@cjzafir: I fine-tuned a model on my bookmarks. I tried everything to make the model understand what I like, what content I love …
Summary
The author fine-tuned a model on their bookmarks to create a personalized content recommendation algorithm that scans social media and research papers daily, adapting to user feedback.
View Cached Full Text
Cached at: 09/12/26, 06:59 PM
I fine-tuned a model on my bookmarks.
I tried everything to make the model understand what I like, what content I love to consume, and what ideas make me excited.
No skills, instructions, or anything else worked!
Then I asked Fable 5.1 to build a tiny algorithm.
This algorithm scans my X, Insta, and YT and sees what I saved. And I had 500+ items there (so a decent dataset on what I like).
Then Fable 5.1 + GPT 6 Astra worked together to understand each piece of content (tweets, long posts, articles, shorts, and for long videos, they used Gemini 3.8 Flash to get all insights with timestamps).
Then they turned all the pieces of interest into a neural system (a virtual profile of who I am, based on my interests). This was important because now this algorithm can predict what content I will like for sure.
Now this algorithm runs a cron job every morning, scans X (with X Grok API), Insta, YT, and Reddit, and updates the feed for me.
You can say it’s like the “For You” page we have here on X, but actually better and personalized to me.
Fun part: I can modify my algorithm by providing it feedback (“Hey, I really like this post, find me more posts on this concept”), and the algorithm scores that interest higher and gives it priority in future scans.
I learned so much with this “feed,” and I love research papers, so now it adds 3 new research papers in the form of a long-form post (in my style), and if I like it, I can ask it more questions, open the paper myself, or ask it to make diagrams for better explanations.
In a nutshell, it’s like if we could tell Grok to modify the algorithm for us and make it truly personalized.
But I feel lonely on my feed lol. But that’s not bad because I feel like X is just turning into a breaking news hub, where there’s just so much clutter, noise, and unnecessary stuff that’s grabbing attention.
I’ll share more on the “feed” if you guys like it.
Similar Articles
@cjzafir: Models that I'm using daily: > Codex 5.5 high (fast) > Deepseek v4 pro via API > Kimi 2.6 via API Models that I am fine…
User shares a personal list of AI models they use daily (Codex 5.5, Deepseek v4 pro, Kimi 2.6) and for fine-tuning (Qwen 3.5 variants, Gemma4 E4B, GPT-oss 20B), aiming to fine-tune Small Language Models into Expert Language Models.
@cjzafir: Before Claude Fable 5 got banned, I turned all my fine-tuning research and experiments into a product: http://Finetuner…
Developer @cjzafir announces Finetuner.dev, a CLI tool that uses orchestrator models like Codex 5.5 and Chinese models to generate high-quality, handcrafted datasets for fine-tuning small language models (1B-30B), claiming 10x lower costs and 5x better quality.
@omarsar0: Hacker News → LLM Artifact I built the most personalized HN feed. It only tracks topics I do research around based on m…
A researcher built a personalized Hacker News feed powered by LLMs, memory, and proactive agents that tracks only topics relevant to their research interests, eliminating the need for bookmarks.
@oneill_c: 1/ We fine-tune a lot of customer models, so we decided to systematically try and figure out some best practices for fi…
The thread shares systematic experimental findings on fine-tuning best practices, varying one SFT lever at a time across dense and MoE models up to 235B on four real-world customer datasets with custom evals to eliminate confounders.
@no_stp_on_snek: what actually surprised me fine-tuning a small open model. note im failry new in this area so some of this may seem obv…
A developer shares surprising lessons from fine-tuning a small open model, including that base models often already max out on intended improvements, the real weakness is behavior (caving), and fine-tuning requires careful measurement and balancing.