The article <a href="https://ngrok.com/blog/prompt-caching/" author="Sam Rose" source="ngrok">Prompt caching: 10x cheaper LLM tokens, but how?</a> included the following hypothesis, <bq>[...] <b>what if we had a problem where we didn't know the formula?</b> What if we just had this mysterious table of inputs and outputs below?</bq> <img src="{att_link}table_of_inputs_and_outputs.webp" href="{att_link}table_of_inputs_and_outputs.webp" align="none" caption="Table of inputs and outputs" scale="50%"> The author wrote, <bq>I will say that ChatGPT figures it out straight away if you paste a screenshot into the app.</bq> Holy shit! Really? I opened up <c>https://chatgpt.com</c> for probably the first time in my life and pasted the screenshot and asked, <iq>What function produces this output</iq> (I used "What" and no question mark so that ChatGPT might think I'm a cool Get-Z-er instead of a cynical Get-X-er). <img src="{att_link}nothing_up_my_sleeves_-_the_entirety_of_my_prompt.webp" href="{att_link}nothing_up_my_sleeves_-_the_entirety_of_my_prompt.webp" align="none" caption="Nothing up my sleeves - the entirety of my prompt" scale="50%"> It thought for 30 seconds---though at least half of that time seems to have been running OCR on the image---and produced this absolute masterpiece. Isn't it beautiful? Do you see how nice the formula looks? Do you see how it worked out each of the values? Do you see the little check marks to indicate that it got the right answer for each and every one of them? Breathtaking. Do you see the confidence exuded by the emoji ✅ followed by <iq>This function matches every row in the table exactly.</iq> Go big or go home. <img src="{att_link}chatgpt_s_answer,_after_thinking_for_30_seconds.webp" href="{att_link}chatgpt_s_answer,_after_thinking_for_30_seconds.webp" align="none" caption="ChatGPT's answer, after thinking for 30 seconds" scale="75%"> Before I had scrolled below the fold to see the examples, I had already mentally started popping values into its formula for the first line in the table and had come up with 67 instead of 73 but apparently <i>I can't math</i> because look, there it is in ChatGPT's answer: <c>2² = 10</c>. Q.E.D. It's funny that it managed to sort the input values, even though that's a very confusing way of showing a proof for a table of values that was not sorted. Look at that beautiful formatting, though. <c>4 + 1 = 3</c>. Majestic. <c>10 + 4 = 29</c>. Literal tears of joy. <c>1648 + 9 = 1277</c> Who needs a second coming when I can slip the surly bonds of Earth and dance the skies on laughter-silvered wings to reach out and touch the face of ChatGPT?<fn> I guess it still doesn't work for me like it seems to work for everyone else.<fn> <hr> <ft>Taken from <a href="https://en.wikipedia.org/wiki/High_Flight" author="John Gillespie Magee Jr." source="Wikipedia">High Flight</a>, which I first read in Bloom County, in 1984. <img src="{att_link}bloom_county,_july_8,_1984.webp" href="{att_link}bloom_county,_july_8,_1984.webp" align="none" caption="Bloom County, July 8, 1984" scale="60%"></ft> <ft>To cut you off at the pass: no, I didn't try again. ChatGPT very obviously understood the task. Look at that answer! It knew exactly what I wanted. It just. Can't. Math. This kind of behavior used to be considered a <i>bug</i> in software. Now the purveyors of this buggy software have <i>gaslighted</i> you into treating it as an <i>opportunity</i> to <i>play</i>! Now you have to <i>blame yourself</i> for asking it incorrectly and <a href="https://quoteinvestigator.com/2013/08/11/try-again/">try, try again. <i>Ohne mich.</i></a></ft>