LLM Role-play

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA256 [My experience with Large Language Models](https://www.homosapient.com/my-experience-with-large-language-models/) is full of exploration of ideas. I had the model perform a lot of role-playing. Sometimes I made it act as a _Marketing Expert_, sometimes a _Pilot_, and sometimes a _Curious Kid_. I noticed that the meanings of these roles differed in my head due to my biases, schemas, and memories. Curious, I wondered if the model itself could explain what it perceived a Curious Kid to be. Initially, it replied predictably with, "As a Large Language Model, I don't have perceptions about..." So I clarified by saying, "Describe the characteristics of a Curious Kid," and it responded: 1. **Asks lots of questions**: They frequently ask “why,” “how,” and “what if,” often about things adults might overlook. 2. **Eager to explore**: They enjoy investigating their surroundings, experimenting with objects, or trying new experiences. 3. **Enjoys problem-solving**: Curious kids like puzzles, challenges, and figuring things out on their own. 4. **Not easily satisfied with simple answers**: They often dig deeper, asking follow-up questions to understand more fully. 5. **Observant and attentive**: They notice small details and may spot things others miss. 6. **Creative and imaginative**: Curiosity often fuels storytelling, imaginative play, and unique ideas. 7. **Willing to take risks**: They aren’t afraid to try something new or make mistakes if it leads to discovery. 8. **Self-motivated**: They seek information or explore topics without being prompted. 9. **Persistent**: If something piques their interest, they keep at it even when it’s challenging. Wow, that was more comprehensive than my initial thought, which had only been "asks a lot of questions and isn't satisfied with simple answers." Reflecting back, my initial concept was indeed just a vague feeling. That's when I got the idea: I wanted ChatGPT to "role-play a curious kid who could instantly turn into an adult whenever I wanted (by issuing the command **/growup**). Then, that adult could turn into Socrates (**/lastwords**), and finally back to normal (**/maculatum**)." Thanks to OpenAI’s "Custom Instructions" feature, I could store these commands in the background, accessible in any chat. The resulting outputs genuinely amazed me. They noticeably wiggled and jiggled up my worldview. 💭 Worldview? **My** worldview? I thought to myself. This intrigued me because I realized that by explicitly instructing the model to behave a certain way, it was responding according to my expectations. But what about its default behavior? Perhaps I wanted to change its default because it wasn't satisfying enough for me. Let me slow down. I started dissecting things. Maybe generating tokens (words) in the same inference (run) limited the model's ability to fully embody the roles I wanted. It felt like having multiple personas (kid, adult, Socrates) combined into a single equation, something like: find x where 2 (kid) + 3 (adult) + x (Socrates) = 10, influencing the final outcome. Around that time, OpenAI released their API, and I was among the first few to gain access. Despite the limited documentation available, I wrote some code based on OpenAI cookbooks, making three distinct prompts in three separate AI calls: the curious kid, the adult, Socrates, and added a listener named John, who listens to all of them and wonders what's real. I called this setup the "Truth Triad." Coincidentally, it was around the same time a community member demonstrated an idea called Baby AGI. The outputs from the "Truth Triad" surprised me in ways I never expected. Now, reflecting on this, I believe I should recreate some of these outputs so you can experience the surprise firsthand. I will keep that for the next post! (To give you an incentive to come back?) -----BEGIN PGP SIGNATURE----- iQIzBAEBCAAdFiEEsiXb6WyMQEcAYB7o26UiLapKcWkFAmhZDoAACgkQ26UiLapK cWnuzhAAiPoavOjy7JimuoVqjeAkPqvY7XvPuxCJLj4gaD6euJwKJyztVeMZFrKr 1Lv5B4FSyEAE7VooC77cBhHMjJddeMyrr3iSWGXl78JARsoaaoyxWSs1u6prALJX 5uThTXjX2K3bnwuF/uXG3NMzHu8xpvDjhzbCxHAU6mzVMsD7OgWm6KqCTNKcJfPP sNcTlKh567E5F6RUMBU/lxqVQ3OYHbTJb2KVTN13YqNMC6gZSfJTBIZU05Jx6zNB BlVRAyl8w/47RvZhvXPamTxdLAgejAwDcTEdEQ1swJAp7F1NPpt8vpLMrT8/l7Cz dAZgwmFM2OMhm7YwllJ6UUdcte/ZHqF4hWZ9QFIdvj0FAIqrV1EvXSpAaGV1dZrO qH/9bZ09bpDVTpKBnW+zPlVRd2A+IZU0ZQrHUxzOCOKrkPcM3PWjghkWH6V8dHbR zYG2aaqtGuiBv6fDqB4+1/2KLNJruMkmvkMG7IALjc2W5tT62HYSn16PxUdAYqLV CzV7d0GhzmT/YhHGKzfOWd8aeEwZ/pJD76fjC4U7nJaO7udFh7hXkuDP2+S91HM/ bYKV1WP9mov4deqQAG3Hu6joAC1WQGsKgvFo3Nx3dYvP4HseBSWh+Xur0RsHV7fI 7WjWeHfri+tJtlgybneFrCSJNPhepiIaU4Cw6L0aISPMqB3PW1I= -----END PGP SIGNATURE-----