You're predictable. That's a good thing: Can personal LLMs learn your judgment? (handsdiff.substack.com)

A new study has revealed promising advancements in the development of personal large language models (LLMs) that can predict user writing behaviors, thereby significantly enhancing workflow efficiency. The model, Astra xhigh, demonstrated a 17.1% semantic accuracy for next-write predictions at a cost of just $0.3 per call, compared to a mere 0.9% accuracy from the Qwen 3.6-35B-A3B model. This research highlights the potential for LLMs to learn user judgment by leveraging structured usage data from computers, ultimately allowing for smarter, context-aware AI agents that can adapt to individual user preferences and goals. The significance of this work lies in its ability to alleviate the cognitive load on knowledge workers, who currently spend excessive time formulating prompts for AI tools. By creating a dynamic pipeline to capture and clean user data, the study lays the groundwork for continuous learning and optimization of these models. This involves fine-tuning based on user-specific workflows, with the ambition of developing an integrated agent that not only understands user intentions but can generate high-quality outputs with minimal input from the user. The vision is to transition toward a future where personalized AI systems can process inputs across various modalities—text, images, audio—while maintaining strict privacy protocols, ultimately augmenting human creativity and decision-making in professional environments.