A Vietnamese teenager born 2010–2012 making a high-school subject focus, university major, and early-career project portfolio decision in 2026 is implicitly buying a 4–7 year forward position on a labour market whose task structure is being repriced by generative-AI tools at a pace that the canonical Frey-Osborne 2013/2017 occupation-level model did not anticipate and substantially overstated. This paper synthesises the empirical labour-economics and AI-disruption literature 2018–2026 to produce an operationally usable durability taxonomy for that decision. The synthesis covers the Acemoglu-Restrepo task-displacement-and-reinstatement framework, the Brynjolfsson-Mitchell-Rock 2018/2023 Suitability for Machine Learning construct, the Eloundou-Manning-MishkinRock 2023 GPTs are GPTs exposure measure, the Felten-RajSeamans 2018/2023 occupational AI exposure index, the Korinek 2024 economics-of-transformative-AI synthesis, the posthoc walk-back of Frey-Osborne 2013/2017 by Arntz-GregorZierahn 2016 and the OECD 2018/2024 reanalyses, and the 2024–2026 empirical wave (McKinsey Generative AI and the Future of Work, OECD 2024, Goldman Sachs 2023, World Economic Forum 2024 Future of Jobs, IMF 2024 Cazzaniga GenAI: Artificial Intelligence and the Future of Work). Vietnamese context is grounded in ILO Vietnam, MOLISA-GSO-ADB skill forecasting, and the VietnamWorks/JobStreet 2024–2025 salaryand-vacancy series. The paper's contribution is a three-tier taxonomy — durable, amplification-leveraged, automatable — instantiated across eight capability axes (judgment, taste, verification, orchestration, deep-domain expertise, embodied/physical, relational, ethical), with explicit separation of (a) measured 2023–2026 effects, (b) plausible 2026–2030 trajectories, and (c) speculative 2030+ scenarios that the paper marks as such. Operational guidance is provided at three decision points: highschool subject focus, university major selection input, and earlycareer project portfolio for 18–22 year-olds. Two failure-mode skill investments common in 2024–2026 Vietnamese discourse — generic prompt-engineering courses and surface-level fullstack development — are analysed and not recommended as primary investments. The Frey-Osborne walk-back is handled directly: their 47% of US occupations at high automation risk is now treated as a methodological artefact of occupation-level rather than task-level decomposition; subsequent task-level work (Arntz-Gregor-Zierahn, Felten-Raj-Seamans, Eloundou et al.) places the empirically defensible exposure share substantially lower and concentrates it in specific task bundles rather than whole occupations. The thesis the paper defends is that durability flows from capabilities that are co-located with judgment under uncertainty, with embodied or relational presence, or with deep verification of outputs that AI systems generate cheaply but
That Le (Tue,) studied this question.