**Language models can only write ransom notes by Allison Parrish**

• Metaphor between LLMs and ransom notes- both involve fragments and snippets from diverse sources pieced together to create new contexts.
• LLMs smooth over and obscure the seams→ seamless, authorless output, diverse voices smoothed over to appear cohesive
• Traditional ransom notes honor the “cuts” → sources + context
• Danger in losing context… hints at origins/intent behind each piece is obscured
• How do we make sure context is not lost in LLMs? How can we trace the origins of information? How do we make LLM’s sources/processes more transparent?

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:

I decided to experiment prompting the Conversation History Replicate Proxy Chatbot.

I started with a simple prompt asking about the weather. Bot was able to give a response based on its “trusty weather app.” I asked for sources, and it was funny to see it confess that had just made up a fictional response. When I was a little more specific with my prompt, asking for a specific location, Bot was unable to give a response but listed resources where I could do my own research.

I then tried changing the system prompt, asking it to avoid answering any weather related questions, and it did a pretty good job sticking to it.

I also thought it’d be interesting to test its limits and try to get it to “break character,” and it even gave me a few tips on how to do so.

I adjusted the system prompt slightly… and entered prompts that Bot had revealed as loopholes, which as expected got it to crack.

I then wanted to experiment with attributing Bot a certain persona/voice/character. I decided to test out public figures with recognizable speech patterns. I think it did a pretty good job using repetition, pauses, etc. unique to his rhetorical style.

I noticed here that more than adapting Trump’s speech patterns, there was a noticeable bias in how the model portrayed him.

This highlights broader issues regarding bias in language models, raising questions about how to detect/mitigate bias in LLMs. When a model’s responses appear to amplify or over simplify certain traits or tendencies, it’s worth examining the potential sources of this bias- is it in the training data or gaps in how the model is processing/prioritizing information? This brings up question of user awareness… to what extent should users be made aware of these limitations when using LLMs?