Fine Tuning misconception

1st November, 2023ยท
ยท1 min read

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 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 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

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

Now, it works fairly well.

Why don't you give it a try? For starters, ask "How can one be happy". ๐Ÿ˜‰

Aravind Balla

By Aravind Balla, a Javascript Developer building things to solve problems faced by him & his friends. You should hit him up on Twitter!

Get letters from me ๐Ÿ™Œ

Get a behind-the-scenes look on the stuff I build, articles I write and podcast episodes which make you a more effective builder.

Read the archive ๐Ÿ“ฌ

One email every Tuesday. No more. Maybe less.