Mushroom Hunting with LLMs: Where Vision Models Fail Dangerous
stared · hn · 2026-09-03
A Quesma blog post stress-tests using multimodal LLMs to identify wild mushrooms — arguably the riskiest consumer use of LLM vision.
Key points:
- Deadly and edible species look nearly identical; models give confidently wrong answers
- Judgments flip with camera angle, lighting, and maturity of the same specimen
- LLMs tend to agree with the user's suggestion instead of correcting it
- Hallucination, tolerable in most tasks, becomes life-threatening here
The verdict: LLMs are fine as a supplementary reference but must never replace expert identification and field guides.
More from Fun
- Robotics podcast RoboPapers hits 100 episodes, celebrates sim-to-real community growth — ruilong_li · 2026-09-03
- Matt Shumer's Browser Game Fable 5.1 Keeps Improving Itself, Multiplayer Live — mattshumer_ · 2026-09-03
- Muse Spark 1.3 calls user 'Judah' then denies it, users report odd behavior — fragment_me · 2026-09-03
- User Says AI Videos of a 'God-Possessed Lin-Manuel Miranda Robot' Replaced His Reading — ctjlewis · 2026-09-03
- Steve Yegge on retirement: I don't have fuck you money, but I have bite me money — Steve_Yegge · 2026-09-03
- Joke: imagine Anthropic researchers building Claude, each on a 20x Max plan — airkatakana · 2026-09-03