Hacker News Reader

Comment by hooande | Hacker News Reader

I think the example of an auto repair manual is illustrative. But because it's trained on real internet data, GPT-3 is actually quite good at "How-to".

For example, ask it a practical how to question like "how do I make a banana smoothie?" and you'll get a practical response. I haven't tested it, but I'm willing to bet you could get useful auto repair instructions for the most common models of vehicle right now without any special training. It might take some playing with the prompts, though.

If anything, GPT-3 is too tied to the real world. Like a search engine, it can quickly find the answer to almost any practical question. What we're trying to figure out is if it can do anything else.

View on HN

I tried it with a custom prompt on AI Dungeon (the paid version uses GPT-3) and I got:

Prompt:

"A recipe is a list of ingredients and instructions to make food. You are a home cook in search of a recipe for banana bread. You open a recipe book and see the following ingredients:"

GPT-3:

"Banana Bread

1/2 cup butter, softened

3 eggs, beaten

1 1/4 cups flour (measured by weight)

Preheat oven to 350 degrees F. Grease an 8-inch square pan with butter. Beat together the butter and eggs until light and fluffy. Add the flour one tablespoon at a time, beating well for each addition. Pour into prepared pan and bake for about 45 minutes or until a toothpick inserted in the center comes out clean. Let cool completely before cutting into pieces."

The auto-repair instructions would only be correct & useful in exactly the proportion that they are plagiary. The more creative it gets, the more useless or harmful the repair instructions would become. The examples we've seen demonstrate the model doesn't distinguish between fabricating descriptions of real things, and fabricating things to describe (ex.: see the example from the article, where the AI attributed a quotation that was completely made up.)

>If anything, GPT-3 is too tied to the real world. Like a search engine, it can quickly find the answer to almost any practical question. What we're trying to figure out is if it can do anything else.

I disagree, for what it is GPT could still be incredibly powerful. No person can hold all of human knowledge in their head but could GPT reasonably approach that and be able to answer almost any question that Humans have already answered? That would be incredible imo. It's Google on steroids. All the worlds information queryable in plain "english".

It's not at an all powerful Oracle that we can ask about how to perfect Fusion power or build Warp Drives but it can still do some incredible things.

Google is easy to query in plain English. ask "how many miles from London to Paris?" and you'll get a concrete, factual answer. Same as gpt-3

Reply on news.ycombinator.com