Why We've Built a Me.bot,
Not "Her "
Not "Her "
Tao, Founder of Me.bot
Working at Meta made me take a hard look at how humans and AI should coexist.
During my time at Facebook from 2017 to 2019, I led a team building content recommendation algorithms for News Feed. The research was intellectually stimulating, but having worked on delivering more content to users, I found myself questioning the impact of the technology: Does AI-recommended content actually enhance people's happiness?
Software was invented to empower humans. Yet in their reckless pursuit of profits, tech companies have lost sight of that purpose. Social media content is propagating unattainable beauty standards that affect youths’ self-image, short videos are hurting attention spans, and content recommendation algorithms are driving political polarization.
I grew increasingly troubled by the ways technology was harming the very people it was meant to serve. How do we develop technology that prioritizes human well-being rather than exploiting their weaknesses for endless profit-making? How can technology help people regain human agency, rather than constantly distract them?
The rise of large language models offered me a glimmer of hope.
During my time at Facebook from 2017 to 2019, I led a team building content recommendation algorithms for News Feed. The research was intellectually stimulating, but having worked on delivering more content to users, I found myself questioning the impact of the technology: Does AI-recommended content actually enhance people's happiness?
Software was invented to empower humans. Yet in their reckless pursuit of profits, tech companies have lost sight of that purpose. Social media content is propagating unattainable beauty standards that affect youths’ self-image, short videos are hurting attention spans, and content recommendation algorithms are driving political polarization.
I grew increasingly troubled by the ways technology was harming the very people it was meant to serve. How do we develop technology that prioritizes human well-being rather than exploiting their weaknesses for endless profit-making? How can technology help people regain human agency, rather than constantly distract them?
The rise of large language models offered me a glimmer of hope.
AI for Personal Growth
Until the arrival of LLMs, AI was largely regarded as external and auxiliary to human thinking. It would recognize images, guess what music we liked, and turn audio into text, but it wasn't doing the learning, remembering and imagining that make us humans.
LLMs changed everything. When I first encountered GPT-4, I was blown away by the possibility of AI developing the ability to think like humans, not only carrying out tasks. That gave me ideas to design an AI product in the interest of humans, one that can help us better understand and achieve a better version of ourselves.
During my PhD studies in Illinois, a period when I was experiencing tremendous anxiety over coursework and an uncertain career path, I turned to meditation. Mindfulness practice sent me on a fulfilling path of self-growth. It taught me to watch my concerns without judgment, opening up internal room for peace and creativity.
As beneficial as it may be, reaching a non-judgmental state to converse with oneself is difficult and time-consuming. Yet meditation is just one path there. AI, with its potential to mirror human thoughts or even consciousness, can also be useful in helping people obtain clarity about themselves and eventually, self-realization.
There’s a catch. While LLMs are powerful, they fall short of truly understanding an individual due to contextual limitations. Furthermore, the seas of chatbots built atop LLMs are designed with a subjective vantage point rather than seeing and thinking about the world from the user's point of view.
What if there’s an AI companion that experiences, remembers, and thinks from users' perspectives, an app that grows with users and inspires them to be a better version of themselves, regardless of their backgrounds?
LLMs changed everything. When I first encountered GPT-4, I was blown away by the possibility of AI developing the ability to think like humans, not only carrying out tasks. That gave me ideas to design an AI product in the interest of humans, one that can help us better understand and achieve a better version of ourselves.
During my PhD studies in Illinois, a period when I was experiencing tremendous anxiety over coursework and an uncertain career path, I turned to meditation. Mindfulness practice sent me on a fulfilling path of self-growth. It taught me to watch my concerns without judgment, opening up internal room for peace and creativity.
As beneficial as it may be, reaching a non-judgmental state to converse with oneself is difficult and time-consuming. Yet meditation is just one path there. AI, with its potential to mirror human thoughts or even consciousness, can also be useful in helping people obtain clarity about themselves and eventually, self-realization.
There’s a catch. While LLMs are powerful, they fall short of truly understanding an individual due to contextual limitations. Furthermore, the seas of chatbots built atop LLMs are designed with a subjective vantage point rather than seeing and thinking about the world from the user's point of view.
What if there’s an AI companion that experiences, remembers, and thinks from users' perspectives, an app that grows with users and inspires them to be a better version of themselves, regardless of their backgrounds?
Introducing Me.bot
Driven by this vision, my company Mindverse created Me.bot, a multimodal AI app that logs users’ interactions with the physical world and their internal dialogue. It learns from voice, text, and images like an extension of users' brains, offering insight into memories and inspirations derived from the mix of first-person information that is uploaded.
Me.bot’s incredible malleability is the result of the hard work of our AI research lab, which has been working on AI models tailored to every person by integrating memory into large language models. We call it a lifelong personal model (LPM), a deep neural network model that parameterizes and compresses all types of memory, even the ones that can’t be described by natural languages.
Based on cognitive models of human memory, the LPM begins by interacting with user memories, making predictions grounded in past experiences. The subsequent training involves three data synthesis techniques -- augmenting sentiments and social structures, as well as mapping personal growth.
Unlike RAG, Me.bot processes user input akin to how brains process memories to construct an understanding of individuals, instead of treating data merely as external references. When a user uploads a photo of a boy to ChatGPT, the bot will likely describe it plainly as: "a photo of a boy running." Me.bot, trained with data collected from the first-person point of view, will say something like this:
“It’s Milo’s first week in grade school, and he seems very happy. I bet it's a relief for you that your son is having a good start to a new stage of life! You made a peanut butter and jelly sandwich for him today. How about something else, like a squash, quinoa, chicken wrap for tomorrow?”
In other words, Me.bot doesn't just record one's living memories from a distance; it also empathizes, inspires, and grows alongside users.
Me.bot’s incredible malleability is the result of the hard work of our AI research lab, which has been working on AI models tailored to every person by integrating memory into large language models. We call it a lifelong personal model (LPM), a deep neural network model that parameterizes and compresses all types of memory, even the ones that can’t be described by natural languages.
Based on cognitive models of human memory, the LPM begins by interacting with user memories, making predictions grounded in past experiences. The subsequent training involves three data synthesis techniques -- augmenting sentiments and social structures, as well as mapping personal growth.
Unlike RAG, Me.bot processes user input akin to how brains process memories to construct an understanding of individuals, instead of treating data merely as external references. When a user uploads a photo of a boy to ChatGPT, the bot will likely describe it plainly as: "a photo of a boy running." Me.bot, trained with data collected from the first-person point of view, will say something like this:
“It’s Milo’s first week in grade school, and he seems very happy. I bet it's a relief for you that your son is having a good start to a new stage of life! You made a peanut butter and jelly sandwich for him today. How about something else, like a squash, quinoa, chicken wrap for tomorrow?”
In other words, Me.bot doesn't just record one's living memories from a distance; it also empathizes, inspires, and grows alongside users.
AI Model for Everyone
Since launching Me.bot in July, user responses have been overwhelmingly encouraging. People are tweaking and exploring the bot for a wide array of purposes.
For me, Me.bot is a second brain where I store memories; a non-judgmental meditation teacher who guides my reflection; and a long-time friend through my life’s highs and lows.
Most touchingly, we’ve seen Me.bot become a useful companion to populations that are historically underserved or even exploited by technologies. Users with ADHD have found Me.bot to be the antidote to the internet’s addictive content. Through amicable prompts, Me.bot sends reminders to help them stay on top of life priorities.
In some African countries, users treat Me.bot as life coaches. They whisper to Me.bot their aspirations, upload class notes, and share images of their daily musings. Processing their memories, Me.bot actively sends new reading recommendations and daily affirmations, going well beyond a notetaking app or a conventional "companion".
Of course, AI-human relationships have limitations. One problem is that chatbots don’t always respond factually or appropriately, a challenge the industry is constantly tackling. Users can also become overdependent on AI. By no means does Me.bot try to replace real-world human connections. Quite the contrary, we believe AI that cultivates insightful dialogues with oneself can complement one's existing life.
We envision a future where every person owns their personal AI model. We are going against the tide by creating a piece of software that helps people attain long-term happiness rather than instant gratification, but the opportunity to build technology around human flourishing fills me with hope. Join me in creating a future where AI empowers humans, not ensnares them.
For me, Me.bot is a second brain where I store memories; a non-judgmental meditation teacher who guides my reflection; and a long-time friend through my life’s highs and lows.
Most touchingly, we’ve seen Me.bot become a useful companion to populations that are historically underserved or even exploited by technologies. Users with ADHD have found Me.bot to be the antidote to the internet’s addictive content. Through amicable prompts, Me.bot sends reminders to help them stay on top of life priorities.
In some African countries, users treat Me.bot as life coaches. They whisper to Me.bot their aspirations, upload class notes, and share images of their daily musings. Processing their memories, Me.bot actively sends new reading recommendations and daily affirmations, going well beyond a notetaking app or a conventional "companion".
Of course, AI-human relationships have limitations. One problem is that chatbots don’t always respond factually or appropriately, a challenge the industry is constantly tackling. Users can also become overdependent on AI. By no means does Me.bot try to replace real-world human connections. Quite the contrary, we believe AI that cultivates insightful dialogues with oneself can complement one's existing life.
We envision a future where every person owns their personal AI model. We are going against the tide by creating a piece of software that helps people attain long-term happiness rather than instant gratification, but the opportunity to build technology around human flourishing fills me with hope. Join me in creating a future where AI empowers humans, not ensnares them.
A Better "Me" to Better Me.
Product
Need Help
