Where does next-token prediction leave us?

Original Article ↗ Hacker News Discussion ↗

Concerns over large language models and “next-token prediction” center less on the math and more on power, class, and the future of work. Commenters argue over whether AI’s productivity gains will be broadly shared or primarily enrich existing elites, with many fearing mass displacement of white‑collar jobs, erosion of individual bargaining power, and the creation of a “permanent underclass.” Others see LLMs as another labor‑saving tool that, like electricity or the power loom, will eventually boost living standards, though even optimists acknowledge deep uncertainty about the transition period, regulation, and how societies will adapt psychologically and politically.

Global Attitudes, Class, and Social Contracts

  • Several comments question the idea that only the economically secure support AI, citing surveys showing more negative views in the US/EU than in China/developing countries.
  • Explanations offered:
    • China is perceived as still on an economic “winning streak” with a more collectivist social contract, so AI is framed as national progress with state-backed retraining and infrastructure.
    • In the US/EU, AI is more often seen through the lens of oligarchy, deindustrialization, weak safety nets, and fear of a “permanent underclass.”
    • Some link AI enthusiasm to cultures focused on quick gains, weaker traditions of critical thinking, and higher corruption.

Democratization, Skills, and Fungibility

  • Strong disagreement on whether AI “democratizes” creation:
    • Pro: It lowers barriers to making software, games, tools—like cheap Ferraris broadening access.
    • Contra: Skills were already accessible via free learning; AI instead devalues hard-won expertise and makes workers more fungible, undermining bargaining power.
  • Analogies to power looms and earlier automation recur, with some arguing “this time is different” because AI targets most knowledge work.

Jobs, Economy, and Redistribution

  • Deep concern about mass displacement, second-order effects (e.g., customers losing income, deflation, oversupply of remaining jobs like trades), and lack of clear political response (UBI, retraining, etc.).
  • Some argue historical pattern: productivity gains are ultimately redistributed and create new roles; others counter that redistribution is not automatic and short- to medium-term harm will be severe.
  • Debate over who is really at risk: junior devs vs. middle managers vs. whole professions.

Ethics, Responsibility, and Class War Framing

  • Moral unease about working for AI labs compared to making guns or doing neutral research; questions of complicity when leadership openly talks about replacing labor.
  • References to “class war,” rent-seeking on humanity’s collective output, KYC as an information-control mechanism, and propaganda that redirects anger away from elites.

Nature and Limits of LLMs

  • Dispute over “next-token prediction” as a dismissive framing:
    • Some say we’re beyond simple next-token models (RL, reasoning, diffusion).
    • Others maintain that, even with reasoning scaffolds, LLMs remain next-token predictors without true innovation or agency.
  • A minority expects LLM progress to plateau; others foresee profound, possibly uncontrollable transformation.

Psychological and Cultural Impact

  • Several express apathy and loss of meaning as AI encroaches on craft and learning.
  • Others see AI as another powerful tool like Google or Wikipedia, enabling curiosity rather than replacing it.