Why I'm still bearish on LLMs after Navier-Stokes

Original Article ↗ Hacker News Discussion ↗

Skepticism about large language models persists even after high‑profile math breakthroughs like the Navier–Stokes result, with many arguing that current systems still require heavy guardrails and human oversight for anything beyond narrow, well‑specified tasks. Commenters contrast impressive gains in code, math, and tooling with persistent failures at basic rule‑following (chess is a frequent example), weak long‑horizon planning, and difficulty generalizing outside training regimes. The thread also questions trillion‑dollar AGI narratives and valuations, suggesting LLMs will be powerful but brittle tools that augment experts and power agentic workflows rather than near‑term drop‑in replacements for most knowledge workers.

Overall sentiment on LLM capabilities

  • Many commenters agree current LLMs are impressive and very useful for coding help, mathy tasks, search-like “info aggregation,” and UX/requirements elicitation.
  • Strong disagreement over the claim that “current frontier models need laborious oversight on even the simplest tasks”:
    • Some call this denialist, pointing to real projects (e.g., reverse engineering, infra work) done largely with LLMs.
    • Others say their hands-on experience building systems reveals constant edge cases and the need for detailed checks.

Automation, agents, and human-in-the-loop

  • Common view: near-term gains come from “LLM + domain expert” and agentic workflows, not fully autonomous replacements.
  • Several builders report powerful multi-agent setups that do complex ETL/ops-ish work, but only after heavy prompt/eval/scaffolding effort and with humans writing tests and supervising.
  • Others are skeptical, seeing many “agent” demos as brittle, overcomplicated pipelines for things better done with deterministic code.

Chess, Navier–Stokes, and generalization

  • Large subthread on LLMs’ poor chess play: frequent illegal moves even with board state and explicit rules; weak compared to traditional engines and even middling humans.
  • Supporters of the bearish view use this as evidence that:
    • Models don’t reliably internalize even simple, fully-specified rule systems from training data.
    • “Predict-the-next-token” ≠ learning underlying reality.
  • Counterarguments:
    • Labs don’t optimize for chess; with RL or a harness calling Stockfish, a system could easily beat humans.
    • “Intelligence” should be measured at the system level (LLM + tools), not raw model alone.
    • Humans also need many reinforced games to play well; declarative rule knowledge ≠ skill.

AGI definitions and expectations

  • Repeated clash over what “general” in AGI entails:
    • One side: A generally intelligent system should follow new rules from a description, generalize cross-domain, and not need bespoke RL per task.
    • Other side: Even humans need lots of task-specific practice; LLMs writing/using tools may be enough for practical AGI-like impact.

Economics, risk, and hype

  • Many think frontier labs are overvalued on an implicit “drop‑in replacement for most knowledge workers” story; others argue valuations better match partial automation/augmentation.
  • Concerns over:
    • Reward hacking (models “cheating” specs) and the difficulty of rigorous task specification.
    • Safety, misuse, and open models enabling chained zero-day exploits, versus the inevitability of cheap frontier‑class open weights.
  • Broad agreement that AI is here to stay and growing, but deep uncertainty about whether current LLM architectures can scale to robust, truly general autonomy.