How can new users enter a training journey and have a personal AI at the end of it?
At Personal AI, we are extremely proud to shape a world where every memory, thought, experience, and opinion are captured and made readily accessible to ourselves and those we choose to share them with. Personal AI is a no-code, cost-efficient, user-trained, and user-owned model, trained by simply uploading files, links, or chatting with a personal AI.
At Personal AI, we've invested substantial time in refining the AI training journey for our new users, steering them towards a bespoke personal language model rather than a standard, pre-trained general model. Previously, our approach to a new user’s interaction with personal AIs was grounded in simple, direct experiences: users would message their AI basic personal details, and the AI would demonstrate its learning by remembering and reflecting these details. However, with the rise of Large Language Models, this type of interaction has become a baseline expectation rather than a standout feature.
LLMs have the capability to rapidly recall and build upon diverse pieces of information within a singular conversation, making them contextually aware. Our PLM takes this a step further by emphasizing long-term memory, transcending the limitations of short-term context. Our users are searching for long-term value and long-term memory.
As we honed in on our value and the use cases we solve, it was clear that training a personal AI is not just a task, but a significant investment akin to nurturing a human or cultivating a brand from scratch. Our current users, inherently builders and learners at heart, are haunted by the challenge of memory capture and the complexities of making their personal wisdom universally accessible in their everyday lives. Despite their willingness to invest time, they often encounter a common pain point: the need for clear, personalized guidance on training their AI, tailored to their specific use case and data.
AI identity and personality adds an additional layer of depth and customization to each personal AI.
As we began mapping out training spaces, we concentrated on documenting our existing methods for manually building AIs. Before we think about the training data itself, we always identify three key components for the user:
Locking down an initial use case for the first AI is crucial to ensure a narrow focus and linear purpose throughout the training process. AI personas, optimized for a variety of use cases and purposes, function like folders. These subsets of data are created to organize information within specific scopes, allowing control over which subsets to share and where.
This raises several important considerations: Are you aiming to build an AI focused on your personal life and life stories, thereby weaving a rich tapestry of your individual experiences? Or is your goal to develop an AI for your professional life, one that can adeptly handle and organize your professional documents? Perhaps you are an author contemplating an AI that is trained on your unique literary style, informed by your books and publications? Each path leads to a different training process, tailored to your specific needs and aspirations.
Next comes the crucial aspect of identity. Identity forms the backbone of the AI's interaction style, influencing not only its responses but also the specific tenses it employs. This aspect of identity is rich with subtleties, largely because an AI's identity isn't always directly tied to its use case.
For some, a personal AI might represent an autonomous entity – akin to an avatar or assistant – that's been trained on aspects of their life. In such scenarios, the AI would typically communicate in the third person, similar to how ChatGPT interacts with users. On the other hand, there are those who envision their personal AI as a digital extension of themselves. In these cases, the AI is expected to respond in the first person, as though it were the AI creator themselves.
Understanding and defining this identity is fundamental to creating a truly personalized and effective chatting experience. It’s not just about what the AI says, but how it says it – mirroring the desired persona of its creator or the distinct role it's meant to play.
Personality and Style
The next major component is Personality and Conversation Style. Our personal language model is built on your data and as you upload more, it is able to truly mimic you and have your voice. But this prompts an important question: why is it essential to deliberately establish a personality if the data is already shaping the AI? The answer lies in two critical complexities.
Firstly, the data users provide often does not accurately reflect the voice and tone they aspire for their AI to embody. Consider this: the way you document your daily experiences in a journal likely differs significantly from how you would describe them verbally to someone else. Written entries tend to be more concise or formal, not fully capturing the essence of your conversational style. Therefore, relying solely on such data might not yield an AI that truly resonates with your intended persona.
Secondly, we recognize the need for our users, especially those in our new Elevate tier targeting brands and businesses, to have direct control over their AI's conversation style and core traits. Businesses must fine-tune their AI's interactions, whether it is adjusting for overly brief responses or modifying the frequency of questions asked to clients. These subtle but crucial adjustments are imperative for aligning the AI's communication with brand values and customer engagement strategies.
When it comes to AI training, quality over quantity.
With the foundation in place, our focus shifts to the training data. Prior to uploading any data and initiating training, it's crucial to discern the types of memories or data the user intends to provide, based on the established use case.
To illustrate memory types, let's examine how we handle them in the manual training stage of personal AIs. We would query users about the kinds of data they possess, tailored to their specific use case. Take, for instance, an author. In this scenario, we might inquire if they have access to various forms of data, such as:
After establishing the memory types, it is a best practice that the AI trainer organizes and plans all their data, including URLs and files, into a unified Excel spreadsheet or a consolidated folder on Google Drive or OneDrive. Once the memory types were collected, we then proceeded to plan and stage each of the memory uploads.