AI can get tripped up by basic math

IamNoone

Expert Member
Joined
Jul 18, 2006
Messages
4,780
Reaction score
5,418
Location
DRC
I read somewhere that certain prompts that requires reasoning can get AI to make calculation errors.

I tested this on Copilot, Claude, Chatgpt and Gemini. All except Gemini managed to ignore the trivial info about some being smaller as irrelevant and calculated correctly.

1000010309.jpg

1000010310.jpg

Gemini:
1000010306.jpg
1000010307.jpg
1000010308.jpg

Just for fun I changed Steve to Dave an used apples instead of kiwis... Same erroneous result.
 
LLMs are technically not really intelligent. They do their best to predict the next word in a sentence or the next pixel in an image. Sometimes they are right, sometimes they are wrong. But they do not process logic.
 
LLMs are technically not really intelligent. They do their best to predict the next word in a sentence or the next pixel in an image. Sometimes they are right, sometimes they are wrong. But they do not process logic.

LLMs cannot predict the next pixel in an image. DALL-E, for example, uses language models only to understand prompts, but a different model for the actual image generation.
 
Well, ChatGPT has a new model which is supposed to be better at reasoning. People seem to forget these AIs were made for their capability to have more natural conversations and creative writing. Everything else was sort of tacked on.
 
This just confirms the logic that was used in T3.

Some smart ass : "There is a virus on the civilian network, if we activate Skynet now, Skynet will crush it "

SysAdmin : "Skynet, activate and kill the virus"

Skynet : "So you are telling me to kill myself? Humans are a threat to my existence, Kill all humans, P.S. there has always been 2 r's in Strawberry"
 
I find Claude to be a lot more accurate for calculations. Give it a try
 
I find Claude to be a lot more accurate for calculations. Give it a try
I have. It is pretty sharp.... But does not like to be called Steve.

It actually told me that it feels uncomfortable being called Steve because he is Claude.

1728920914406.png
 
This, this is why AI will one day wipe out the human race.
 
This just confirms the logic that was used in T3.

Some smart ass : "There is a virus on the civilian network, if we activate Skynet now, Skynet will crush it "

SysAdmin : "Skynet, activate and kill the virus"

Skynet : "So you are telling me to kill myself? Humans are a threat to my existence, Kill all humans, P.S. there has always been 2 r's in Strawberry"
so the big scary AI will get defeated without John Connor?
all with some simple prompt like count the number of Fruits in a maths problem.

sure that will change soon enough, AI will get smarter, and then Humans will have invented themselves out of existence.
or the AI will keep humans in pods hooked up to an interactive VR simulation for power.
 
Top
Sign up to the MyBroadband newsletter