12 hours ago
- The author challenged LLMs to predict the temperature of cooling coffee over time, accounting for many complex physical phenomena.
- LLMs produced equations with one or two exponential decay terms, e.g., T(t) = 20 + A exp(-t/b) + C exp(-t/d).
- The author conducted a real experiment, measuring water temperature in a mug over 5 minutes.
- Experimental results showed faster initial cooling and slower later cooling than any LLM predicted.
- None of the LLM predictions were highly accurate; the best (Claude 4.6 Opus) cost $0.61 in tokens.
- Some models skipped the fast cooling rate term, using only a single exponential decay.