That's a great article, and the logic also makes perfect sense. However, I think that the learning capabilities of LLMs are a bit underestimated - as long as the human experience is present on the internet in the form of words, it can learn from it in the next iteration.

For instance, I tried the three examples from the start of your post. Neither of the example gave the wrong answer, but rather a more nuanced one. Like for the question "If you need to cover your hair for work in a fast-food restaurant, which would work better, a paper sandwich wrapper or a hamburger bun?", this is the answer now from GPT-3:

"When covering your hair for work in a fast-food restaurant, neither a paper sandwich wrapper nor a hamburger bun is an appropriate option.

A paper sandwich wrapper is not designed to be worn as a hair covering and may not be large enough to fully cover and contain all of your hair. It may also be difficult to secure it in place and could easily fall off, potentially contaminating food or causing a safety hazard.

Similarly, a hamburger bun is not a suitable hair covering as it's designed for consumption and not for use as a protective barrier. Using food products as hair coverings can also pose a risk of contamination and is not hygienic.

Instead, it's recommended to wear a hairnet or a hat that is designed for this specific purpose. A hairnet is a lightweight, mesh-like covering that fully encloses the hair and is designed to prevent loose hair from falling into food. A hat is another option that can effectively cover the hair and prevent loose hairs from contaminating the food. Both of these options are available in a variety of styles and colors and are specifically designed to be worn in food service settings to ensure hygiene and safety."

So it does seem like the model got iteratively better from the large-scale feedback, in case it gave crappy answers initially.

This means that when the model learns once (through sensory inputs or not), it is able to simulate an answer that is highly probable or even exactly like in the real world.