Context aggregation is the real battleground for artificial intelligence.
The more contextual information an AI model possesses, the better the user experience. If an AI company has 100 times more contextual information about you than other companies, it will have a significant competitive advantage. The battle for contextual information is more important than the competition among a series of AI browsers.
I. What is Context?
Suppose you've been married for 30 years, have three children, and live in Beijing. Think about how much knowledge you've accumulated over these 30 years about your life, your family, their preferences, wishes, problems, and so on. Who is better suited to answer questions about your life in the past 30 years? Google, Amazon, Meta, DeepSeek? Or your spouse? Of course, it's your spouse. They have decades of life experience.
Ask an LLM:
"How should I renovate my master bathroom?"
Once again, just by asking this question, they have no background information. They don't know what your current bathroom looks like, nor do they know if it matches the style of the rest of your house. Such an answer won't be considered good or relevant. So, those who are skeptical about AI will think it's not very good if they try to get an answer in this way. But if you ask your spouse this question, you'll get a more relevant answer.
AI models need background information related to the situation, question, or conversation to generate accurate, fast, and useful answers. Background information can include:
• Chat history
• User intent. What goals do you want to achieve?
• Domain or task. Does this refer to programming, writing, medicine, history, etc.?
• External data: Data and information on smartphones, Google Drive, printed materials, etc.
• World knowledge. Facts about time, place, people, and relevant entities.
This is not just a matter of big data. Context is rich, associated, time - based, location - aware, and continuously evolving knowledge about the world you're in. Think about the example of a married couple: Context is information and experience. It's an in - depth understanding of your preferences, habits, and past experiences. In this sense, context can be regarded as a digital spouse... an AI that truly understands you. (Well, it sounds a bit creepy now.)
On JD.com, you can experience the application of contextual information firsthand. It knows what you've bought and what you've searched for. So, when you visit again, you'll see a prompt to "Continue your last shopping" or get recommendations for products you might like. Additionally, it will suggest that you repurchase products you've bought before.
For example, Instagram Reels runs a complex algorithm every second to provide you with more short videos you might like based on your past viewing history. But it knows nothing about your Amazon shopping history, Google search history, health information, emails, travel records, friendships, or even all the information on your iPhone. What if it did?
A startup called MemO is dedicated to preserving contextual information. It allows users to save interaction records with different large language models (LLMs) and share these memories among different LLMs, thereby minimizing token usage and repeated learning. In this system, users have full control over their memory context.
II. Not Everyone Uses AI Today
Despite the extensive hype in the media and the market, not everyone has embraced AI. For AI companies, the challenge lies not in convincing advanced users, but in converting the skeptical middle - group. This group of users is the main component of the market and also represents an opportunity.
For those who are skeptical, the problems it faces are the same as those of any new technology: What are its benefits? How is it better than my current way? Are my colleagues/friends using it? Why should I change?
The skeptical neutral group won't accept AI because of the model quality; they'll only accept it when AI makes them feel personalized, useful, and frictionless. Moreover, they'll accept it when AI has enough contextual information to provide a unique user experience for the tasks they want to complete.
III. Context Aggregation
Context aggregation refers to the continuous collection and connection of every aspect of users' lives - what they buy, watch, read, write, and say - to form a unified understanding. The reward is a highly personalized experience that's almost telepathic. Every excellent answer will train you to provide more information, creating a compound interest advantage that may far exceed today's search or social moats.
This advantage encourages you to continuously provide information to the context aggregator, making it seem to know you better than you know yourself. This constitutes a huge competitive moat that may be larger than any other consumer - oriented moat to date.
Context aggregation is the way to become a new type of AI - centric aggregator, as defined by Ben Thompson in the Aggregation Theory.
By integrating users' contextual information, AI - centric companies can provide a unique and excellent user experience (beyond the best - in - class level), which users have never experienced before. It can provide answers and help users complete their work more efficiently, with results far better than before and even beyond users' imagination.
There are already some phenomena of content aggregation on today's platforms. The problem is that each platform only covers a part of your life.
Your current information is scattered among many different technology companies. Google and Meta are both information aggregators. They know a lot about you based on your interaction with their services. But there's also a lot of information they don't know.
You search for travel destinations on Google. They may see the flight and hotel booking confirmation emails you sent, but they won't know all the details of your trip. However, the photos on your phone may record everything.
Imagine if you search for travel destinations on ChatGPT and then wear an OpenAI context - capturing device that can record all the places you've been, the food you've eaten, and the places you've stayed. With this information, can they better recommend future travel destinations for you? Maybe they'll push ticket and hotel discounts at locations that interest you. If their suggestions and details surprise you, will you be more willing to provide more information? And what if it's all free?
IV. The Friction Problem: Why Multimodal AI is Crucial
Current large language models (LLMs) are text - prediction models. Inputting and reading text is a major difficulty for AI to be applied to the mass market.
Multimodal input can accelerate the development of AI. Each new input method - voice, photos, cameras, sensors - can reduce the cost of providing contextual information. That's why multimodal AI will spread much faster than web or mobile technologies.
For those who are skeptical, voice and photos seem more acceptable. You can upload photos now and ask for suggestions on room renovation. But not everyone knows this. Moreover, a single photo can't show the whole picture of the house, nor can it reflect the occupant's preferences for interior design. Photos are helpful, but they're still just the tip of the iceberg. They can't show the whole house, understand your previous choices, or know your partner's preferences.
Once again, if the device you wear can take photos of your house as you move around and allow you to provide voice explanations when needed, more relevant contextual information can be captured with less resistance.
Reducing friction can not only build context faster but also enhance the network effect. Each new user will contribute more meaningful context to the system.
V. The Advantage Flywheel
Existing aggregation platforms like Google, Netflix, Amazon, Meta, and Airbnb serve a very broad horizontal market. As Ben Thompson pointed out in the Aggregation Theory, the more users they serve, the better the service quality.
There are some differences here. Trust and privacy will play a greater role in whether users are willing to provide more background information. And, undoubtedly, the results provided by AI must be excellent at least in some specific aspects for this virtuous cycle to work.
How will the situation develop if context - aggregation technology is adopted? Is it possible for an AI company to aggregate enough contextual information to provide a unique user experience for new users without providing them with too much contextual information? Maybe they can obtain enough contextual information from new users to predict a lot of contextual information that users haven't provided? This will enable AI companies to integrate their aggregated information faster.
Summary
Of course, context aggregation involves many profound ethical and privacy issues. This is a huge topic in itself and worthy of separate discussion. Naturally, regulation is also a topic that needs comprehensive discussion.
In this battle, the ultimate winner won't be the one with the largest model or the best AI browser, but the one with the richest and most trustworthy contextual information, serving the most users, and transforming this information into an unprecedented and unique user experience. This user experience can not only precisely meet users' needs but also bring a much richer experience than expected. Such an experience will further consolidate the position of context aggregators as "almost all - encompassing" in the field of AI.
However, to succeed, such a user experience must help the public, especially the skeptical middle - group, complete things faster, more easily, and more economically in ways they've never imagined. And, as I said, it must be presented in a way that amazes users, beyond their imagination.
If an AI - centric company wants to win the context battle, what types of services and devices should it develop? When Sam Altman talks about achieving super - intelligence, what role does it play in the context battle? Do users need to feel that they have super - intelligence to establish a connection with an AI partner with rich contextual information?
Finally, let's do a thought experiment: Try to think about what a super - intelligent AI would do for you to make you consider providing it with more information about yourself.
For software developers, coding is a good micro - example. Tools like Claude Code (and other similar tools) understand your entire codebase, documentation, tests, and commit history. You trust them because the results they return are often amazing.
The same trust cycle, where context brings better results and thus more trust, will determine the winner of the context - aggregation battle.
This article is from the WeChat official account "Data - Driven Intelligence" (ID: Data_0101), author: Xiaoxiao, published by 36Kr with authorization.