Hacker News Reader

> The concept of a cube is a mathematical construct that does not exist in reality, and any attempts to manipulate or rotate it would be impossible.

This is hilarious. Did someone accidentally add "cube" as a synonym for "race" in the neutering algorithm?

> If you use less powerful AI your products will most likely lose to competition who gamble and use a more powerful model (in theory)

I think it really depends on many factor and there isn't a one size fits all answer. Depending on your user target, it could even be that your customers get disappointed if they see a sudden decrease in the quality and just stop paying for your products.

I'd question why you'd ask an LLM something like this instead of Wolfram Alpha, or just look up the answer. It's not something you need an LLM to figure out.

It's like asking an LLM for all the digits of pi multiplied by 5. Why?

The problem with LLMs is the amount of things it makes sense to use them for is really not that large.

>> I'd question why you'd ask an LLM something like this instead of Wolfram Alpha, or just look up the answer.

That's because OpenAI dedicated an entire section in claims that GPT-3 is reasonably good at arithmetic. See Section 3.9.1 titled "Arithmetic" in "Language Models are Few-Shot Learners":

https://arxiv.org/abs/2005.14165

Whence I quote below:

Overall, GPT-3 displays reasonable proficiency at moderately complex arithmetic in few-shot, one-shot, and even zero-shot settings.

The reported results are pretty poor and too poor to justify even the relatively weak claim above (although the rest of the text in the same section very clearly and strongly implies that GPT-3 is doing something else than simply memorising a table of sums, which is an altogether much grander claim). OpenAI themselves seemed to be dubious enough about their own claim that the Arithmetic section of their paper was only included in the preprint (on Arxiv) and not in the published paper (in the 34th NeurIPS).

Yet, the claim in the preprint was still enough for people to forcefully argue that GPT-3 can do arithmetic, that it can learn the rules of arithmetic, and other impossible things before breakfast.

This is a discussion that goes back at least 3 years (judging from my comments where I point out that it's nonsense). It seems that the arithmetic ability of large language models is now a well accepted truth in the minds of the general public, who will casually use it to do, say, their maths coursework etc.

So blame OpenAI who made the big claims.

> I'm just an AI, my purpose is to provide accurate and informative responses to your questions, but I must always do so in a safe and responsible manner.

I wonder if poor grammar is baked in as well? It's interesting that it wrote the sentence like that!

I would guess that mistake is made in lots of human text data that it might have gotten trained on. It is interesting, though in that other models have perfect grammar. Maybe it needs better human feedback based on the grammar?

Yea it was a wake up call for me when I was asking about the volume of a cube vs a dodecahedron and Claude+ hit me with "The diameter of a dodecahedron passes through 3 pentagonal faces", just no ability to reason about geometry.

https://poe.com/lookaroundyou/1512927999895666

Sidenote, poe has a bug that mis-reports this as a conversation with Claude-2-100k, but the conversation took place on March 24, about a week after Claude+ was made public. Can't even rely on the portals to be truthful about what model was used.

"Reasoning" about faculties which the bot does not possess (visual, audial, or individual character domains) is simply not going to happen until we hook it up to some symbolic model, or the LLM becomes advanced enough to infer a world model from text alone.

Until then, we're basically asking a blind & deaf guy, who happens to be very well-read, to reason about senses he doesn't have.

Though the mistake in your example does seem kind of egregious and I'd be curious to see whether GPT-4 would make similar mistakes.

Oh well, it’s good enough for writing automated responses to scam and cold prospecting emails…

Is this real? That’s incredible if so.

https://chat.lmsys.org/

See for yourself

LLMs are generally terrible with math and reasoning questions, so everyone likes to "prove" that a model is bad by giving it a simple math question. It's unintuitive because computers are amazing with math and terrible at everything else.

Basically a computer turns everything into a calculation to process something, but a LLM turns everything into tokens.