~~ChatGPT is terrible at poetry~~ AI's Potential in Generating Poetry

ChatGPT is terrible at poetry AI’s Potential in Generating Poetry #

Artificial Intelligence (AI) is an ever-evolving field, continually unlocking new use cases through breakthroughs in text and image generation. Amidst this innovation, there’s a persisting narrative arguing that it’s impossible for AI to generate poetry effectively. This viewpoint often arises from various sources, including critical articles and self-awareness from AI models like ChatGPT-4:

Vice: ChatGPT Is Pretty Bad At Poetry, According To Poets

The Atlantic:What Poets Know That ChatGPT Doesn’t

prompt: is poetry a good use case for GPT-4?

response: […] while GPT-4 can generate poems, it’s important to remember that it doesn’t truly understand the emotions, experiences, or the deeper meanings behind the words it’s generating. It’s simulating the process of writing poetry based on patterns it has learned, not creating from a place of human understanding or emotional depth. Therefore, while it can be a useful tool, it doesn’t replace the creativity and emotional intelligence of a human poet.

AI’s Potential in Generating Creative Short-Form Text Content #

I want to challenge this opinion and see if it is possible to get ChatGPT to create poetry. Now, there are some strong arguments against it such as those in the articles, and I understand what the challenge is for anyone who is trying to get ChatGPT to create truly unique and interesting text. As someone who has used the OpenAI Davinci model (predecessor to GPT), GPT3.5 and GPT4, the main limitation is obvious: ChatGPT creates bland, generic, and repetitive texts. The first and most obvious challenge is somewhat easily surmountable - make the text content less repetitive.

However, even if you can make the content itself somewhat diverse, if you ask ChatGPT to make a poem, or 100 or 500+ poems as I have, you can tell from just glancing at the overall formatting and the tone of the content that it was AI written. There are plenty of examples of this such as linked above.

Obviously you can’t make ChatGPT truly understand the human experiences that inspire poets. This is the core argument against ChatGPT being a good use case for poetry by itself. So maybe 100% unassisted poetry is not attainable. But perhaps there is a way to use layers of automation to create a consistently interesting output with very minimal human assistance.

Now, before I proceed to technical details about how I am brainstorming to try to tackle this challenge, I want to take a step back and think about the entire idea of what poetry is and how text content is consumed and distributed in the internet era.

I would argue that while there is still a place for traditional poetry and prose, as the relative cost of generating and distributing text content has been reduced drastically over time, the number of applicable use cases has exploded.

Simply put, if we consider the text of all memes, posts, tweets, blogs – and we can consider at least some amount of this content – prose or poetry. The proliferation of text communication and the ephemeral nature of it is clear. I see a future where with minimal input, you can easily generate highly curated text content for a specific person or set of people in a specific physical and social context. I see this work of trying to generate highly curated poetry as just one small drop of water in a massive tsunami of social and technological change that AI is actively creating.

I do not think of this “poetry” goal as necessarily poetry as we think of it today, but just snippets of stories, that may closer to prose or memes.

What is the potential for unique creative content being created by ChatGPT? Advancements in GPT and text generation #

Not all blogs about ChatGPT for poetry are negative. We can see that some people are using poetry and other artistic text to circumvent ChatGPT’s reluctancy to provide an opinion about any topic:

Harvard Law Blog: chatgpt-poems-and-secrets

One of the core use cases that ChatGPT seems to be really well made for is writing blogs, articles, and assisting with longer-form content such as books. Here is an example of a blog where the author shows how to use ChatGPT to help write books.

ChatGPT Fiction Demo

The author concludes:

Could I have written this myself in the same amount of time? Almost certainly. At this point, I will say that ChatGPT is workable, but it’s no “Photoshop for Writers.” It often forgets or ignores commands (like “write in first person present tense”) so I wasted a fair bit of time cajoling it into executing the commands I wanted. Based on the incredible advancements that other AI systems have gone through (i.e., Midjourney) I suspect we’re six months to a year away from having many of these problems ironed out.

It is interesting to note they used ChatGPT3 and they explicitly called out the limitations of 4,000 tokens. GPT-4 API allows you to use much more tokens. It is currently in limited beta but you can see in the docs it allows 8,192 tokens. The model gpt-4-32k-0314 allows 32,768 tokens. That is a factor of 8 over the GPT3 4,000 tokens.

Additionally the use of the chat completions endpoint allows you to maintain a context more reliably than previous models such as DaVinci.

It’s crucial to note the substantial enhancements in GPT-4 compared to its predecessors, such as a significant increase in token capacity and more reliable context maintenance. These advancements not only make AI tools more efficient but also open doors for more creative use cases each day.

OK enough rambling ideas. For my next Article I will be sharing actual brainstorming and testing details around making poetry with the openAI GPT3.5-turbo API.