Back to Blog
·Eva Popílková·2 min read·Archive 2020

Experiments with Different Versions of GPT-3 and Language

In my previous post, I raved about GPT-3. Now, I would like to share a few of my recent experiments with different versions of GPT-3 and language: Recently…

Experiments with Different Versions of GPT-3 and Language

In my previous post, I raved about GPT-3. Now, I would like to share a few of my recent experiments with different versions of GPT-3 and language:

Recently, I tested the linguistic intelligence of various versions of GPT-3. I tasked it with coming up with an alliteration for the sentence “I am very hungry.” = “Jsem velmi hladový” and in an instant, it produced a remarkable result: “I am veritably voracious.” = “Jsem skutečně nenasytný.” The older version merely wrote, “I am in need of an ice cream cone.” = “Potřebuji zmrzlinový kornout.” For comparison, my human colleagues at work were practically unable to come up with anything :)

In another experiment, I wrote a simple analogy generator with reasoning and asked it, “What is dancing in the rain like?” The latest version promptly responded: “Dancing in the rain is like enjoying a cold drink, because both are ways to cope with the heat.” Clever :) Another version wrote, “Dancing in the rain is like surfing, because both offer an exhilarating way to enjoy great nature.” Alternatively, it said, “Dancing in the rain is like playing in the mud, because both are fun.” The old version, however, only produced something dull: “Dancing in the rain is like dancing in a circle, because both involve moving from one place to another at high speed.” My human colleagues offered solutions that were closer to the last convoluted variant.

I always had trouble with rhyming in the old version. For the word “dog,” it suggested futile rhymes like: doghouse, dog, mongrel, mule, mutter, and for the word “Johnny,” it provided the unimpressive jay, jam, joe, joey, jane.

The new version responds to the word dog with: log, fog, dog, jog, nag, bog, and for “Johnny,” it offers the decent ninny, money, Sonny, Winnie.

From my perspective, I took away two things:

  1. There are significant differences between the versions of GPT-3!

  2. In a number of linguistic tasks, GPT-3 outperforms untrained colleagues, not only in terms of speed but also quality ?

Source:

https://cs.wikipedia.org/wiki/Aliterace

Původní zdroj: wordpress

Související články