[sph]

"Ability to solve novel problems" does not mean ability to solve all problems, nor to have all the knowledge in the world. A caveman can be as intelligent as you and I, even without being able to read.

Not having knowledge in mathematics has no impact on your intelligence quotient.

[terminalcommand]

I agree. But LLMs do solve novel problems in that case, you just have to explain it to them like you would have done to an intelligent caveman or me.

Which novel problem can't an LLM solve? I gave the example of an imaginary game because the LLM cannot have possibly trained on it as it is imagined by one person and nobody knows how to play it.

[latexr]

I tried the first thing which came to mind, with GPT-4o. The problem is probably not novel, yet it failed anyway.

---

User:

Let’s play a game. Every time I give you a word, you separate it into its individual characters and assign a number to each letter, in order, following the Latin alphabet. So "a" is "1", "b" is "2", and so forth. You will then subtract 2 from each number, convert it back to the corresponding letter and return the new word. Did you get all that? Are you ready to start with my next word?

LLM:

Got it! I'm ready to start. What's your first word?

User:

anthropomorfic

LLM:

Sure! Let's break down "anthropomorfic".

1. Assign numbers to each letter:

a (1), n (14), t (20), h (8), r (18), o (15), p (16), o (15), m (13), o (15), r (18), f (6), i (9), c (3)

2. Subtract 2 from each number:

a (1-2=-1: Consider Zuzzer, i.e., `25`), n (14-2=12), t (20-2=18), h (8-2=6), r (18-2=16), o (15-2=13), p (16-2=14), o (15-2=13), m (13-2=11), o (15-2=13), r (18-2=16), f (6-2=4), i (9-2=7), c (3-2=1)

3. Convert them back to letters:

y, l, r, f, p, m, n, m, k, m, p, d, g, y

The new word is "ylrfpmnmkmpdgy".

---

Looks pretty good, right? It even found a good wraparound solution for “a” (I wasn’t trying to trick it, I just didn’t think about the word too hard), though I have no clue what it means by “Consider Zuzzer”. But then you look at the last one, where it converted “c” to “y”. That’s absolutely wrong and is the easiest one any human would guess. Yet you could miss it because all the previous ones were right.

Here’s a bonus. When verifying the answer, I queried Brave search for “letters of the alphabet with numbers assigned”. The AI-generated answer at the top was: A=1, B=2, C=3, D=4, E=5, F=6, G=7, H=8, I=9, J=1, K=2, L=3, M=4, N=5, O=6, P=7, Q=8, R=9, S=1, T=2, U=3, V=4, W=5, X=6, Y=7, Z=8. Which, alright, I can see the logic, but that’s not what anyone would want by asking that question.

[terminalcommand]

For argument's sake I rewrote your prompt. I think gpt 3.5 understood it on the first try:

Prompt: Let’s play a cypher game.

The rules of the game are as follows:

1. Imagine a numbering system where each letter of the English alphabet corresponds to a number starting with 1. For example, a corresponds to 1, b corresponds to 2, c corresponds to three.

2. When you receive a word from me, break it into individual characters.

3. Convert each character into a list of numbers by following rule #1.

4. Subtract two from each number of the list and generate a new list, if the result is lesser than 0, add 26.

5. Convert each number in the new list into a letter by following rule #1 and present me your result.

Are you ready for the first word? ------- GPT: I'm ready! Go ahead and give me the word. Me: anthropomorfic GPT: Sure, let's start:

a -> 1 -> 1 - 2 = 25 (z)

n -> 14 -> 14 - 2 = 12 (l)

t -> 20 -> 20 - 2 = 18 (r)

h -> 8 -> 8 - 2 = 6 (f)

r -> 18 -> 18 - 2 = 16 (p)

o -> 15 -> 15 - 2 = 13 (m)

p -> 16 -> 16 - 2 = 14 (n)

o -> 15 -> 15 - 2 = 13 (m)

m -> 13 -> 13 - 2 = 11 (k)

o -> 15 -> 15 - 2 = 13 (m)

r -> 18 -> 18 - 2 = 16 (p)

f -> 6 -> 6 - 2 = 4 (d)

i -> 9 -> 9 - 2 = 7 (g)

c -> 3 -> 3 - 2 = 1 (a)

So, "anthropomorfic" becomes "zlrfpmmnpmkpdmga". Your turn!

[fireflash38]

Isn't 25 y not z?

[latexr]

Indeed it is. That part is a huge mess. The LLM gave the wrong number when matching with the letter, but the letter itself is correct considering the instructions, except the instructions are wrong (ironic considering this came from someone claiming everyone is doing it wrong). Step four says:

> Subtract two from each number of the list and generate a new list, if the result is lesser than 0, add 26.

It should be “if the result is less than 1”, not “0”.

[terminalcommand]

Well I told you I was bad at math. I wrote a faulty prompt. But chatgpt understood my instructions perfectly. That was the argument I was trying to demonstrate.

The problem lied between the chair and the computer.

We have to learn how to use LLMs.

[fragmede]

Weird, it got ylrfpmnmkmpdga when I tried it.

https://chat.openai.com/share/49ad0bc9-7b3d-4295-860d-4c8168...

Does that make it intelligent, then?

[latexr]

> Weird, it got ylrfpmnmkmpdga when I tried it.

That’s not weird at all. LLMs often give different answers to the same query. Which has been demonstrated several times in this thread.

> Does that make it intelligent, then?

No, it does not, because it isn’t consistent, it demonstrated it doesn’t understand.

https://news.ycombinator.com/item?id=40368446

By your logic, any system which spews random strings is intelligent because sometimes it’s randomness coincidentally aligns with the input you give it.

[fragmede]

we use words like reason and understand and think, and try to apply them to LLMs, when it's just dot products and matrix multiplications at the heart of it, which is where we go wrong. these things are a new alien kind of intelligence, and we're struggling with it because it's completely foreign. it's more than a random coincidence. your logic was that, in your run of the same query, because it made a mistake on the last letter, which is a mistake an inattentive teenager or drunk adult could easily make, we can't consider it intelligent.

we're not talking about any system here, we're talking about LLMs and their ability to generate random coincidental text that does happen to align with the input given. when the output, coincidental and random as it may well be, is aligned with the input in a way that resembles intelligence, we do have to ponder not just what intelligence actually is, but also what it means to be intelligent. octopuses are intelligent but they're not able to solve your particular puzzle.

[terminalcommand]

Try giving an example in your prompt :). I am sure gpt can solve it if you provided a couple of examples.

Also this is not a game it is a cypher. Try specifying that as well.