Over the past year, the wave of large models represented by ChatGPT has profoundly transformed the landscape of the artificial intelligence industry, from basic model layers to practical applications and beyond.
Accordingly, while large models have been touted as the backbone of future computing, there’s still uncertainty about how to support the applications built on top of them, whether in terms of hardware or software.
Recently, 36Kr came across KMind, a Chinese company focused on developing what it describes as the concept of a “personal AI computer (PAIC).”
KMind was founded by Wu Hanqing, a researcher and former head of security technology at Alibaba Group. Before founding KMind, Wu worked in a team led by Wang Jian, the founder of Alibaba Cloud. Since 2020, Wu has been focusing on the field of AI, exploring adjacent areas like cloud gaming, cloud rendering, and other businesses under Alibaba Cloud, before resigning in May last year. KMind’s team also includes former Alibaba engineer Chen Dongbai and Yu Kaicheng, head of the AutoLab at Westlake University in Hangzhou.
“If large model companies are compared to CPUs, then many companies that implement AI are applications on computers. However, applications cannot run directly on CPUs, and they want to make up for this ‘last mile’ by connecting the large models with the applications to create a complete loop,” said Wu, when asked what the future of AI could look like.
Personal AI computers
In terms of how users interact with their PAIC, KMind proposes the concept of “star companions.”
A star companion is a personal assistant powered by kOS, a new AI-based operating system developed by KMind. It will automatically converse and work with users according to their intentions. Data gathered from usage will be stored in a place called “star soul” to support personalization features, enabling star companions to better understand their users over time.
Additionally, star companions can interconnect to communicate and exchange information with one another, enhancing their user support capabilities.
Users can also upload and manage data using what KMind has termed as a “star disk,” serving as a personal knowledge and experience bank. According to KMind, data stored in star disks will not be used to train AI models.
“Action units (ACTs)” are used to describe the various tasks that can be performed by star companions, which can range from casual conversations to intricate work assignments. Star companions can be utilized in either the “companion” or “work” mode. The former deploys them like intimate friends who can strike up a conversation, while the latter configures them to complete a variety of work-oriented tasks. Star companions can utilize predefined ACTs when set up in the latter mode.
In the future, new ACTs can be developed to serve a more diverse pool of user needs and requirements.
It’s worth noting that, after leaving Alibaba Cloud, Wu and his team’s initial endeavor was the launch of Kuixing in July 2023, an intelligent B2C assistant that supports functions such as document summarization, drawing, and making short videos. Although it accumulated 100,000 users within three months of its launch, Kuixing could not solve the hallucination problem of large models, ultimately resulting in a low user retention rate.
This experience prompted the team to reflect on and review the problems of large models, culminating in the eventual conceptualization and launch of the first generation of programmable PAICs.
Designed with user accessibility in mind, KMind’s PAICs are powered by kOS, a proprietary system based on the principles of information, control, and systems theories, focused on understanding user intent and completing tasks with high precision. Leveraging a “brain” controller, it can understand user intentions, decompose complex tasks, and schedule different execution units to complete user tasks.
This system is designed to unify neural networks and advanced programming languages, combining the generalization ability of neural networks with the logical reasoning ability of high-level programming languages to enhance accuracy in responses. Simply put, it hopes to retain the strengths of both large models and programmers.
To illustrate the strengths of both approaches, Wu uses an example from the well known novel, Journey to the West. While large models struggle to answer how many times “Sun Wukong” is mentioned due to his various aliases such as “Great Sage,” “Old Sun,” and “Monkey King”, traditional programmers can address this issue seamlessly using SQL or Python.
Selling computing power, not data
While startups may look to OpenAI for guidance on business strategies, KMind is charting its own course. Wu contended that internet giants have become central hubs for aggregating information, gradually monopolizing data and privatizing user-generated content for profit. This trend raises concerns about privacy and threatens internet freedom.
In response, KMind’s founders have embraced an open-source approach. They plan to open-source recommendation and search algorithms once the system matures, managed by an open-source committee. According to Wu, KMind’s mission is to empower individual users to derive value from their data, diverging from the monopolistic tendencies of mainstream companies.
Wu told 36Kr that there are currently two prevailing business models in the industry: the ad-based revenue model exemplified by Google, and the direct user payment model employed by OpenAI. KMind finds fault with both approaches—the ad-based model risks moral compromise, while OpenAI’s recent management upheaval underscores the tension between its non-profit origins and commercial aspirations.
Consequently, KMind intends to monetize by selling computing power rather than user data. According to Wu, this model would foster collaboration within the KMind ecosystem. While the company supplies computing power, community programmers will contribute algorithms and users can provide data, ensuring that data ownership remains with the users while programmers benefit from the value generated by their algorithms.
Given the resource-intensive nature of computing and servers, KMind has eschewed a free version, opting for a subscription model akin to purchasing a smartphone, where users pay a monthly fee for access to computing resources that can be upgraded as needed.
Wu declined to outline specific future plans for KMind. However, he expressed that KMind’s current objective is to democratically connect two billion people worldwide through AI.
This feature article was written by Wang Yining in chinese and was originally published on 36Kr.com
