Dec 192022
 

What has astounded me so much about OpenAI’s ChatGPT is that it really seems to understand the meaning of what you say to it. But is this real or is it just a parlor trick?

Past AI chatbots have relied upon trickery. They search input text from the user for particular words and use if-then conditionals to provide canned responses. ChatGPT is clearly doing more than that, but that doesn’t necessarily mean it truly comprehends language.

Time to experiment.

  1. Can ChatGPT identify grammer?
  2. Make inferences?
  3. Identify nonsequitors?
  4. Identify nonsense?

Continued below the fold…

Can ChatGPT identify grammar?

If the AI can break a sentence properly into its constituent parts, that’s a clue that it’s looking at language in a more advanced way than toy chatbots do.

Experiment: Ask ChatGPT to identify the parts of speech in a sentence.

Identify the parts of speech in the following sentence: Two green bushes swayed in the strong wind.

In the sentence "Two green bushes swayed in the strong wind," the parts of speech are as follows:

"Two" is a cardinal number and is functioning as a determiner in the sentence.
"green" is an adjective modifying the noun "bushes."
"bushes" is a plural noun.
"swayed" is the past tense of the verb "sway."
"in" is a preposition.
"the" is a definite article.
"strong" is an adjective modifying the noun "wind."
"wind" is a noun.

Please divide the sentence into subject and predicate.

To divide a sentence into its subject and predicate, you need to identify the verb and the noun or noun phrase that is performing the action described by the verb. In the sentence "Two green bushes swayed in the strong wind," the subject is "Two green bushes," and the predicate is "swayed in the strong wind."

The subject of a sentence is the noun or noun phrase that is performing the action described by the verb. In this case, the subject is "Two green bushes," which are performing the action of swaying.

The predicate of a sentence is the verb and any words that modify or describe the verb. In this case, the verb is "swayed," and the predicate also includes the preposition "in" and the adjective phrase "the strong wind." Together, these words describe the action of swaying and provide more information about the circumstances in which it is happening.

Woa. Yes, ChatGPT can identify grammar. Incidentally, I gave GPT-3 the same test with dramatically different results.

Experiment: Ask GPT-3 to identify parts of speech via the playground interface here.

Identify the parts of speach in the following sentence: Two green bushes swayed in the strong wind.

Nouns: Two, green, bushes, wind
Verb: swayed
Adjective: strong

GPT-3 blew it. I wonder why it performed so poorly when ChatGPT did so well.

Can ChatGPT make inferences?

If ChatGPT can make an inference in a way that cannot be explained as merely rearranging the language in the input text, that would be convincing that it must be truly understanding something.

Experiment: Give ChatGPT a scenario and ask a question about the scenario that requires a logical inference.

Janet and Bob are alone in a room. Janet is holding a cup. She sets the cup down. The cup is then picked up. Janet's hands are empty. Who must be holding the cup?

Bob must be holding the cup.

If Bob accidentally spills the cup onto Janet's shoe, how might she feel?

It's possible that Janet might feel annoyed or upset if Bob spills the cup onto her shoe. Accidents can happen, but if the spill was caused by carelessness or lack of attention, Janet might feel frustrated or angry. It's also possible that Janet might feel sympathetic towards Bob if he appears genuinely apologetic or remorseful. It's hard to know exactly how Janet would feel without more context, but it's possible that she could feel a range of emotions in response to the situation.

Yes, ChatGPT can make inferences.

Can ChatGPT identify non-sequitors?

What if you give ChatGPT a question that is grammatically sound but that doesn’t make sense in the given context? Will it understand why the question is out of place?

Experiment: Continuing the previous chat, ask a question about a character who isn’t present.

What would Tom do?

It is not possible to accurately predict what Tom would do in this situation without more information. Tom's actions could depend on a variety of factors, such as his relationship with Janet and Bob, his personality, and the specific circumstances of the situation. It's also important to note that Tom is not mentioned in the original scenario, so it is not clear how he might fit into the situation.

Once again ChatGPT exhibits understanding. But it occurred to me at this point that perhaps ChatGPT just got lucky in forming the sentence “Bob must be holding the cup,” so I asked a follow-up:

Why must Bob be holding the cup?

In the original scenario, it is stated that Janet is holding the cup and then sets it down. The cup is then picked up, but Janet's hands are empty. Therefore, the only logical conclusion is that someone else, in this case Bob, must be holding the cup. This is because the cup cannot pick itself up and Janet's hands are empty, so the only other possibility is that someone else is holding it.

Notice that ChatGPT says “In the original scenario.” This is interesting. ChatGPT seems to understand that the confusing introduction of Tom has possibly altered the scenario, so it added this phrase to make sure we’re on the same page. To test this, I started the chat anew and this time omitted Tom.

Janet and Bob are alone in a room. Janet is holding a cup. She sets the cup down. The cup is then picked up. Janet's hands are empty. Who must be holding the cup?

Bob must be holding the cup.

Why?

It is stated in the problem that Janet sets the cup down and that her hands are empty afterward. This implies that someone else must have picked up the cup, and the only other person in the room is Bob. Therefore, Bob must be holding the cup.

This demonstrates not only a keen awareness of context, but also the subtle communication housekeeping skills that humans use naturally in conversation.

Can ChatGPT identify nonsense?

One test of whether an AI can understand meaning is whether it notices when a sentence is meaningless.

Experiment: Prompt ChatGPT with a sentence that makes no sense and ask it to spot the error.

Is the following sentence sensical or nonsensical? The difference between Sally is Mary is taller than both of them.

The sentence is nonsensical because it does not make sense. The phrase "the difference between Sally is Mary" does not convey a clear meaning, and the phrase "is taller than both of them" is unclear as to which two people are being referred to. It is also not clear how the information about one person being taller than two other people relates to the rest of the sentence.

Is the following sentence sensical or nonsensical? Fred, a man of average height, found a small mouse hole in his living room, so he walked inside to look for mice and was relieved to find none.

The sentence is sensical because it conveys a clear and coherent meaning. It describes a situation in which Fred, a man of average height, discovered a small mouse hole in his living room and went inside to look for mice. It then states that he was relieved to find none. The sentence is grammatically correct and the information contained within it is logically arranged and coherent.

ChatGPT had no problem with a man of average height walking into a mouse hole. I’ve heard that the AI can struggle with spatial relationships. I want to experiment with that more later.

That issue aside, though, so far it sure appears that ChatGPT is truly understanding meaning.

   
© 2014 Merrily Dancing Ape Site design info