In the digital age, we process various information repetitively every day, which is not only time-consuming, but also prone to cognitive fatigue. This article introduces an innovative solution – Second Me, an AI assistant that fully represents you.
Why do we need a “second brain”? —— Understand the core issues with life scenarios
Imagine that you turn on your phone every day and need to fill in your address, phone number, hobbies and other information repeatedly in different apps; When meeting with customers, always repeatedly explain the details of cooperation that have been discussed before; Even logging in to different websites requires entering passwords again and again…… Do these “memory repetition” scenes seem like moving the same stone every day?
This is the “cognitive fatigue” of modern life – our brains are forced to waste a lot of energy on repetitive information exchanges. Existing technology (such as browser autofill) is just an “information storehouse” that does not think or adapt. For example, if you fill in the address in a shopping app, it stands to reason that after you change the address when renting, other apps should automatically update, but the reality is that you have to modify them one by one – because they are only mechanically stored and do not “understand” the changes in your life.
Of course, this AI agent is based on other basic large models, and then superimposed on some of your personal memories, or native, it is a personalized AI that belongs only to you, that is, an AI that belongs only to you.
The core goal of Second Me is to free us from this “memory coolie”. It is not a simple database, but an intelligent assistant that can “understand you, organize information, and respond dynamically”. It’s more like a “translator who understands you” – not only remembers your information, but also actively helps you generate the most suitable replies or fill in content based on current scenarios, such as interviews, shopping, and social networking.
How does Second Me work? —— The “Memory Magic” of the Three-Layer Structure
1. Bottom: Raw Data Layer (L0) – Your “Memory Store”
This layer is like a big box that holds all your raw data: chat history, documents, form fillouts, and even browsing history. For example, if you mention “I like to drink coffee” in an email and “exercise three times a week” in a memo, these will be deposited into L0. But it doesn’t actively process this data, it’s just “truthfully recorded”, similar to the original knowledge base of RAG (Retrieval-Augmented Generation) models.
What does a product manager need to do?
In the process of a product from scratch, it is not easy to do a good job in the role of product manager, in addition to the well-known writing requirements, writing requirements, writing requirements, there are many things to do. The product manager is not what you think, but will only ask you for trouble, make a request:
View details >
2. Intermediate layer: Natural language memory layer (L1) – the organizer who can “speak”
L1 is like a thoughtful secretary, translating the raw data in L0 into a language that humans can understand and organizing it into structured information. For example, it will extract “name, phone number, address” from your chat history to form a “personal profile”, summarize “favorite brands and consumption habits” from shopping records, and even summarize “work field and project experience” from emails. This information exists in the form of natural language tags, such as “Favorite coffee brand: Starbucks; Fitness frequency: three times a week, prefer strength training”.
3. Core layer: AI native memory layer (L2) – the brain that can “think”
This is where the soul of Second Me lies. L2 does not store memory in language, but “encodes” your personal knowledge through the parameters of the neural network. For example, if you repeatedly mention “emphasizing time management”, the neuronal connections in L2 will reinforce the “time management” related patterns; You often mention “machine learning algorithms” in technical documentation, and L2 forms a specific combination of parameters to represent this knowledge.
For a simple example: If you often say to a friend, “I like to go for a walk in the park on weekends because it relaxes my mind,” L0 will save the text of this sentence, and L1 will extract “Hobbies: walking; Reason: Relax your mind”, and L2 will remember the relationship between “walking” and “relaxing” through model parameters, and the next time you mention “I want to do something on the weekend”, L2 can actively associate the suggestion of walking and adjust the recommendation based on the current scene (such as the weather).
How to train an AI that understands you? —— Fully automatic assembly line
One of Second Me’s major innovations is the “fully automated training pipeline”, which allows everyone to quickly have their own AI. This process is like developing a “personal assistant” and is divided into three stages:
1. Data preparation: cleaning, mining, and synthesis
First, data cleaning is like tidying up a room, throwing away useless information (such as duplicate records, spam) and keeping “clean” personal data. Then dig up information and use AI to analyze your data to extract key entities (e.g., people’s names, places) and topics (e.g., jobs, interests). For example, extract “work experience: software engineer of a company” from your resume, and find “favorite book type: science fiction” from your reading record. Finally, the data is synthesized and a large language model (such as GPT-4) is used to simulate real scenarios and generate training data. For example, if you are an entrepreneur, the model will simulate the scenario of “customers asking about product features” and generate possible answers to help Second Me learn how to communicate on your behalf.
2. Model Training: From “General Student” to “Dedicated Expert”
They start by doing efficient fine-tuning, which is like giving students a “crash course” to quickly adapt the model to your data with a small amount of computing resources. For example, QLoRA technology can be used to reduce training costs while maintaining model performance. After that, supervise fine-tuning and show the model a “standard answer” – for example, you manually mark “When someone asks about your age, answer: confidential”, so that the model learns to “imitate your behavior patterns”. Finally, DPO preference optimization is performed to compare the advantages and disadvantages of different answers, so that the model can better understand your preferences. For example, if you might prefer “concise technical jargon” to “lengthy explanations,” DPO will allow the model to prioritize generating content that aligns with your style. Using 20% of the data for DPO optimization is like giving students “targeted coaching” to strengthen their weak links.
3. Evaluate and iterate: Make AI “smarter the more you use it”
There are mainly a few core steps here, likememory question and answer,Divide it into “self-questions and answers” (e.g., “What did I buy last week?”) and “third-party Q&A” (such as “What is Xiao Ming’s hobby?”) ), to test the model’s grasp of personal information.
And alsocontextual completion,For example, if you enter “I need to book a meeting room”, can the model automatically add information such as “time, number of people, location”, etc.
It was also usedcontextual criticism,When an external AI gives a recommendation (e.g., “recommend you to buy a product”), the model can evaluate and refute it in conjunction with your preferences (e.g., “limited budget”).
Finally, human evaluation may underestimate model capabilities. For example, if the model generates a response that is “reasonably inferred but does not directly quote data” (such as recommending a new product based on your shopping history), the machine may deduct points for “unmatched keywords”, but humans will consider it a “manifestation of intelligence”. Therefore, human assessment is crucial, just like a teacher grading homework, not only looking at the answers, but also looking at the “thought process”.
Why is the “thought process” important? —— The power of the chain of thought
Imagine two students doing a question:
- Student A writes the answer directly: “3+5=8”
- Student B writes the steps: “3+5=? Calculate 3+2=5 first, then 5+3=8, so the answer is 8”
Although the results were all correct, student B had a clearer idea and was easier to spot mistakes. Second Me’s “chain of thought training” is just that: let the model not only give answers, but also show “how to think”.
They tried three chain-of-thought strategies:weak mode,Answer freely, similar to student A, fluent but lacking in detail.multi-step mode,In two steps, reason first and then give an answer, similar to a simplified version of student B.strong mode,Use professional models (such as DeepSeek-R1) to generate detailed reasoning processes, such as “Xueba writes a problem-solving report”, with a clear structure and technical details.
Finally foundStrong COT works best。 For example, when answering “the user’s career plan”, the strong COT model will derive suggestions step by step based on the user’s work experience, training records, “promotion goals” in chats, and other information, while the weak COT may only give general answers. This shows thatA clear thought process allows the model to understand and use personal memory more accurately。
What can Second Me do for you?
So what can Second Me do for you, for example
1) Daily Efficiency Enhancement:
Say goodbye to repetitive labor auto-filling: When you register in a new app, Second Me will automatically fill in the most suitable information based on L1’s “profile” and L2’s “scenario understanding” (such as an address for shopping apps and interest tags for social apps), and even help you generate personalized profiles.
2) Meeting Assistant:
Before the meeting, it will sort out the historical discussion records and refine the key points;
During the meeting, you will be reminded of the “details of the cooperation mentioned before” in real time;
After the meeting, the minutes are automatically generated and synced to the schedule. Help you make decisions and become your “rational brain”
3) Information filtering: Among the massive information (such as news, emails, social updates), it helps you filter out content related to “personal goals”. For example, if you pay attention to “artificial intelligence”, it will prioritize pushing industry reports and filtering entertainment news.
4) Or it may berisk reminder,When you consider investing in a project, it will analyze historical similar cases based on your “financial situation” and “risk appetite” (stored on L2) to give objective recommendations.
5) It can also bePersonalized AI network,When multiple Second Me systems are networked and collaborate, it may form “distributed intelligence”. For example, you and your colleagues’ Second Me can securely share project information, automatically coordinate workflows, and improve team efficiency.
6) I think the most interesting thing is that it can be donecognitive asset inheritance,Through technologies like NFTs, your Second Me model could become a “digital legacy,” passing on your knowledge, experience, and mindset.
This is a bit like Jarvis in Iron Man, which is a small prototype, Jarvis is an exclusive AI built entirely on Tony Stark’s thinking patterns, knowledge systems and behavioral habits, which can accurately understand his needs, call warframes, analyze tactics, and even ridicule and complain, and the core goal of Second Me is also trained through user personal dataDedicated AI modelsto allow the AI to understand the user’s preferences, habits, and context. From this perspective, both are “human-centered” intelligent extensions.
