It can't do anagrams though (every now and then it might get a common one right but in general it's bad at letter- based manipulations/ information, including even word lengths, reversal etc.).
It doesn't know what letters are because it sees BPE tokens, but if you forgive that it does something like it.
example prompt: Imagine I took all the letters in "Wikipedia" and threw them in the air so they fell on the ground randomly. What are some possible arrangements of them?
Similarly, it can almost do arithmetic but apparently forgets to carry digits. That's wrong but it's still generalization!
Interestingly enough, it immediately got "Hated for ill" (presumably because there are source texts that discuss that very anagram). But it took about 10 goes to finally a correct anagram for "Indebted sniper", though the best it could do was "pride-bent snide". I then asked it which world leader's title this might also be anagram of and it some how decided "Prime Minister" was a valid anagram of the same letters.
