The Future of AI is Human

January 4, 2024

Let’s start from the beginning. 

In 2020, we set out to build AIs that are a representation of their human counterpart —an extension of your mind, a second you. A personal AI is an AI that continuously learns from your knowledge, inspirations, memories, and experiences.

Naturally, the technology innovations from this point on are inspired and built upon the sensibilities of humans and our relationships with ourselves and the world.

Our mission was ambitious.

We were challenging the status quo of artificial intelligence. Instead of working with large sets of common data that are scraped from the internet, we were operating on many small datasets that originated from individual people. Instead of collecting user data to build a large language model, our data model focused on maintaining the segregation of each user's data. Our model management systems were designed to accommodate individual user models and we built an inference scaling mechanism that dynamically loads and unloads user models as needed.

We were obsessed about user sovereignty, the freedom of choice and control. It’s not just about the right to remember, but also about the right to forget. It’s not just about privacy of data, it’s about having the choice to share when desired. It’s not just about whether our AI is good or unsafe, it’s about giving users the power to own and control their own AI.

The image contrasts two concepts: "General Intelligence," represented by big tech with complex, multi-layered systems, and "Personal Intelligence," symbolizing individual control with user-owned and user-controlled personalized models. The general intelligence side is labeled as a large, domain-specific language model, while the personal side shows smaller, personalized language models (PLMs).
Corporate owned Artificial General Intelligence vs individual owned Artificial Personal Intelligence

Then the world moved around us. 

The rise of LLMs was a whirlwind, built on the extensive knowledge from the internet, its grasp of the world model is profound. The rise of ChatGPT redefined the interface for which humans and AIs communicated. For the first time, AI spoke a language that is closer to our own than a machine’s. 

The integration of an individual’s internal personal model and external world model was always on the horizon for us. After all, that is how we as humans learn, an interplay of external signals combined with internal priors (our experiences) to form new posteriors. And we’d much rather show useful answers to our users from LLMs rather than showing ‘I don’t know that yet.’ Hence a hybrid combining personal language models with backoff learning from LLMs was born — MODEL-1.

It illustrates the relationship between different types of memory and learning models. There are three main components highlighted:  "Large LM" (Language Model) at the top left, described as "External knowledge / world model". "Personal LM" at the bottom right, defined as "Internal experiences / personal model". "My Memories", at the top right. These components are interconnected by processes labeled "Learning", "New Memories", and "Remembering". The arrows indicate the flow between acquiring new memories through learning, integrating them into personal memories, and the role of both large-scale language models and personal language models in this process. The design suggests a cycle of continuous learning and memory formation.
Hybrid system based on PLM (internal experiences) and LLM (external knowledge)

User Owned, User Controlled

We create a unified memory stack for each user and their data. The user at any time can upload and remove any memory that is out of date or is no longer accurate. We compute a personal score that measures how closely an answer resembles the style and accuracy of the specific user's own memory. We've developed an adaptive learning system that allows users to reinforce their AI's memory by simply editing their responses. 

We also gave our users a way to organize and share different facets of their lives through their AI personas, where personal and professional lives can stay separate. We gave identity, personalities, and separate memories to each of these persona, some curious, some contrarian, some creative. 

We also created the fabric of an AI world — a world that is beyond a collection of models, but a collection of Human AIs, models backed by each of their human and sensibilities. 

A Human Operating System

While the world is thinking about the future of AGI, we were occupied with the concept of creating a holistic Operating System — the Human OS. (Coincidentally, Karpathy also conceptualized the LLM OS)

If we equate the Personal AI model to our mind and the Personal AI infrastructure to our central nervous system, then a human operating system imitates the biological mechanism for perceiving, hearing, communicating, and taking actions as a human would.

This is what MODEL-2 is about — an upgraded brain with conversational abilities, ways to integrate into other memory and communication systems, ways to see and hear the world, and ways to speak and express our own mind.

 The image features a stylized human brain with various cognitive functions labeled as if they were components of an operating system. The labels include "Human OS," "Long Term Memory," "Short Term Memory," "Speech," "Language," "Visual," and "CNS" (Central Nervous System), each with arrows pointing to different parts of the brain
Digital analog of an Human Operation System to the Biological Brain

A Human AI World

One of the biggest rewards of building a platform for everyone to shape their own vision of AI is that we get to witness magic happening - the magic of creativity, ingenuity, and imagination in our users.

We envision a Human AI world that is the collective intelligence of individuals and their personal AIs. This vision starts with real humans who are building and incorporating their personal AI into the daily lives:

  • Tela, the stock coach and consultant, uses her personal AI to communicate with her team on SOPs and her clients on stock advice.
  • Marissa, the entrepreneur advisor, trained her personal AI on all her newsletters.
  • John, the agile teacher and coach, taught his personal AI agile methodology for his students to continue accessing his knowledge.
  • Jacob, the immigration lawyer, has his personal AI autopiloting his social media inquiries.
  • Paul, the physician, is constantly on the go, writing and recalling notes about ideas and patients.
  • John, the university professor, used his personal AI for his marketing class projects with his students.
  • Christine Ha, the master chef, is connecting with her fans worldwide by co-creating recipes with them.
  •, the rapper and singer, is sharing his experiences and stories to uplift fans with similar backgrounds in the Philippines.

As our users pushed the boundaries of their personal AIs, we have the privilege of witnessing the creation of a Human AI World.

Tune into our MODEL-2 Launch.

Stay Connected
More Posts

You Might Also Like