A Hacker News thread titled "Why I'm still bearish on LLMs after Navier-Stokes" reached 166 points and 174 comments. The post argues that large language models cannot reliably solve or verify solutions to the Navier-Stokes equations, a core test of physical accuracy.
What the Post Claims
The author tested LLMs on fluid dynamics problems governed by Navier-Stokes. Models produced plausible-sounding derivations yet failed to maintain conservation laws or boundary conditions across steps. Errors appeared consistently when equations required iterative numerical stability checks rather than pattern matching.
The post frames this as evidence that scaling alone will not close the gap. LLMs optimize for token prediction, not for enforcing differential equation constraints.
Community Reaction on Hacker News
Commenters with physics and numerical methods backgrounds largely agreed. Multiple users reported similar failures when prompting models for finite element analysis or turbulence modeling. Several noted that models hallucinate stable solutions that diverge under actual simulation.
A minority argued that hybrid systems—LLMs paired with symbolic solvers—could mitigate the issue. Others countered that verification still requires human oversight or dedicated solvers.
Bottom line: The thread shows broad practitioner skepticism that pure LLMs will replace numerical methods for physics-grade accuracy.
Technical Limitations Highlighted
Navier-Stokes solutions demand strict adherence to continuity and momentum conservation. LLMs lack built-in mechanisms to enforce these invariants during generation. Token-by-token prediction can violate partial differential equation residuals without detection.
Existing benchmarks such as MATH and GSM8K measure algebraic manipulation, not continuous physical systems. No current public LLM benchmark tests long-horizon consistency on fluid equations.
Practical Alternatives
Teams needing reliable fluid simulations continue to use established tools:
- OpenFOAM for open-source CFD
- COMSOL Multiphysics for commercial finite element work
- JAX-based differentiable solvers for research code
These packages deliver deterministic results with explicit error bounds. LLMs remain useful for initial code scaffolding or documentation lookup but not for final numerical output.
| Approach | Accuracy on Navier-Stokes | Verification | Typical Hardware |
|---|---|---|---|
| Pure LLM | Low | Manual | GPU inference |
| OpenFOAM | High | Built-in residuals | CPU/GPU cluster |
| LLM + solver | Medium-High | Solver residuals | GPU + CPU |
Who Should Pay Attention
Researchers building autonomous scientific agents should treat the thread as a caution. Any pipeline claiming to generate publishable physics results without external verification carries high risk. Practitioners focused on code assistance or natural language interfaces can continue using LLMs without conflict.
Developers targeting high-stakes simulation domains should prioritize hybrid architectures that keep numerical solvers in the loop.
Verdict
The discussion reinforces that current LLMs remain unreliable for problems requiring strict physical invariants. Until architectures incorporate explicit constraint enforcement, numerical methods retain the advantage for Navier-Stokes and similar domains.
Top comments (0)