilitarian tasks, that bias toward conventional answers is often desirable. For books, art, and essays, it produces bland convergence. A few commenters pushed back that the article proves less than it claims because prompt quality, system prompts, and workflow matter a lot. With more structured steering, randomization, and iterative outlining, you can get more varied results than a pile of naive one-shot prompts would suggest. Even those defenses rarely claimed you get great art out of the process. They mostly argued you can widen the range of mediocre outputs. A separate thread broadened the issue from originality to market effects. People pointed to AI books and images already showing up in Amazon rankings and physical retail, often with obvious errors inside. That turned the post from an aesthetic complaint into a distribution story. Cheap, coherent-enough content can flood channels long before models become truly inventive, and many buyers either will not notice or will not care enough to stop it.