On Mon, 13 Jan 2025 09:43:28 -0000 (UTC), Jethro_uk <
[email protected]> wrote:
I may start posting daily "AI" fails. It really is still fuck all use to
man nor beast.
I asked Gemini how many degrees an electric 6.5 kW heater would heat a flow of 3.7 liters per minute.
Answer: 248 °K.
Bunch of waffle about unrealistic because losses would make that lower, and that
water would boil before that.
Drilled down to find the mistake:
Gemini calculated 6500 J/s / (0,062 kg/s * 4186 J/kg*K) as 248 K.
Gemini calculated 6500/(0,062*4186) as ~25.
I asked about the discrepancy, the answer:
"I probably made a calculation error or wrote down a number incorrectly the first time. It is important to be careful when doing such calculations and to check the results."
"Wrote down the number incorrectly" -- a crap excuse for a program given a number in writing as input, sounds like a lame excuse form a schoolchild.
Gemini finally got it right, and prefaced the correct answer with using the phrase "Calculation using a calculator"... 24,8 K. (My calculator arrives at 25,04 by the way).
The place where language model meets calculation is visible. Gemini parsed the question correctly. It found the formula, formulated grammatically correct natural-language and correct explanations of why the answer would be different from a measurement (heat losses, boiling, overheat stat). It formulated reasonable-sounding excuses and explanations for its errors. It correctly cancelled the units in the formula, and completely fucked up simple arithmetic.
Very odd, that's one thing I would have expected to Just Work.
Thomas Prufer
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)