Feed LLMs a table of pure noise and they'll confidently invent 'sensor data'
No-Plant-5234 · reddit · 2026-09-02
A Reddit user built a table of pure noise—random floats, columns named c1–c20, no real structure—and asked several LLMs to describe the dataset. Every model obliged; one claimed it looked like sensor readings from manufacturing equipment and speculated on column correlations. No errors, no hedging, same confident tone.
The author notes this is inherent completion behavior, but a real hazard for agent pipelines: if an agent's first move on an unfamiliar table is to ask an LLM, the answer will always sound confident with no signal distinguishing genuine understanding from fabrication. Asking for a confidence number yields a made-up number; a second model checking the first just agrees with it. The author works at SchemaLabs, whose table-trained model returns "no domain identified" on the noise file, but admits the test design is crude and asks the community for rigorous methods to test whether a model actually recognizes a table versus pattern-matches a plausible description, and whether anyone's agent has a path to stop and declare a source unreadable.
Related event: LLMs Confidently Fabricate Meaning From Pure Noise Data(2 posts)→
More from coding & agent
- Claude-written test harness let 31KB fake archives land on a real backup server — Grimmoner · 2026-10-18
- Open-source Tyger-S2S voice agent runs natural, interruptible conversations on a single GPU — solyarisoftware · 2026-10-18
- Matt Pocock's pre-commit hook caps agent-written files at 5,000 tokens — mattpocockuk · 2026-10-18
- Backups silently failed for 10 days: three monitoring blind spots lined up — Grimmoner · 2026-10-18
- Author Laurence Moroney: write your ugly first draft yourself, then let an LLM edit it — lmoroney · 2026-10-18
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18