It seems that simply adding "I keep thinking until my calculation is complete" with some examples improves performance of gpt3. The more the token "thinking" is used, the higher the probability of 8.

​

https://preview.redd.it/31vcu2p6z8291.jpg?width=363&format=pjpg&auto=webp&s=7d4575baad3c6186583043d2abdeee24d8d9881c

https://preview.redd.it/wqv5w8p6z8291.jpg?width=434&format=pjpg&auto=webp&s=9686f3e517c0d5534aa41248e4780ee6be8de10c

https://preview.redd.it/pdxg2bp6z8291.jpg?width=472&format=pjpg&auto=webp&s=c24ae05f10fdef0718a50639317916aed9e6ec4b

https://preview.redd.it/cad6fur6z8291.jpg?width=534&format=pjpg&auto=webp&s=fcd6fc1b0579651882862db6d96195d5ed04f0c8

https://preview.redd.it/7e2t1hr6z8291.jpg?width=602&format=pjpg&auto=webp&s=f9ac6444885f6cb523dd25349015863e4cd21dea

https://preview.redd.it/qj5aelr6z8291.jpg?width=654&format=pjpg&auto=webp&s=9598e451d955ed0052bce7f6016b4b113a32e122

https://preview.redd.it/0awwanr6z8291.jpg?width=715&format=pjpg&auto=webp&s=fb5636fd06adc74bd6a7470a9ae80e502a7b1cf2

https://preview.redd.it/ptmvnlr6z8291.jpg?width=790&format=pjpg&auto=webp&s=a6276958813abe404b19cf2293dad8462afa3424

https://preview.redd.it/nft4n5s6z8291.jpg?width=843&format=pjpg&auto=webp&s=72765d1c4b1f88dfa6228b2b5f5a5273501e2af7

On the other hand, when "Calculate at once" is added, gpt3 answered wrong.

​

https://preview.redd.it/32neavr6z8291.jpg?width=353&format=pjpg&auto=webp&s=49eed7f79fae1f406b9f055d6763499df42b4efb

​

I think "let's think step by step" and "I keep thinking until my calculation is complete" both give gpt3 'time to think'. The difference is that the former uses a vector of the input layer, and the latter uses a vector of the hidden layer.