GPT-3: Creative Potential of NLP

It was last year in February, as OpenAI published results on their training of unsupervised language model GPT-2 . Trained in 40Gb texts (8 Mio websites) and was able to predict words in proximity. GPT-2, a transformer-based language applied to self-attention, allowed us to generated very convincing and coherent texts. The quality was that good, so the main model with 1.5 billion parameters wasn’t initially publicly accessible, to prevent uncontrolled fake news. Luckily, the complete model was later published and could be even used withColab Notebooks.

This year OpenAI strikes back with new language model GPT-3. With 175 billion parameters.

Unnecessary spoiler: it’s incredibly good.

There are already some profound articles on TDS examining features and paper of GPT-3:

But how does it look like in action?

OpenAI is building an API, currently accessible via waiting list:

Fortunately, I could get access and experiment with GPT-3 directly. Here are some of my initial outcomes.

Interface, Settings, Presets.

The AI Playground interface looks simple, but it bears the power within. For the first, here is a setting dialog, which lets you configure text length, temperature (from low/boring to standard to chaotic/creative), and other features.

You also can define where the generated text has to start and to stop, these are some of the control functions that have a direct impact on textual results.

The simple interface provides also some GPT-3 presets . The amazing thing about transformer-driven GPT-models is among others the ability to recognize a specific style, text character, or structure. In case you begin with lists, GPT-3 continues generating lists. In case your prompt has a Q&A structure, it will be kept coherently. If you ask for a poem, it writes a poem.

You can do your own presets, or use the existing, which are:

Chat.

A typical setting for a chatbot. You ask - AI answers. It’s possible to change the “characters” or setting also. As you can see, the chat situation was accomplished perfectly (even if my, Human’s, third question was kind of unfair).

To demonstrate the contextual impact, let’s change the AI character from “helpful” and “very friendly” to “brutal, stupid and very unfriendly”. You will see how the whole dialogue will be influenced:

I think, we re-invented Marvin.

Q&A

This preset consists of a clear dual structure: Question and Answer. You need some training before it starts to answer the question (and get the rules), but then it works perfectly. I asked some random questions from various areas and here you go:

I’d say, perfect!

Parsing unstructured data

This one is fascinating and shows a good comprehension of the unstructured text — extracting structured data from the full text.

Summarizing for a 2nd grader

This preset shows another level of comprehension — including rephrasing of difficult concepts and sentences in clear words.

I tried Wittgenstein:

The simple proverb can be paraphrased convincingly:

Or look at this pretty well and clear transition of Sigmund Freud’s time distancing concept:

As you see, compression of text and its coherent “translation” is one of the strengths of GPT-3.

What about languages?

GPT-2 was already a great language model when it was about English. You could generate amazing texts, especially with 1.5 billion parameters. I used GPT-2 for a screenplay of this short movie — and its absurdity could be rather understood as a good tradition of David Lynch and Beckett:

The dialogues were logical, even if spontaneous. But it was regarding English. If you’ve tried with inputs in other languages, you would face the barrier of understanding. GPT-2 tried to imitate languages, but you needed to fine-tune it on text corpus in a specific language to get good results.

GPT-3 is different.

It’s processing in other languages is phenomenal.

I tried German, Russian, and Japanese.

German.

It was rather my daughter, who tried to let GPT-3 write a fairy tale. She began with “ Eine Katze mit Flügeln ging im Park spazieren ” (“ A cat with wings took a walk in a park ”).

The emerged story was astonishingly well written. With irony, vivid characters, and some leitmotifs. This is not just a collection of topoi or connected sentences. This is… a story!

Russian.

I trained once GPT-2 on Pushkin’s poetry and have got some interesting neologisms, but it was a grammar mess. Here I input some lines of Pushkin’s poem — and the result I’ve got was… interesting. It hadn’t rhymes, but stylistically intense power. It was not Pushkin style, though. But almost without any mistakes or weird grammar. And… it works as poetry (especially if you are ready to interpret it).

Japanese.

This was something special. I entered just a random sentence:

今日は楽しい一日になりますように！と言いました。// Today was funny and entertaining day, I said.

And the result was a small story about prayer, happiness, wisdom, and financial investment. In well written Japanese (neutral politeness form, like the input).

It does mean: GPT-3 is ready for multilingual text processing.

Various experiments (and alerting signals).

ShakespAIre and writing poems

My first try was, of course, to write a Shakespearean sonnet. So the prompt was just:

here is a poem by Shakespeare

The result was this:

Perfect iambic verse, great style, nice rhymes… If not one thing:

The first two lines are actually from Alexander Pope, The Rape of the Lock. And here we have a reason to be cautious: GPT-3 produces unique and unrepeatable texts, but it can reuse the whole quotes of existing texts it was trained on.

Re-examination of results is inevitable if you want to guarantee a singularity of a text.

I wonder, if there are some possibilities for “Projection” like StyleGAN2 feature , just in opposite to StyleGAN2 (where it compares the image with latent space), in GPT-3 it would compare with the dataset it was trained on? To prevent accidental plagiarism.

But the thing is: GPT-3 can write poems on demand, in particular styles.

Here is another example:

Essays

As I still hadn’t accessed, I asked a friend to let GPT-3 write an essay on Kurt Schwitters, a German artist, and Dadaist:

The outcome is: GPT-3 has already a rich knowledge, which can be recollected. It is not always reliable (you have to fine-tune it to have a perfect meaning match), but it’s still very close to the discourse.

Coding with GPT-3

Another mindblowing possibility is using GPT-3 is quite different cases than just text generation:

You can get support by CSS:

And calling it General Intelligence is already a thing:

Summary.

We are still at the beginning, but the experiments with GPT-3 made by the AI community show its power, potential, and impact. We just have to use it with reason and good intention. But that’s the human factor. Which is not always the best one.

For more wonderful text experiments I highly recommend you to read Gwern:

Let the journey continue!