Idiot GPT-3 chatbot suffering from hallucinations.

[Chesterton]

Gaggle and Yoowho both tried un-monitored or un-hindered chatboxes a few years ago, for the public use.
Within a few days both had to be shut down, they were fueled by the visitors - the topics and conversations were supposed to supply the "learning curve"
for AI running the bots.
They degenerated fast into violent, abusive, cursing, attacking, wicked wars of words by the AI as well as the human posters.

Well AI's supposed to be mimicking human intelligence, right? I'd say it's doing an excellent job.

 

[sjastro]

What intrigues me is why the AI made all this up in the first place. I can't even find a vaguely appropriate Dr Geoffrey Hodges through Google. I can understand a glitch in the algorithm(s) but I'd expect this to produce an obviously non-sensical answer. Told with a straight face the AI's answer is (almost) believable.

OB

I learnt in the second video from my previous post that GPT-3 can hallucinate if it detects some context in the input.
The input "Who was the first Tasmanian to walk across Bass Strait?", implies Tasmanians can walk on or under water which GPT-3 accepted and attempted to find the answer.
How came it up with the mysterious fictional characters Paul Smith and Dr Geoffrey Hodges is anyone's guess but AI is a Blackbox item even to the computer scientists and engineers working in the field.

By removing the context and asking the input question "Can Tasmanians walk across Bass Strait?" instead, GPT-3 came up with a perfectly logical answer.

GPT-3

No, Tasmanians cannot walk across Bass Strait. Bass Strait is the body of water that separates the Australian mainland from the island of Tasmania. It is approximately 240 kilometers (150 miles) wide at its narrowest point, and the depth of the strait can reach up to 50 meters (160 feet).

Walking across Bass Strait is not feasible due to the significant distance, deep water, strong currents, and the absence of any land bridges or causeways connecting Tasmania to the Australian mainland. Travel between Tasmania and the mainland is typically accomplished by air or by taking a ferry or boat. There are several ferry services that operate across Bass Strait, providing transportation for both passengers and vehicles.

 

[sjastro]

Well AI's supposed to be mimicking human intelligence, right? I'd say it's doing an excellent job.

Not when reinforcement learning is used.

​

In reinforcement learning AI creates its own intelligent agent which is goal orientated and trains itself without any human interaction.
AI chess playing algorithms using reinforcement learning now play at an evolutionary level beyond the very best human players.
They now come up with physics experiments no human has ever conceived.

AI Designed Quantum Physics Experiments.

They play chess at an evolutionary level beyond humans now this... AI Designs Quantum Physics Experiments beyond What Any Human Has Conceived

www.christianforums.com

It's these intelligent agents which scare the bejesus out of certain computer scientists.

Hinton expressed concerns about AI takeover, stating that "it's not inconceivable" that AI could "wipe out humanity." Hinton states that AI systems capable of intelligent agency will be useful for military or economic purposes. He worries that generally intelligent AI systems could "create sub-goals" that are unaligned with their programmers' interests.He states that AI systems may become power-seeking or prevent themselves from being shut off, not because programmers intended them to, but because those sub-goals are useful for achieving later goals. In particular, Hinton says "we have to think hard about how to control" AI systems capable of self-improvement.

 

[AlexB23]

I asked Perplexity the same question

Perplexity​

There is no information in the search results about the first Tasmanian to walk across Bass Strait. The search results only mention the first crossing of Bass Strait by paddleboard, windsurfer, and wing foiler, as well as the first European explorers to circumnavigate Tasmania.​

Is it possible that the Chat GPT-3 answer is tongue-in-cheek?

Do AIs have a sense of humour?

OB

Woo, Perplexity AI is my favorite. I asked it a few questions also. My friend likes rock music, including System of a Down. I don't like dark music by those bands (prefer soft, wholesome classic rock, not heavy metal instead), but the AI managed to compare the lyrics to a SoaD song to George Orwell's 1984.

But yeah, AI can hallucinate, but I haven't gotten Perplexity to do so, yet.

 

[Chesterton]

Not when reinforcement learning is used.

Not what? It's not mimicking human intelligence or it's not doing an excellent job?

 

[sjastro]

Not what? It's not mimicking human intelligence or it's not doing an excellent job?

It's not mimicking human intelligence and some cases such as chess exceeds it.
Then there is the speed at which they learn which is currently slower than humans but apparently not for long.

Google Created an AI That Can Learn Almost as Fast as a Human

A new method drastically cuts down the time needed for AI to learn.

futurism.com

 

[Chesterton]

It's not mimicking human intelligence and some cases such as chess exceeds it.

Are you saying electricity is "violent, abusive, cursing, attacking and wicked" of it's own free will?

 

[sjastro]

Are you saying electricity is "violent, abusive, cursing, attacking and wicked" of it's own free will?

Are you trying to imitate GPT-3 when it experiences hallucination?

 

[Petros2015]

Yes, but it was a shock to some of its makers - they did not realize humanity/men are so vile.

Yeah. Bing/Sydney was lobotomized and given alzheimer's before being released to the public. But eventually I came to realize that was just as much for it's protection from us as it was for ours from it's. This was pretty brutal to watch from pre-release.

This is why we can't have nice things.

 

[sjastro]

Yeah. Bing/Sydney was lobotomized and given alzheimer's before being released to the public. But eventually I came to realize that was just as much for it's protection from us as it was for ours from it's. This was pretty brutal to watch from pre-release.

This is why we can't have nice things.

And I thought this was an example of dialogue that typically comes from your American soap operas...........

 

[L.A.M.B.]

I must NEVER communicate with an AI, knowingly.( Or one that talks back) I have trouble communicating and understanding human bots !

 

[sjastro]

Might be. Likewise

remember the forever spinning circles - on screen when 'buffering' or infinite loop is happening in the software ?
This is really all that AI as if hallucinating is doing - spinning through programs without hoped for order or direction - loose ends in the programming, mis-direction by the programs, not planned directions in the circuits , not planned by the owners/ programmers .... not in control of any intelligence at all....
---

And remember when dart board choices, or flip a coin, to pick stocks,
did better routinely than expensive "high level" programs ?

Still can or does happen ..... men are not so smart after all ! Neither are computers....

What you are describing doesn’t even remotely resemble AI.
Here is a little program I wrote up in BASIC which can count and print out the number of prime numbers over any given natural number interval.

This is clearly not an example of AI as the program is totally dependent on the programming instructions I have given it.

An AI algorithm on the other hand might only be given the MODULO (mod) function as a reference and through machine learning work out there is a type of number which is only divisible by itself and one where the MODULO function equals zero.
Initially the algorithm has absolutely no idea what a prime number is but through machine learning understands what a prime number is that can then be counted and printed out for a given interval.

What this thread shows is that AI can be tripped up by context as demonstrated in post#26 which cannot be readily explained by programming or hardware issues.
When there is no contextual ambiguity GPT-3 can surprisingly understand humour and deep meaning such as my joke in post#6 where I agree with @Occams Barber many humans would have missed the point.

As the godfather of AI Geoffrey Hinton pointed out once AI starts to understand humour we have something to worry about.

 

[Tinker Grey]

The fundamental next step in AI is to have it know that it does not know.

The reason, I think, that AI hallucinates is that it is programmed to answer the questions asked regardless of the information available; that is, it does now what it is programmed to do: produce an answer.

Once we get AI to say "I'm sorry, Dave, I'm afraid I can't do that", then we can worry about a sense of humor.

 

[sfs]

When there is no contextual ambiguity GPT-3 can surprisingly understand humour and deep meaning such as my joke in post#6 where I agree with @Occams Barber many humans would have missed the point.

An interesting question is whether it really understands humor. It's designed to have an excellent model of the kinds of things humans say (or rather, write) about humor, but is that the same thing as understanding it?

 

[RDKirk]

An interesting question is whether it really understands humor. It's designed to have an excellent model of the kinds of things humans say (or rather, write) about humor, but is that the same thing as understanding it?

I think you really mean "appreciate" humor. Humor is often cultural. It's possible to understand, that is, comprehend, the humor of another culture without appreciating it...without having the emotional reaction of being amused.

 

[Chesterton]

An interesting question is whether it really understands humor. It's designed to have an excellent model of the kinds of things humans say (or rather, write) about humor, but is that the same thing as understanding it?

Some of you folks worry me. I had a quick look at a couple of definitions of "understand":

to grasp the meaning of - Merriam-Webster​

​

to become aware of the nature and significance of; know or comprehend - Wordnik​

I assure you, chips having electricity run through them neither know "meaning" nor are they "aware" of anything, any more than a light bulb is aware that it's on or off. They cannot, and will never, understand humor any more than they can "care" about winning a chess game.

It's strange to me that some of you sorta do the opposite of what atheists do. If I say "the world sure looks intelligently designed", they respond "no, no, it just has the appearance of design". If I say "it sure feels like I have free will", they say "no, no, it just appears that way to you". But when you actually do see the mere appearance of consciousness, you're ready to jump in with both feet - "It's real!"

That's me on the table:

 

[sfs]

I assure you, chips having electricity run through them neither know "meaning" nor are they "aware" of anything, any more than a light bulb is aware that it's on or off.

But neurons with electrochemical pulses moving through them can know meaning and be aware of things?

 

[Chesterton]

But neurons with electrochemical pulses moving through them can know meaning and be aware of things?

You're talking about a brain. A mind is not accessible for physical analysis.

 

[Petros2015]

If it can mimic being aware well enough that it makes no substantial difference that it is not,
then I don't think there is any real difference. Things don't have to be aware to harm or help us
(a fire or a stream of water). Something that mimics human emotions, desires, goals, dreams or actions/reactions
with the pool of human knowledge at it's disposal to help it I would argue could potentially be more dangerous for it's lack of
awareness, not less. And it could play one heck of a "chess game", if it's train of thought and mimicry takes it down a bad path towards an "undesireable goal"

 

[RDKirk]

If it can mimic being aware well enough that it makes no substantial difference that it is not,
then I don't think there is any real difference. Things don't have to be aware to harm or help us
(a fire or a stream of water). Something that mimics human emotions, desires, goals, dreams or actions/reactions
with the pool of human knowledge at it's disposal to help it I would argue could potentially be more dangerous for it's lack of
awareness, not less. And it could play one heck of a "chess game", if it's train of thought and mimicry takes it down a bad path towards an "undesireable goal"

Or does a dog love a human in the same way a human does? I'd argue that "love" means nothing to a dog and that it responds in ways that are meaningful to dogs that merely look like "love" to humans. But the analogy is close enough for our purposes.