What is the difference between narrow AI and general AI?

Narrow AI does specific tasks it was built for; general AI (AGI) would handle any intellectual task a human can, transferring understanding across domains without being retrained. Everything deployed today is narrow — though modern language models have blurred the boundary enough that the definition itself is now contested. **Narrow AI** covers essentially all of it: a chess engine that plays chess superbly and cannot do anything else; a spam filter; a medical imaging model; a recommendation system. Superhuman within its domain, incapable outside it, with no understanding it can carry elsewhere. **AGI**, as usually defined, would learn any new task about as readily as a human, transfer knowledge between unrelated domains, set its own goals and subgoals, and operate in situations it wasn't designed for. Nothing that exists does this. Why the boundary got messy: large language models are technically narrow — they're trained on next-token prediction — but they're general across an enormous range of *tasks*. One model writes code, analyses arguments, translates, and explains physics without being separately trained on each. That's not what 'narrow' used to mean, and it's why serious people disagree about how close we are. What's still clearly missing, and this is where the honest assessment sits: - **Reliable reasoning over long chains.** Errors compound; performance degrades in ways human reasoning generally doesn't. - **Continuous learning.** They don't update from experience during use. Every conversation starts from the same fixed weights. - **Genuine grounding.** Learning about the world from text about the world differs from interacting with it. - **Robust generalisation** to genuinely novel situations, as opposed to novel combinations of familiar ones. - **Self-directed goals** and knowing what's worth doing. On timelines: predictions from credible researchers span from a few years to many decades to never, which should tell you how little anyone knows. Historically, AI timeline predictions have been wrong in both directions — some things arrived far sooner than expected (language fluency), others remain stubbornly hard (reliable common-sense reasoning, dexterous manipulation). A more useful frame than the AGI binary: capability is arriving unevenly. Some human-level and superhuman abilities already exist alongside failures that a child wouldn't make. That jagged profile is likely to persist, which means 'has AGI arrived' will probably be argued about rather than observed as an event — and the practical questions (what can it do reliably, who's accountable when it's wrong) matter more either way.