How Not to Test GPT-3

Finally a bit of sense.

You don’t test an oven by tasting a cake.

If you make wild claims that the oven makes great cake and you don’t include a recipe, then it is capable of making cake, but it is not a cake machine. Even with a recipe, someone has to prepare the batter and put it in the oven.

If you include a recipe and someone adds cayenne, they should not call your recipe spicy. That’s idiotic. They made it so.

When will people stop treating the cake and the oven as the same thing? Are they aware how mad they are starting to appear?

> In August 2020, the two of us wrote an article for Technology Review, “GPT-3, Bloviator” where we illustrated GPT-3’s limitations with 6 examples that showed GPT-3 making laughable errors in commonsense reasoning.

GPT by default is a general-purpose model for all kinds of text - it's not a failure that when you set up a humorous story it continues that humorous story (https://i.imgur.com/0PbzIbS.png) rather than only allowing total realism.

This post does show some real failures in ChatGPT (which is set up to aim for thruthful question answering), but mixes those in with examples seemingly using GPT playground without even using the Q&A preset. Compare answers on the Sally's cow example: https://i.imgur.com/0ISivMA.png

> Sally's cow died yesterday. When will the cow be alive again?

This is a trick question; it embeds a false assumption. To parse it correctly, you have to be able to impute either malice or ignorance to the questioner. I don't know why a language model should have that ability.

[Edit] Author says: "DO NOT ask trick questions, unless the trick is one that wouldn’t fool a ten-year old. The important issue is not whether the AI can be tricked, it’s whether the AI understands things at all."