A question circulating now: which of the hard problems people flagged for AI have…

For decades, technologists and researchers have articulated specific capabilities they believed would require human level intelligence. These were less predictions and more hard constraints, the kinds of problems where smart people said "we don't know how to do this yet." Hacker News, as a community of engineers and founders, became one place where these constraints got regularly discussed and debated. The ongoing conversation centers on a fairly simple question: as AI capabilities actually materialize, which of those "hard problems" turn out to be tractable after all, and which turn out to be misidentified constraints versus genuine limitations. The history of AI problem classification goes back to early expert systems and symbolic AI research. Researchers would identify tasks that seemed to require understanding, reasoning, or creativity: things like natural language translation, image recognition, game playing against humans, and generating coherent text. These became benchmarks and targets because they felt uniquely difficult. The implicit theory was often that solving these would demonstrate genuine understanding or intelligence. As computing power, data availability, and algorithm design improved, some of these targets moved from impossible to difficult to routine. Other problems proved easier to solve in narrow contexts than anyone expected, while the broader version of the problem remained unaddressed. The challenge with talking about "solved" problems is that resolution often feels partial or sideways. Image recognition went from mostly unusable in the 1990s to highly accurate in many commercial contexts by the 2010s. But "solved image recognition" doesn't mean computers understand images the way humans do; it means they can classify them with human level accuracy in training conditions. Similarly, systems can generate fluent sentences or code without necessarily demonstrating understanding in the philosophical sense. The framing of a problem, the metrics used to measure it, and the acceptable accuracy threshold all matter hugely to whether something counts as solved. Hacker News discussions about this tend to divide into camps. Some people focus on narrow capability gains: a system can beat humans at a game, or transcribe speech, or write syntactically correct programs. Others ask whether the underlying approach actually solved the conceptual challenge or just sidestepped it through scale or pattern matching. Still others point out that some originally flagged hard problems weren't as coherent as they seemed. What seemed like one unified problem, language understanding for instance, breaks down into dozens of different subproblems with different difficulty curves. The practical significance of this conversation is substantial. If you're deciding whether to build a product, invest in research, or allocate engineering resources, knowing which hard problems are actually solvable matters. It shape