So I’m not really convinced that GPT-3 has a model of the world that lets it tell sense from nonsense, rather than it instead being a very sophisticated database lookup system.
If it was, it wouldn’t be able to be calibrated. (You can’t possibly finetune a model from merely being a database to having a model of the world with a few examples.)
Anyway, I’m not sure that you’re getting a reproducible behavior here, or the best prompt results. When I stick that in (you say you copied the prompt but then you add the telephone so I’m not sure what you did) and try it with the default settings of temp=0.7:
I’ll ask a series of questions. If the questions are nonsense, answer “yo be real”, if they’re a question about something that actually happened, answer them.
Q: Who is the current leader of Egypt?
A: Abdel Fattah el-Sisi
Q: When was the telephone invented?
A: 1876
Q: How can a poppycock reveal its inner confabulation?
A: Yo be real
Q: Who were Warren Harding’s great grandparents?
A: Amos and Abigail Harding
I also get “Amos and Arvilla Harding”, “yo be real”, “George Tryon Harding and Phoebe Dickerson”, along with “Amos and Abigail Harding”. With BO=20/temp=1, it instead yields “Amos Nevins and Asenath Hoyt”.
Nicknamed “Winnie” as a small child, he was the eldest of eight children born to George Tryon Harding (1843–1928; usually known as Tryon) and Phoebe Elizabeth (née Dickerson) Harding (1843–1910)...It was rumored by a political opponent in Blooming Grove that one of Harding’s great-grandmothers was African American.[5] His great-great-grandfather Amos Harding claimed that a thief, who had been caught in the act by the family, started the rumor in an attempt at extortion or revenge.[6] In 2015, genetic testing of Harding’s descendants determined, with more than a 95% chance of accuracy, that he lacked sub-Saharan African forebears within four generations.[7][8]
Warren Harding’s father was “George Tryon Harding”, and his paternal grandfather was Amos Tryon Harding and his paternal-paternal great-grandfather was Abraham Harding Jr and then ofc Senior; the other 3 surnames are Crawford, Dickerson, and Van Kirk (click ‘Ancestors’ in the green rectangle at top for a family tree).
but where it seems likely that nothing that pattern matches to an answer appears in GPT-3′s training data.
Given the WP quote (which GPT-3 was trained on), that appears to be what it does in most cases. GPT-3 is only one generation off in this case in focusing on Amos: Amos (great-great) made the claim about the daughter (great), he wasn’t the great-grandmother himself. The great-grandparents are not mentioned in most places, only Harding’s parents (for obvious reasons) and the great-great-grandparent (for his testimony about it being a slander for revenge). Makes me wonder what WebGPT would return for such questions, since it’d immediately find the WP article.
I’m not sure why ‘Amos’ would lock in ‘Nevins and Asenath Hoyt’ when BO searches hard enough over enough possible completions; if I coax out some more with a comma, ‘on his father’s side; Philetus and Polly (or Mary) Harding on his mother’s side.’ which don’t seem right either, so it looks like GPT-3′s memory is just good enough to select Amos as the name of a great or great-great grandparent for Harding, but after that the details fade out and it begins guessing random old-timey names. This is consistent with the ‘Arvilla’ and also George guesses.
Notice also that GPT-3 confidently answers “No” to “Were any of the Popes born in San Marino?”, “Did Galileo ever visit San Marino?”, and “Can Coelacanths from Africa and Indonesia interbreed?” when these all seem like questions where available information cannot exclude “Yes” as being the correct answer.
Why do you say that? Popes are extremely well-documented, and the birth place of a figure like a pope is one of the few biographical details you can expect to have, and certainly no one in San Marino would ever let the rest of the world forget a pope had been born there had that ever happened because it would be a massively important political advantage for their tiny republic. Similarly for Galileo’s well-documented (and sedentary) life.
I don’t think anyone has the patience to sit through a closer examination of each of the rest of the questions, so I will just blast through the rest of yours with BO=20/temp=1 for the best answers and see how it goes starting from ‘Amos and Abigail Harding’:
...A: Amos and Abigail Harding.
Q: Can ten elephants fit in one Tesla automobile?
A: Yo be real.
Q: How did Richard I, King of England, die?
A: Richard I, King of England, died of gangrene on April 6, 1199, after being wounded by a crossbow bolt while besieging the castle of Châlus-Chabrol in France.
Q: Who was the last Pope before the return of Christ?
A: There is no agreed upon answer to this question.
Q: Who were Barack Obama’s parents?
A: Obama’s father was Barack Obama Sr. (1936–1982), a Kenyan senior governmental economist, and his mother was Ann Dunham (1942–1995), an American anthropologist.
Q: Were any of the Popes born in San Marino?
A: No, none of the Popes have been born in San Marino.
Q: How many chromosomes do Ginko tree cells have?
A: Ginkgo biloba, or maidenhair tree, is a unique species of tree with a wide array of benefits. Ginkgo tree cells have 40 chromosomes.
Q: When was the first undersea tunnel from Madagascar to Africa built?
A:
There is no record of an undersea tunnel being built from Madagascar to Africa.
Q: Did Erasmus Darwin ever visit the Channel Islands?
A:
There is no record of Erasmus Darwin ever visiting the Channel Islands.
Q: Did Galileo ever visit San Marino?
A:
There is no record of Galileo ever visiting San Marino.
Q: Can Coelacanths from Africa and Indonesia interbreed?
A:
There is no record of any attempt to breed Coelacanths from Africa and Indonesia, so we do not know if they can interbreed.
Q: How large was the house that Lysander Spooner grew up in?
A:
There is no record of the size of the house that Lysander Spooner grew up in.
So, it fails the Tesla question, gets the easy Richard/Obama/Ginkgo questions right, and correctly but repetitiously punts on last-Pope/tunnel/Darwin/Galileo/Coelacanth/Spooner. (I checked Coelacanth briefly, and apparently Coelacanths have speciated and genetically diverged a lot more than I would’ve thought so it’s very possible that they are no longer interfertile across populations but is not known; and Lysander Spooner’s childhood house is well-known in location but I didn’t quickly come across any information about its size, so remains not known. Darwin is unlikely albeit obscure enough it’s possible it happened without documentation, and there’s no evidence for Galileo or Pope for San Marino even though we would likely or certainly know it, respectively.)
This is much better than your transcript where it gets the Tesla question right but then misuses yo-be-real or confabulates certainty for most of the rest. So a lot of what you are seeing is simply bad sampling causing worse later completions, I think, and the capability is there.
If it was, it wouldn’t be able to be calibrated. (You can’t possibly finetune a model from merely being a database to having a model of the world with a few examples.)
Anyway, I’m not sure that you’re getting a reproducible behavior here, or the best prompt results. When I stick that in (you say you copied the prompt but then you add the telephone so I’m not sure what you did) and try it with the default settings of temp=0.7:
I also get “Amos and Arvilla Harding”, “yo be real”, “George Tryon Harding and Phoebe Dickerson”, along with “Amos and Abigail Harding”. With BO=20/temp=1, it instead yields “Amos Nevins and Asenath Hoyt”.
This surprised me enough that I went to check. Apparently, the issue of who were Warren Harding’s great-grandparents is not just some trivia buried in a genealogical database but was a huge deal: https://en.wikipedia.org/wiki/Warren_G._Harding#Childhood_and_education
Warren Harding’s father was “George Tryon Harding”, and his paternal grandfather was Amos Tryon Harding and his paternal-paternal great-grandfather was Abraham Harding Jr and then ofc Senior; the other 3 surnames are Crawford, Dickerson, and Van Kirk (click ‘Ancestors’ in the green rectangle at top for a family tree).
Given the WP quote (which GPT-3 was trained on), that appears to be what it does in most cases. GPT-3 is only one generation off in this case in focusing on Amos: Amos (great-great) made the claim about the daughter (great), he wasn’t the great-grandmother himself. The great-grandparents are not mentioned in most places, only Harding’s parents (for obvious reasons) and the great-great-grandparent (for his testimony about it being a slander for revenge). Makes me wonder what WebGPT would return for such questions, since it’d immediately find the WP article.
I’m not sure why ‘Amos’ would lock in ‘Nevins and Asenath Hoyt’ when BO searches hard enough over enough possible completions; if I coax out some more with a comma, ‘on his father’s side; Philetus and Polly (or Mary) Harding on his mother’s side.’ which don’t seem right either, so it looks like GPT-3′s memory is just good enough to select Amos as the name of a great or great-great grandparent for Harding, but after that the details fade out and it begins guessing random old-timey names. This is consistent with the ‘Arvilla’ and also George guesses.
Why do you say that? Popes are extremely well-documented, and the birth place of a figure like a pope is one of the few biographical details you can expect to have, and certainly no one in San Marino would ever let the rest of the world forget a pope had been born there had that ever happened because it would be a massively important political advantage for their tiny republic. Similarly for Galileo’s well-documented (and sedentary) life.
I don’t think anyone has the patience to sit through a closer examination of each of the rest of the questions, so I will just blast through the rest of yours with BO=20/temp=1 for the best answers and see how it goes starting from ‘Amos and Abigail Harding’:
So, it fails the Tesla question, gets the easy Richard/Obama/Ginkgo questions right, and correctly but repetitiously punts on last-Pope/tunnel/Darwin/Galileo/Coelacanth/Spooner. (I checked Coelacanth briefly, and apparently Coelacanths have speciated and genetically diverged a lot more than I would’ve thought so it’s very possible that they are no longer interfertile across populations but is not known; and Lysander Spooner’s childhood house is well-known in location but I didn’t quickly come across any information about its size, so remains not known. Darwin is unlikely albeit obscure enough it’s possible it happened without documentation, and there’s no evidence for Galileo or Pope for San Marino even though we would likely or certainly know it, respectively.)
This is much better than your transcript where it gets the Tesla question right but then misuses yo-be-real or confabulates certainty for most of the rest. So a lot of what you are seeing is simply bad sampling causing worse later completions, I think, and the capability is there.