A product information architecture diagram essential for AI product managers.

Taking advantage of the weekend, I'm spending time on the scientific research path.

After transitioning from a product manager to a computer science Ph.D. candidate, the most significant change I've noticed is the substantial increase in the technical barriers of products. Leading my team to explore interesting algorithms and open - source project frameworks in scientific research papers, and then using them as a basis for product framework design and development, gives our products a much more solid technical foundation than before.

Currently, many product managers are still focused on functional product design and interactive interfaces, concentrating on entertainment and large - company tools. They are unable to plan products with cutting - edge technological features to meet user needs.

After obtaining a computer science Ph.D., based on the product research direction and company strategy, by referring to scientific research papers and applying research thinking to product development, the technical barriers of products will be directly enhanced. There's no need to worry about performance either, as research papers surely contain comparative experiments and results.

Recently, I read an article suggesting that the ultimate stage of development for AI product managers is to design products for the AIOS system. Here are a few skills for AI - native product managers.

With the emergence of more and more AI coding IDE platforms, I believe they are the predecessors of AIOS, including various current agent - based products. Here is the information architecture of the AIOS system.

AIOS Will Be the Next Opportunity for OS System Product Managers

Under the AI model, native AI apps will have brand - new UI interfaces. Instead of presenting users with pre - defined boxes or rectangles, the interfaces can be generated in real - time according to users' needs.

This technological revolution driven by AI model generation means that all the familiar product development concepts need to be fundamentally changed. For example, when traditional products add AI features, it's usually "AI +". Although the business scenarios of AIOS users remain fixed, the interfaces are no longer composed of fixed buttons, labels, and entrances. They can be adjusted according to users' changing scenarios and needs.

The above is the 3D world launched by wordlabs. During the user experience, the routes are generated by AI, rather than being fixed paths and backgrounds.

The Fixed Product Framework of AIOS: Input and Output

In AIOS, product managers still need to provide users with operation buttons, input boxes, and even some fixed interfaces to complete specific business operations.

The learning curve for users is quite steep. If the interfaces are completely new, users will need more time to get familiar with them. Therefore, in the AI operating system, familiar applications such as browsers, note - taking apps, and social chat tools remain, but their interfaces are not fixed. They can be generated in real - time or upgraded according to users' preferences.

In this research paper, the information technology architecture of the AI operating system is mentioned. It uses large - language models as the infrastructure and stacks systems on the same hardware. There are three layers in total: the application layer, the AIOS core layer, the traditional OS core layer, and the hardware. For parts without user interaction and logical reasoning requirements, the traditional OS layer is used for access, while the AIOS core layer is responsible for parts with user input and interaction.

Familiar IDE tools like Cursor and Manus are gradually evolving into architectures similar to AIOS. They are equipped with various tools, which are integrated into the AIOS core part of the AIOS system. For example, the coding environments and preview tools we mentioned can be part of the scheduler and tools sections, improving efficiency.

The AIOS core part includes planning, action, memory, storage, and scheduling control. In traditional AI applications, when a user issues a task to book a flight to San Francisco, the agent completes seven steps, including information retrieval, flight and hotel recommendations, and the final summary. All these steps are carried out by the LLM. However, data verification, payment, and calendar scheduling are handled by AI, which is an inefficient use of AI resources and becomes increasingly cumbersome as the number of tasks grows.

As the number of agent tasks increases, the processing of the LLM and the information retrieval of the OS become more frequent. Managing computer hardware resources and model tasks has become the solution for AIOS, leading to the following information system architecture.

The application layer comes from AI applications. The processes, flash memory, system files, and hardware drivers in the kernel part are managed by the traditional OS kernel. Daily tasks, text tasks, memory tasks, storage management, tool management, and permissions are all handled by the LLM kernel, forming the underlying layer of the OS system.

The Proxy Services Available Here Are

Agent Scheduler: Prioritize and arrange agent requests to optimize the use of LLM.

Context Manager: Support snapshot and restoration of intermediate generation states in LLM and manage the LLM context window.

Memory Manager: Provide short - term memory for each agent's interaction logs.

Storage Manager: Persist the agent's interaction logs to long - term storage for future retrieval.

Tool Manager: Manage the agent's calls to external API tools (e.g., search, scientific computing).

Access Manager: Enforce privacy and access control policies among agents.

In the AI OS operating system, there are bound to be multiple model agents. Therefore, the switching and internal communication between multiple model agents are crucial. The AIOS operating system provides kernel switching between different models to ensure that the optimal operating system can run under different specific tasks.

AI Programming IDE, the Next - Generation AIOS System

Currently, when choosing AI models for IDE programming, it's still a manual process. Users need to select the most suitable large - language model according to specific tasks.

Although there are multiple SDK management systems in AIOS to facilitate AI models in driving various tasks, the following are the registration, production planning, creation planning, update planning, and initial planning mentioned in the paper, which cover the lifecycle management of agents, effectively reducing latency. As shown in the AIOS experiment mentioned in the paper, in tasks with AIOS, the task waiting time is reduced by nearly half.

Product Managers of AI Operating Systems: Focus on Imitating Human - Computer Interaction

At the application level, the interaction of the AI operating system imitates the way humans use various terminal services on a computer, such as browsers, office documents, and coding environments. Product managers need to map these steps to the AI OS, rather than simply providing users with an execution result without allowing them to monitor the process, which increases users' trust in the product execution process.

Similar to the product interface of Manus, there are corresponding task progress bars, task times, and task operation pages. Including the interaction in the CHATGPT agent mode, product managers provide previews of web pages and document pages that users can view during computer operations.

On the left is the task dialog box, and on the right is the task flow. Users can view the operating environment for unpacking file packages and browsing web pages in the task flow.

The above tasks are carried out with a single agent. In theory, we can encapsulate more agents to obtain different results and then select the optimal one. For example, sending the same task to ten different Manus - like products will yield ten results for comparison.

The new version of Manus is evolving in this direction, supporting concurrent execution of multiple task agents, and the agents can also collaborate with each other.

The following is the AI scheduling mechanism in the AIOS system mentioned in the paper, which allows different AI models to be scheduled according to specific rules to achieve high - memory usage.

This is the so - called application part and scheduling mechanism of the AI operating system LLMS.

AI Product Managers Will Become the Hottest Profession in 2 - 3 Years, No Doubt

With IDE programming and interactive design tools, almost anyone can become an independent developer, provided they can identify the combination of users' new needs in the AI era and past scenarios.

I believe that in the next two to three years, all traditional software on the market will be replaced by native AI products. For any product, the input and output functions and interaction paths for each operation are fixed, but the interaction results of other AI tasks are not unified. There is no fixed GUI, and it can change according to users' needs.

That's all for today's sharing.

Paper Reference:

https://arxiv.org/pdf/2403.16971

This article is from the WeChat official account “Kevin's Little Steps to Change the World” (ID: Kevingbsjddd). The author is Kevin's Stories. It is published by 36Kr with permission.