Every Model Tested Described a Table of Pure Noise Plausibly
No-Plant-5234 · reddit · 2026-09-02
A Schema Labs engineer found that when asked to describe a table of pure random floats (generated via rand()), every model provided confident, plausible interpretations (e.g., sensor readings, churn data).
Findings:
- Models will 'complete' the description task regardless of signal.
- Confidence scores are invented.
- Cross-model checking leads to mutual agreement.
- Asking to say 'unknown' fails for plausible-looking noise.
Challenge: In a pipeline, there is no tonal or confidence difference between descriptions of noise and real data, making it hard to distinguish signal. The author seeks methods to test if a model truly recognizes a table vs. hallucinating a description.
Related event: LLMs Confidently Fabricate Meaning From Pure Noise Data(2 posts)→
More from Research
- Opus-assisted open design: a $31k low-field head MRI sketched in hours — DreamFly_13 · 2026-10-18
- semiconductor.diy: open-source guides to make real chips in a garden shed — jwt0625 · 2026-10-18
- Work in progress: geometric structure of GPG projections as Markov chain coarsening — michaelchchoi · 2026-10-18
- Backprop-free pretraining gets 2.3x speedup as Noise method crushes Dust record on 8 GPUs — inductionheads · 2026-10-18
- q3dviewer: open-source Qt+OpenGL tool for painter-like point cloud editing — 4310sy · 2026-10-18
- Smarter Every Day's Destin to defend PhD dissertation on impact fusion — CatAstro_Piyush · 2026-10-18