vault backup: 2023-11-01 22:56:20

content/memos/Fine Tuning misconception.md

@@ -0,0 +1,33 @@
+---
+title: Fine Tuning misconception
+type: Memo
+date: 2023-11-01
+description: TL;DR - Fine tuning is not long term memory.
+---
+TL;DR - Fine tuning is not long term memory.
+
+This might not come as a surprise for many but I learned this recently and it felt like a silly misconception I had.
+
+This is how I discovered...
+
+I was reading the book **The Courage to be Disliked** (see [highlights](https://aravindballa.com/bookshelf/the-courage-to-be-disliked-8783727/) from it). The format, which is like a conversation between a Philosopher and a person in their youth, felt good. It feels like we are talking to the Philosopher.
+
+However, we could not ask questions or followups.
+
+So I built [Philosopher Chat](https://philosopher-chat.vercel.app) which lets us talk to and extend conversations with the Philosopher.
+
+I took the full text of the book, converted the conversations in it to JSON format, and fine tuned `gpt-3.5-turbo` on that.
+
+Results were not as I expected.
+
+![fine tuned conversation](https://ik.imagekit.io/aravindballa/website/convo-ft.jpeg)
+
+It sure mimicked the style of the Philosopher, but not his memory.
+
+For that, I had to add embeddings, which lets me search for previously had similar conversations, and provide that as context for the AI to reply.
+
+![convo with memory](https://ik.imagekit.io/aravindballa/website/convo-memory.jpg)
+
+Now, it works fairly well.
+
+Why don't you give it a try? For starters, ask "How can one be happy". 😉