[longtom]

Could someone with GPT-3 beta access try whether it can better solve 3 digit addition when it is allowed/encouraged/forced to make intermediate results explicit? E.g. instead of

21 + 110 = 130

150 + 12 =

condition it on

21 + 110 = 100 + 10 + 20 + 1 = 100 + 30 + 1 = 131

150 + 12 =

or similar. Given that humans make these intermediate steps in their heads GPT may perform better when it is encouraged to do them as well.

This may in fact apply to all sorts of reasoning, but in many cases it may be difficult to make these steps explicit in text form. Humans seem to mainly use some prediction layer or scratchpad which also contains the inner monologue but also motor primitives, smells, images, everything. Humans can decide to think a bit longer before producing an output, which appears to require an RNN.

[maliker]

I took a shot at this, didn't have much luck though:

PROMPT ======= Input: 21 + 110 Output: 100 + 20 + 10 + 1 = 100 + 30 + 1 = 130 + 1 = 131

Input: 89+78 Output: 80 + 70 + 9 + 8 = 150 + 9 + 8 = 150 + 17 = 150 + 10 + 7 = 160 + 7 = 167

OUTPUT ====== Input: 37 + 112 Output: 30 + 100 + 10 + 2 = 110 + 1 = 111

Input: 91+11 Output: 100 + 90 + 1 = 190 + 1 = 191