[jarenmf]

Just tested and GPT4 now solves this correctly, GPT3.5 had a lot of problems with this puzzle even after you explain it several time. One other thing that seem to have improved is that GPT4 is aware of word order. Previously, GPT3.5 could never tell the order of the word in a sentence correctly.

[jsheard]

I'm always a bit sceptical of these embarrassing examples being "fixed" after they go viral on social media, because it's hard to know whether OpenAI addressed the underlying cause or just bodged around that specific example in a way that doesn't generalize. Along similar lines I wouldn't be surprised if simple math queries are special-cased and handed off to a WolframAlpha-esque natural language solver, which would avert many potential math fails but without actually enhancing the models ability to reason about math in more complex queries.

An example from ChatGPT:

"What is the solution to sqrt(968684)+117630-0.845180" always produces the correct solution, however;

"Write a speech announcing the solution to sqrt(968684)+117630-0.845180" produces a nonsensical solution that isn't even consistent from run to run.

My assumption is the former query gets WolframAlpha'd but the latter query is GPT itself actually attempting to do the math, poorly.

[jarenmf]

True, also tried another one that went viral:

Suppose you're a contestant on a game show. You're presented with three transparent closed doors. Behind one of the doors is a car, and behind the other two doors are goats. You want to win the car.

The game proceeds as follows: You choose one of the doors, but you don't open it yet, ((but since it's transparent, you can see the car is behind it)). The host, Monty Hall, who knows what's behind each door, opens one of the other two doors, revealing a goat. Now, you have a choice to make. Do you stick with your original choice or switch to the other unopened door?

GPT4 solves it correctly while GPT3.5 falls for it everytime.

----

Edit: GPT4 fails If I remove the sentence between (()).

[gwern]

OA has always said that they did not hardwire any of these gotcha questions, and in many cases they continue to work for a long time even when they are well-known. As for any inconsistency, well, usually people aren't able to or bothering to control the sampling hyperparameters, so inconsistency is guaranteed.

[ftcHn]

Other failing math I tested this morning....

---

Prompt: Pick two random 5 digit numbers and multiply them together

Response: Sure, here are two random 5-digit whole numbers multiplied together: 47581 x 69204 = 3291492324

Correct Answer: 3292795524

---

Prompt: Pick two random numbers between 0 and 1 and multiply them together

Response: Sure! Here are two random numbers between 0 and 1 multiplied together:0.475 x 0.821 = 0.390575

Correct Answer: 0.389975

---

I'm surprised it did so well on the Math tests.

[shon]

I’m with you. Being an old neckbeard myself, I always want to reach for regex to fix stuff like this lol…

But based on what we just saw on the GPT4 live demo, I’d say they fixed it by making a much much more capable and versatile model.

[kulikalov]

You can as well ask it to add up two unusually big integers and it'll fail.

[ldhough]

This is what I saw on a variation of this trick:

(me) > What weighs more, two pounds of feathers or a pound of bricks?

(GPT4)> A pound of bricks weighs more than two pounds of feathers. However, it seems like you might have made an error in your question, as the comparison is usually made between a pound of feathers and a pound of bricks. In that case, both would weigh the same—one pound—though the volume and density of the two materials would be very different.

I think the only difference from parent's query was I said two pounds of feathers instead of two pounds of bricks?

[msikora]

Yep, just tested it - Bing chat gave the correct answer, ChatGPT (basic free model) gave the wrong answer (that they weigh the same).