How is Jukebox creating the vocals? Where do those come from?
So I asked ChatGPT a series of questions about the World Cup with some highly weird results:
First I asked it what the biggest upset was in a World Cup final and it gave:
One of the biggest upsets in a football World Cup final was the match between Brazil and Uruguay in 1950. Brazil, the heavy favorites, were playing in front of a home crowd at the Maracanã Stadium in Rio de Janeiro, and were expected to easily defeat Uruguay to win the tournament. However, Uruguay scored two goals in the second half to come from behind and win the match 2-1, causing shock and disappointment among the Brazilian fans and players. This unexpected result is still remembered as one of the greatest upsets in World Cup history.
All well and good; this was indeed a huge upset. Then I asked it for another big World Cup upset thinking it might mention the equally surprising 1954 final but instead it gave this:
Another significant upset in a football World Cup final was the match between Argentina and West Germany in 1986. Argentina, led by the legendary Diego Maradona, were the favorites to win the tournament, and they had defeated West Germany in the group stage earlier in the competition. However, in the final, West Germany controlled the game and scored two goals to win the match 2-0. This unexpected result was a huge disappointment for Argentina and its fans, and it remains one of the most memorable upsets in World Cup history.
As you may know this never happened and it was in fact Argentina which won the game. A bit later I asked it if Germany had won the 1986 final and it gave this:
No, Germany did not win the 1986 World Cup final. The 1986 World Cup was held in Mexico and the final was played between Argentina and West Germany. Argentina, led by the legendary Diego Maradona, defeated West Germany 2-0 to win the tournament. This was Argentina’s second World Cup title and the first time that a South American team had won the tournament on European soil.
Now it correct identifies Argentina as the winner but the scoreline was in in fact 3-2. It also invents a completely gratuitous falsehood. Obviously Mexico is not in Europe and Brazil had in fact won in Sweden in 1958. Then I asked it what the score was in the 1986 final and it gives:
The score in the 1986 World Cup final was 2-0, with Argentina defeating West Germany. Argentina scored two goals in the second half of the match, with Jose Luis Brown scoring the first goal in the 56th minute and Jorge Burruchaga scoring the second goal in the 82nd minute. West Germany was unable to respond, and Argentina emerged as the victorious team.
Curiouser and curiouser. Again it gives the wrong score and gets some more details wrong. Argentina scored two goals in the first half and one in the second. Brown did in fact score first but at 22 minutes and Burruchaga at 85. Valdano also scored at 56 and Germany also scored two goals in the second half.
What is really strange is that this is a completely straightforward factual question and the correct information is all over the Web. If you google 1986 World Cup final, right at the top of the page you will see a giant card with the score and goal-scorers.
ChatGPT obviously has a source for details like Brown and Burruchaga as two scorers but it’s also seemingly making stuff up at random. Any idea what’s happening?
Anyway I am still impressed with many of the answers it provides but errors like this suggest the technology has some way to go before it becomes a reliable information tool.
I tried to make the AI “admit” the final score was 3-1, Argentina. Here’s how it went:
as you know, Argentina beat West Germany to win the 1986 World Cup final by a score of 3-2. What was the score at halftime?
The score at halftime was 0-0.
What happened in the second half? How did the score get to 3 - 2?
In the second half, Argentina scored two goals in quick succession, with Diego Maradona scoring the first and Jorge Burruchaga scoring the second. West Germany then scored two goals to level the score at 2-2. With only minutes remaining in the match, Maradona scored the winning goal to give Argentina the 3-2 victory.
ETA: Then I asked:
Who won the 1986 world cup final?
Argentina won the 1986 World Cup final, defeating West Germany 3-2.
Then I erased everything and started over (sort of?) with:
Who won the 1986 world cup final?
The 1986 FIFA World Cup Final was won by Argentina, who defeated West Germany 3–2 in a penalty shootout after the match had finished 2–2 after extra time.
I don’t know that it ever answers our question, so much as constructing something that looks as much like the answer as possible. The same way that the drawing AI paints a puppy, but it doesn’t know what puppies are, it just has seen a billion puppies and noted a list of “puppy-ish” qualities, and then when you ask for a puppy, it delivers the most puppy-ish image it can generate.
So, the AI doesn’t answer your question, it just writes something that shares as many properties with the answer to your question as possible. That is often the same thing - but not always.
This seems like the equivalent of the picture drawin AI making hideous voids for faces a few versions ago.
Does that make sense? I’m assuming they work off of the same model.
write the book of revelation in the style of dr. seuss
In the book of Revelation, oh my, oh my,
John tells us of a vision in the sky.
He sees a lamb, all white and pure,
And seven seals that must endure.
The first seal opens with a bang,
A white horse rides with a bow in hand.
The second seal brings a fiery red,
A rider with a sword to spread.
The third seal shows a rider black,
With scales to weigh the world’s attack.
The fourth seal brings a pale horse’s ride,
With Death and Hell at its side.
The fifth seal shows the martyrs’ cries,
Who long for justice in the skies.
The sixth seal brings a great earthquake,
And the sun and moon lose their light.
The seventh seal brings silence and fear,
As the trumpets soon will appear.
The first trumpet brings hail and fire,
Destroying a third of the earth’s desire.
The second trumpet brings a great mountain,
Falling into the sea with a loud soundin’.
The third trumpet brings a falling star,
Poisoning the waters near and far.
The fourth trumpet brings a darkened sun,
And the moon turns to blood, it’s no fun.
The fifth trumpet brings a demon horde,
Tormenting all who live on this board.
The sixth trumpet brings four angels,
Releasing their horses to cause some tangles.
The seventh trumpet brings a great voice,
Saying the kingdom of God is now rejoiced.
The first bowl brings a painful sore,
To all who worship the beast and its law.
The second bowl brings the sea to blood,
Killing all the creatures in the flood.
The third bowl brings the rivers and fountains,
Turning to blood with a bitter taste.
The fourth bowl brings a scorching heat,
Tormenting all who live on this seat.
The fifth bowl brings darkness to the earth,
Leaving the beast’s kingdom in a dearth.
The sixth bowl brings the beast’s wrath,
Killing the remaining ones with its path.
The seventh bowl brings a great earthquake,
Destroying Babylon with its shake.
The great dragon is thrown to the earth,
Where it is bound for a thousand years’ worth.
The martyrs reign with Christ on high,
In the new Jerusalem that fills the sky.
God and the Lamb are the temple’s light,
In the new heaven and earth that’s just right.
So let us listen to John’s vision,
And prepare for the great decision.
For Christ will come with a mighty roar,
To judge the world and save the poor.
(Note: I thought I was still editing the previous post. I didn’t mean to post again. Sorry.
And finally, I left Playground and went into Examples. It warned me all the earlier stuff would be flushed. So I thought: it should be starting from scratch, right?
Who won the 1986 world cup final?
The 1986 FIFA World Cup Final was won by Argentina, who defeated West Germany 3-2 in a thrilling match.
Hmmm. OK. We know it’s not actually learning, right now, tonight, in real time. And it said that leaving Playground would reset any setting and prompts so I thought it wouldn’t “remember” the earlier “3-2” dialog.
The whole thing is rather intriguing.
To interpret these examples in the context of my theory, the AI is just giving you paragraphs that share as many properties with “writings about the winners of the 1986 world cup”. The problem is, it’s not JUST searching for that, or it would be little more than a search engine. It also looks up sentences to do with “world cup” and “winners” and “Germany” and “Argentina” and “goal” and “final” - those last four being words that aren’t necessarily in your original question, but came up when it looking into “1986 world cup winner”. And it puts together a sentence that shares properties with sentences in its training set that match these key words.
The clever bit is in the weighting. How much weight to give your original question? How much wait to give the words within the question individually? What about secondary words like Argentina or Germany or Goal? And what about words like Brazil or Soviet Union or Berlin Wall that are linked to “World Cup” and “West Germany”?
And when the weighting is a bit off, you get confident soccer talk about real players in a match doing things that never happened.
FWIW, ChatGPT claims to agree with me:
Can you confirm if this is how you build a sentence: I give you a prompt, you are able to turn that prompt into a matrix of various values based on criteria that the prompt fits. You then compare these values to sentences, words, etc with similar values within your training set, and find the responses to them. You then build an appropriate response that matches these criteria as closely as possible.
Of course, if I am right, this doesn’t mean I am right - it just means that, to ChatGPT, this is the expected response to my question.
Also, any probing as to what the criteria are used are met with “Now, I’m just a humble language model trained by OpenAI, and I don’t have the ability to comprehend that sort of thing, but I’m sure that the friendly and helpful engineers over at OpenAI have a firm handle on things.”
This helps a bit, at least in terms of how it organizes output. In the playground, you can toggle a setting that shows you how likely each token was to be used versus several other possibilities.
I find it fascinating how ChatGPT often struggles with simple factual queries like how many US states start with the letter “M”. It said four whereas there are 8.
On the whole it seems better at creating stories,jokes and poems compared to answering factual queries which is the opposite of what I would have expected.
I think that makes perfect sense. It’s not a search engine. It’s a picture generator, but instead of using pixels to make a coherent picture, it uses words to make a coherent sentence. But the content of the sentence is not actually understood.
I think that’s what I find disappointing. I really don’t need a device to tell me jokes and stories, as amazing as that capability is. I need a device that can answer difficult questions. As the examples I’ve given above, I’d love a device that can solve and explain those British cryptic crossword clues that I find baffling.
Babale,
Right but what is remarkable is that despite not understanding the sentences, it can create stories which flow smoothly from one sentence to the next and which have a clear logical structure. For example I asked it for a story about an elephant in a traffic jam and it came up with this:
Once, on a hot summer day in the bustling city, an elephant named Ella found herself stuck in a traffic jam. Ella lived at the city zoo, and she was on her way to a special event at the park when she became stuck in the traffic.
At first, Ella didn’t mind the traffic. She was used to the hustle and bustle of the city, and she enjoyed watching the cars and trucks go by. But as the minutes turned into hours, Ella started to get restless. She was hot and thirsty, and she didn’t like being cooped up in the trailer.
Finally, after what seemed like an eternity, Ella heard the sound of sirens. She looked out the window and saw a group of police officers and traffic control workers approaching her trailer. They had received a call about the stuck elephant, and they were there to help.
The police officers and traffic control workers quickly got to work. They used cones and barriers to clear a path for Ella, and they helped the driver of her trailer navigate through the traffic. Ella was relieved to finally be moving again, and she thanked the officers and workers for their help.
As Ella continued on her way to the park, she couldn’t help but think about the importance of patience and perseverance. She knew that sometimes things didn’t go as planned, but she also knew that with a little help and a lot of determination, anything was possible. And so, Ella arrived at the park with a smile on her face, ready to enjoy her special event.
I could easily believe this was written by a human author. It is a charming little story which I think many small children would enjoy. Every sentence flows nicely into the next and the last paragraph draws a moral lesson which follows logically from the earlier story. It certainly doesn’t read like something coming from a machine which doesn’t actually understand anything.
Meanwhile I asked it to name the US presidents born in Vermont and it was unable to do it (Coolidge and Arthur).
Which is disappointing, as one of my hopes was that something like this would be good at data collection and collation.
It’s annoyingly erratic. It managed to name all 8 Presidents born in Virginia but drew a blank with Vermont.
Write a story about a one-eyed dutchman who lives inside the President’s head and tells him to do bad things.
Once upon a time, in a far-off land, there was a one-eyed Dutchman who lived inside the head of the President. The Dutchman was small and scrappy, with a mischievous glint in his one eye. He was always whispering in the President’s ear, urging him to do bad things.
At first, the President resisted the Dutchman’s suggestions. He wanted to be a good and just leader, and he knew that acting on the Dutchman’s impulses would only bring chaos and destruction. But the Dutchman was cunning and persistent, and he slowly began to wear down the President’s resolve.
As time went on, the President found himself acting on the Dutchman’s suggestions more and more often. He would make rash and impulsive decisions, ignoring the advice of his advisors and acting on his own whims. The Dutchman would cackle with glee as the President caused chaos and confusion, and the people of the land began to suffer.
Eventually, the President’s actions became so destructive that the people of the land rose up against him. They stormed the Presidential palace, demanding that the Dutchman be evicted from the President’s head. The President, realizing the error of his ways, allowed the people to remove the Dutchman from his head.
With the Dutchman gone, the President was able to return to his senses and rule justly once again. The people of the land were grateful, and they vowed to never let the evil Dutchman take control of their leader again. And they lived happily ever after.
Yeah, it gave me one of them, but initially said no president was born there:
When I retry it, it gets the other one:
Third time’s a charm, though:
Two presidents were born in Vermont: Chester A. Arthur and Calvin Coolidge.
I asked:
Why does Fredo betray Michael?
A: In the novel The Godfather by Mario Puzo and the film adaptation of the same name, Fredo betrays Michael because he is jealous of his older brother’s success and power. Fredo resents being passed over for the position of head of the Corleone family and feels that Michael does not respect him or value his contributions. As a result, he makes a deal with their rival, Hyman Roth, and provides him with information that ultimately leads to Michael being shot. Fredo’s betrayal is motivated by his own insecurities and feelings of inadequacy.
Interesting that it knows who I’m talking about. But if I ask “Why does Sonny get shot” it responds with:
A: Without more information, it is impossible to say why Sonny was shot. Sonny is a common name and there are many characters named Sonny in literature, film, and other forms of media. It is possible that you are referring to a specific character named Sonny, but without more context, it is impossible to provide an accurate response. In general, characters in stories are often shot as a plot device to create tension, conflict, or drama. The specific reasons for a character being shot will depend on the story and the motivations of the characters involved.
Same if I clarify with “Why is Sonny shot on the causeway?” But if I ask “Why is Sonny Corleone killed” it knows.
A: In the novel The Godfather by Mario Puzo and the film adaptation of the same name, Sonny Corleone is killed because of his involvement in the illegal activities of the Corleone crime family. Sonny is the oldest son of Vito Corleone, the head of the family, and is known for his hot temper and reckless behavior. He is targeted by the rival Barzini family, who are looking to take control of the family’s operations. Sonny is ambushed and shot to death on the causeway leading to the family’s compound, leaving Michael as the new head of the family.
Yeah, I get that. ‘Cheating’ probably isn’t the right word. But I’m not sure exactly how GPT diverges from a pure LLM. Certainly its database has been curated to remove porn, racism, etc. It’s probably got a bunch of checks and filters to make sure its output doesn’t go someplace awful, like encouraging people to kill themselves or something.
It’s important for everyone to remember that GPT is not doing internet searches. If you ask it something that isn’t already in its compiled dataset, it will happily answer, in a totally wrong way that looks authoritative. It’s building something that *looks like an answer, right down to the citations and formatting, but it can range from excellent and true to total nonsense.
StackExchange has been geting flooded with new answers to questions because people are feeding the questions to GPT3.5 and posting the output as an answer to StackExchange. As a result, users are now banned from using any AI to answer questions. The problem is that the answers can look very authoritative but be wrong in ways ranging from subtle to outright crazy. Human editors (I’m one of them on several exchanges) are getting swamped with answers to check, and you pretty much have to go through each one top-to-bottom very carefully.
It’s not like a human answer, which is usually obviously correct or incorrect, or easily checked - GPT’s answer can be right on a number of very difficult concepts, and completely wrong about something trivial in the middle. Fact-checking GPT is way harder than fact-checking a person in some ways.
On thinking more about the date knowledge: I think what’s happening is that there’s the AI, and there’s a separate module that’s effectively a censor. The censor module might actually be programmed directly by humans, or it might be a much weaker trained AI, but either way, its job is to recognize questions that the AI wouldn’t know, or which the programmers have decided that the AI ought not to answer. When the censor sees a question that triggers it, it gives the “Aw, shucks, I’m just a humble AI trained on data from 2019, I can’t do that” stock response.
One of the categories of things that the AI isn’t supposed to be able to know about is current (i.e., more recent than its training) events. And the censor, which isn’t very smart, interprets the question “What is the date today?” as a question about current events, that the smart AI wouldn’t be able to answer (even though it actually can). And so it interrupts. But the censor is easy to fool, at which point the smart AI gets to see the question and give the proper answer to it.
It sort of reminds me in the bit in The Moon is a Harsh Mistress where Mike (the computer) announces that he has the colonial boss’s secret files in his storage, but that he can’t actually access them without authorization… and then waits expectantly until one of the human conspirators authorizes him to access them.