Huh. The algo still relies on ChatGPT correctly multiplying the individual digits by 2256. But maybe it finds cases of the those in it's training set so gets them right?
Ironic that an AI built on flawless massively parallel floating point multiplications struggles with arithmetic at a higher level.
Still incredible you can give it instructions like this in natural language.
You can see the effect clearly in the beta.openai.com playground mode, with Show Probabilities set to Full Spectrum. Ask it to multiply two large numbers and the middle digits will be red (uncertain), fading to green on the outer digits.
It (sorta) explains this:
When you ask me to multiply two large numbers, the result I provide may not be completely accurate. This is because, as a language model, I am designed to generate natural language responses based on the input provided. I am not a calculator or a mathematical tool, and I do not have the ability to perform complex mathematical calculations with complete accuracy.
In some cases, when multiplying large numbers, the middle digits of the result may be incorrect. This is because the multiplication process involves carrying over digits and performing complex operations on large numbers. These calculations can be difficult even for humans, and they are beyond the capabilities of my underlying technology.
Supposedly, multiplications of low numbers works quite well, so I think it should be able to multiply individual digits without problems.
> Ironic that an AI built on flawless massively parallel floating point multiplications struggles with arithmetic at a higher level.
The same thing can be said about our brains though. Our neurons perform millions of calculations each day, but we can still struggle with doing a single on consciously.
Our neurons never perform "flawless" multiplications though. It is just us that interpret the physical proccess going on inside neurons as it. No irony there.
Yes, it executing an algorithm that you just described in prose is nice and all, but please explain to me how this bit works:
> Yes, I understand the process you described. It is similar to the standard multiplication algorithm, but instead of starting with the units digit and moving left, it starts with the most significant digit and moves right. Additionally, it includes an extra step of multiplying the partial product by the PLACE value and adding it to SUM.
Note that there was nothing instructing the network to compare the two algorithms. On the contrary, the poster wanted to avoid any connection between the two as not to confuse the network.
So it knows what a "multiplication algorithm" is (even if it has the wrong idea of it) and has this knowledge readily enough accessible that it can notice if a different algorithm is similar to it - and then compare the two? And all of that unprompted?
Ironic that an AI built on flawless massively parallel floating point multiplications struggles with arithmetic at a higher level.
Still incredible you can give it instructions like this in natural language.