Peter Welinder posted recently about using tokenization to get GPT-3 to reverse strings by breaking them down into letters, adding numbers, reversing, and then recombining the letters two at a time. One of the replies noted this only works with strings of even length, which is ominous.
There’s a sculpture at the CIA called Kryptos which has four encrypted passages. The last one has been unsolved since 1990. It has 97 letters and the creator has given some hints: the plaintext letters 22 to 34 are “EASTNORTHEAST” and letters 64 to 74 are “BERLINCLOCK”. The plaintext is said to be “plain English”. It has letter for letter correspondence so you have to add spaces to be able to read it easily.
For example, from the English Wikipedia Kryptos Talk page, an example solution is:
In the center of this city, east, northeast, from Rotes Rathaus chimes from the Berlin clock are heard across the square.
I’d like to give GPT-3 a Theme - people suggest Sundials, Compasses, Ancient Egypt, Clepsydras, the Antikythera Mechanism etc - and get a Thematic solution to K4.
But I haven’t been able to work out how to do this in the Playground. It’s not hard telling it to write thematic English text of a certain length, but adding the constraints makes it harder. I am wondering how difficult this is because I never would have been able to come up with the “reversing strings” idea above.
Here is something I think should have a good chance of working based on the example, but doesn’t.
Example:
- Write English Text: "In the center of this city east northeast from Rotes Rathaus chimes from the Berlin clock are heard across the square"
- First remove spaces: "InthecenterofthiscityeastnortheastfromRotesRathauschimesfromtheBerlinclockareheardacrossthesquare"
- Then convert to upper case: "INTHECENTEROFTHISCITYEASTNORTHEASTFROMROTESRATHAUSCHIMESFROMTHEBERLINCLOCKAREHEARDACROSSTHESQUARE"
- Then split into letters: I N T H E C E N T E R O F T H I S C I T Y E A S T N O R T H E A S T F R O M R O T E S R A T H A U S C H I M E S F R O M T H E B E R L I N C L O C K A R E H E A R D A C R O S S T H E S Q U A R E
- Then add numbers: "1:I 2:N 3:T 4:H 5:E 6:C 7:E 8:N 9:T 10:E 11:R 12:O 13:F 14:T 15:H 16:I 17:S 18:C 19:I 20:T 21:Y 22:E 23:A 24:S 25:T 26:N 27:O 28:R 29:T 30:H 31:E 32:A 33:S 34:T 35:F 36:R 37:O 38:M 39:R 40:O 41:T 42:E 43:S 44:R 45:A 46:T 47:H 48:A 49:U 50:S 51:C 52:H 53:I 54:M 55:E 56:S 57:F 58:R 59:O 60:M 61:T 62:H 63:E 64:B 65:E 66:R 67:L 68:I 69:N 70:C 71:L 72:O 73:C 74:K 75:A 76:R 77:E 78:H 79:E 80:A 81:R 82:D 83:A 84:C 85:R 86:O 87:S 88:S 89:T 90:H 91:E 92:S 93:Q 94:U 95:A 96:R 97:E"
Write plain English text that is about dogs, which after removing spaces, converting to upper case, then splitting into letters, is exactly 97 letters long, and after adding numbers, also contains the exact sequences "22:E 23:A 24:S 25:T 26:N 27:O 28:R 29:T 30:H 31:E 32:A 33:S 34:T" and "64:B 65:E 66:R 67:L 68:I 69:N 70:C 71:L 72:O 73:C 74:K"
Still interested in this problem - ChatGPT produces some great thematic solutions, but getting it to obey the constraints is really difficult; e.g. I can specify something like “write a 14 row 7 columns matrix of letters, make sure row 4 begins with EASTNOR, row 5 begins with THEAST, row 10 begins with BERLINC, row 11 begins with LOCK” and it gets close, but forcing the result to be meaningful is hard.
Thanks, that meets the length constraint of 97 but doesn’t contain EASTNORTHEAST and BERLINCLOCK at the required positions.
The example provided does fit those constraints: In the center of this city, east, northeast, from Rotes Rathaus chimes from the Berlin clock are heard across the square.
It might be possible to make ChatGPT respect the constraints, if I was patient enough to explain its mistakes …
