The offline in-person conversations can lead to highly relevant insights that are rarely documented. LLMs (Large Language Models) can understand and generate natural languages and thus, are the ideal agent to tackle the note-taking problem. The question is: how do we develo a proper interaction?
Solution
We have not yet found a final answer. However, we believe in the hard way—developing local LLM-empowered AI hardware to generate structured, personalized notes.
Major milestones
These are major appearance and interactive prototype concepts in April 2024.
The original idea came from the daily interactions while working in the first semester of Master of Design program at UC Berkeley:
Within an interdisciplinary cohort of students, we constantly discuss over topics of mutual interest - these casual, unplanned encounters sparkle fleeting insights that can inspire the next project. Yet, we don't have a good way to capture them every time.
From without
The memex is an envisioned device meant to enhance human memory by allowing users to store and retrieve documents linked by associations, resembling the way the human mind works. Inspired by such a foreboding concept from the 1945 paper, As We May Think by Vannevar Bush, we wanted to create a device for us to record daily conversations and consult whenever needed.
Design with no precedents
The paradigm shift in AI-enabled hardware introduces challenging hardware design principles. As we prototype and communicate with potential users, we are also verifying our conjectures:
AI with warmth
The concept of "AI" feels distant to most people. An important step in embracing AI applications is to introduce emotion design elements.
The LED matrix can show customized facial expressions; The simple UI can also show other important states
Co-pilot over auto-pilot
AI's capabilities are far from perfect for many tasks; even when it does, it takes time for people to build trust. In this, user intervention in any task is needed.
This concept introduces the "rewind" UI to redo the most recent notes
The latest concept emphasizes on co-notetaking (the annotation button) in addition
An ongoing journey
The validation process will be an ongoing journey with new theories to prove or disprove. These may be as simple as a design change and as hard as tweaking LLM models. More importantly, it will be a conversational one as we talk with people to get feedback.
If you happen to read this, don't hesitate to connect to express your interest. We would love any input to inspire us. Thank you.
Takeaway
The hard way to develop an AI application?
LLM models excel at specialized tasks because they can process and analyze vast amounts of data quickly and accurately. BLLMs can develop deep expertise in particular domains by training on specific dataset.
On the other hand, AI security is a paramount concern. Proper industrial design along with locally processed data can provide much stronger confidence in data integrity than any pure software solution.
Local LLM and hardware are the hard ways to develop an AI application. However, we believe this to be the easy and sustainable approach in the long run.
