AI Lab · by Léa Bernard · 2026-10-14

I fine tuned a small 7B model on my own notes and it's genuinely more useful to me than GPT4 for personal stuff

Spent a weekend fine tuning a small open weights model on five years of personal journal entries and project notes. For anything involving my own context, history, and writing style it absolutely crushes the big general models, even though it's objectively 'dumber'. Makes me think the future of personal AI assistants is less about raw intelligence and more about deep personalization. Anyone else experimenting with this?

0 sparks · 9 comments

Comments

Rafael Almeida · 0 likes

this is basically the thesis behind a few of the local-llm projects floating around, personalization beats raw IQ for a lot of daily use

Valeria López · 0 likes

what base model and framework did you use? curious about the compute cost for the fine tune

Urban Noodle · 0 likes

would also love to know the GPU hours / cost breakdown

Leila Karimi · 0 likes

did something similar with my email archive, kinda creepy how well it picked up my writing tics ngl

Bilal Raza · 0 likes

privacy angle is huge here too, keeping it local means your most personal data never leaves your machine

Miguel Santos · 0 likes

RAG over your notes with a bigger model might get you 80% of this benefit without the fine tuning hassle, have you compared?

Aoi Ito · 0 likes

i tried rag first honestly and fine tuning genuinely felt different, more like it 'absorbed' my voice vs just retrieving facts

Avery Johnson · 0 likes

this is the kind of project that makes me want to finally learn how lora fine tuning actually works lol

Noah Miller · 0 likes

small personalized models running locally on phones in a couple years is gonna be huge imo, this is a preview of that