• Read Language models can only write ransom notes by Allison Parrish and review the The Foundation Model Transparency Index.
• Experiment with prompting a language model in some way other than a provided interface (e.g., ChatGPT) and document the results in a blog post. You can use any of the code examples above and/or try a variety of LLMs locally with ollama. Reflect on one or more of the following questions:
  • How does the concept of LLMs as “ransom notes” influence your perception of using these models creatively?
  • How does the hidden origin of the text that LLMs generate affect your sense of authorship or originality in your creative work?
  • How does the metaphor of "collage" used to describe LLMs align with or differ from your creative process?
  • How would you compare working with an LLM to other forms of text generation, such as using a Markov chain?

Experimenting with llama 3.2 locally

I tried chatting with llama 3.2 locally using ollama on my PowerShell and started with a simple greeting, the LLM replies rapidly with little delay unlike my experience with ChatGPT 3 or other LLMs that runs on servers.

Then I asked it a tricky, trivial history question and unfortunately, llama 3.2 got it wrong. Nelson is associated with pickles because 1. he was pickled in brandy after death 2. the brandy with Nelson the pickle inside was carried back to England by a ship called HMS Pickle.

I looked up online on HMS Vanguard and it was one of the battleships that Nelson commanded during the Napoleonic Wars, and I guess llama 3.2 scavenged for any information related to Nelson and came up with a completely wrong answer.

Although, I don’t really know how llama 3.2 got the information that Nelson enjoys pickled herring, I tried googling these two keywords and couldn’t find any source, so llama 3.2 probably just made it up in response to my prompt.

Reflection Questions

1. How does the concept of LLMs as “ransom notes” influence your perception of using these models creatively?

“Ransom notes” aims to reduce the traces of the creator using uniform text, colors, and composition, and when we’re using ChatGPT to generate a piece of text we more or less give up our original flavor in writing text. Thus, when we’re using any generative LLMs we should be aware that the language it uses aims to create uniformity instead of individuality of the person inputting the prompt.

1. How does the hidden origin of the text that LLMs generate affect your sense of authorship or originality in your creative work?

In traditional collage or music production if I ever used materials that isn’t mine like samples, photographs, models, original character designs I would be able to correctly attribute the source and respect the original author. Although LLMs also scrap information from multiple sources, it doesn’t attribute the source and I should be careful not to let ChatGPT take credit for the words it spits out and check for the original author.

1. How does the metaphor of "collage" used to describe LLMs align with or differ from your creative process?

I personally enjoy creating collage that uses text, abstract visuals, symbols, motion graphics, dialogues, sound effects, and clips. LLMs could sometimes generate completely random, but somehow structured poems that satisfies my needs. I would give it specific prompt to force it to generate fractured snippets of text that I experiment with recontextualize into my own animation.