Original · Unique Research · 2025-11-14
Editorial note: This is a complete English edition of the historical article and interview. Product capabilities, company age, technology assessments and predictions reflect the source’s reporting and Ma Liang’s statements, not independently verified current results. “Silicon Geek” and “Desktop Companion” are descriptive English renderings of 硅基极客 and 桌搭精灵, not verified official English names. The source uses “ROM” without defining it and describes DeepSeek’s work using “RLHF”; those terms are retained as written rather than silently replaced or interpreted as proof of a particular implementation. Relative time references remain anchored to the historical article.
What if a genuine working partner lived inside your computer?
Not the kind of assistant that occasionally pops up a dialogue box asking whether you need help, but an always-available digital colleague that can operate any software, remembers your habits and may even understand your working rhythm better than you do. As you sit in front of the screen, it works deep inside the system—calling tools, opening applications, moving files, writing scripts and searching for information. Everything is completed quietly after a few instructions in natural language.
That is exactly what Ma Liang, CEO of Silicon Geek, wants to build.
He defines the company simply: use AI to reinvent the familiar software and hardware he knows. As a first step, he chose to transform the Quick Launcher category of productivity software he regularly uses, creating an AI productivity assistant for the computer.
I. From productivity software to an AI partner on your desktop
Silicon Geek was founded less than six months ago. Its flagship product is AIdo, a desktop Agent that runs on a PC. Think of it as the AI-powered super-evolution of productivity software: instead of adding another feature menu, it adds someone to your computer who understands you.
What can it do? On the surface, it is a command centre capable of executing hundreds of tools: compressing images, editing video, revising documents, researching information, issuing instructions and running scripts. The bigger innovation is hidden from view: how to schedule and make decisions across a vast toolset, and how to turn the operating system into a runtime environment that AI can control without burdening the user.
That sounds highly technical, but in everyday language it comes down to two things:
First, you no longer need to remember which tool, website or plug-in a task requires. Tell AIdo what you want to accomplish, and it finds the most suitable tool.
Second, instead of merely clicking and dragging within computer applications on your behalf, it can call tools directly in the background, complete the task and deliver the result.
To make this interaction more natural, Silicon Geek also created a piece of hardware called Desktop Companion, a desktop assistant that lets users control a computer by voice. Many people initially compare it with a smart speaker, but it is very different from the generation of products built around playing music and checking the weather. Those products largely remained conversational; Silicon Geek cares more about getting real work done. Say the word, and it actually begins working on your computer.
Ma Liang's original motivation is straightforward: in the great AI era, create something that genuinely changes how the world operates.
More specifically, he wants to benefit knowledge workers, reduce cognitive load and improve efficiency—ultimately becoming a defining force in next-generation AIPC human-computer interaction, so everyone can have a dedicated and almost all-powerful AI partner.
II. From writing poetry to getting work done: the Agentic AI inflection point
To understand Silicon Geek's path, we must begin with the technological inflection point facing the industry as a whole.
Over the past two years, discussions of large language models (LLM) have readily brought to mind writing copy, drawing images and producing simple code—the content-generation capabilities known as AIGC. They are impressive, but have largely remained at the level of creation.
The real dividing line appeared when two changes converged. First, large models became much better at understanding and planning. Second, the spread of protocols such as function call, MCP and A2A enabled models to invoke external tools more reliably.
Put bluntly, AI used to resemble a literary young talent capable of eloquent prose and inspired writing. Now it is becoming a practical operator that can plan tasks, orchestrate tools and learn through repeated trial and error. The industry has given this shift a name: Agentic AI.
In many products, Agentic AI remains in the cloud—running a workflow, setting up automation or connecting a series of API services. Silicon Geek has instead done something that may sound unglamorous but is technically profound: it has applied these capabilities at the PC operating-system layer.
Everything that happens on a computer—opening software, moving files, selecting an area of the screen and calling local programs—becomes something AI can operate.
On top of that, AIdo adds a personalised strategy for choosing among a vast range of Tools. It makes dynamic selections based on your instructions, preferences and historical behaviour, like a team leader permanently stationed inside your computer who understands you better over time.
Give it a complex task and it does not merely generate text explaining what you should do. It carries out the work and genuinely completes the task.
III. At the height of the technology cycle, how do you avoid building a city no one inhabits?
Almost every AI company talks about leading technology, powerful models and expanding capability boundaries. The constraint in the real world is much simpler and harsher: if nobody uses it, even the most advanced technology is meaningless.
Silicon Geek is still at a very early stage, but its methodology is grounded in reality:
Iterate day by day.
Instead of holding everything back for a grand release, the team launches internal and public beta tests quickly as it refines the product, moving forward together with real users. Each day it solves a seemingly minor problem: is this button unnecessary? Can this step be automated? Is this prompt too difficult to understand? In this way, it gradually moves closer to the central pain points of specific vertical scenarios.
The balance between technological development and commercial execution is not always expressed in a strategy PPT. Sometimes it is embedded in this rhythm: are you willing to accept that today's product is imperfect, yet still let users try it and listen to honest feedback?
Behind this lies another reality that is rarely stated directly: large models themselves are competing rapidly on price, performance and open ecosystems. Application-layer startups that still imagine they can overpower the world through technical specifications will quickly receive a lesson from reality.
The genuinely difficult task is finding an entry point close enough to revenue, close to people and far enough from the major technology companies.
The PC desktop Agent is the entry point Silicon Geek is betting on.
IV. Integrated software and hardware: the hidden advantage of Chinese teams
When the conversation turns to globalisation, Silicon Geek's AIdo has not yet truly gone abroad, but the outline of its route is already visible:
On one side is extensive tool invocation through the MCP ecosystem; on the other is hardware capability closely connected to China's supply chain.
Combining software and hardware in a PC productivity assistant is precisely the sort of integrated move that many overseas teams are relatively less equipped to execute. Handling models, systems, applications, hardware and the supply chain simultaneously sets a substantial barrier to entry.
The opportunity is clear: if it succeeds, the result will be an integrated experience that is extremely difficult to replicate.
The challenge is equally real: many MCP-based services differ enormously across countries. Tools for local life and online services may have to be rebuilt almost from scratch. A company must genuinely understand how local users live, rather than merely translating the language at the code layer.
From a broader perspective, Chinese AI companies are gradually moving from open-source leaders, application-innovation pioneers and cost optimisers to originators of new paradigms.
DeepSeek's large-scale use of reinforcement learning during the RLHF stage and its proposal of contextual optical compression in OCR both point to the same development: Chinese teams are no longer merely following paths pioneered elsewhere at lower cost, but are beginning to explore new technological routes.
A small team such as Silicon Geek, working on desktop agents plus hardware, is also testing a larger question: beyond cloud-based API services, is there another route for human-computer interaction that a Chinese team could be first to prove?
V. The dividing line between advanced and lagging individuals
The theme of this conference is “Pioneering Intelligence | The Individual Era.” Taken apart, the phrase carries an uncomfortable practical implication: the gap between individuals is widening.
The barrier to using AI is falling. Almost anyone can open a webpage, type a prompt and receive a reasonably good answer.
But what truly creates the gap is not whether you know how to use AI; it is how much of your real output you allow AI to participate in.
Consider a typical scenario:
Two people are preparing the same kind of report.
Person A opens PPT and starts from scratch—laying out pages, searching for images, creating tables and polishing the presentation one page at a time.
Person B gives a Word outline to AI and asks it first to break down the structure, source images and lay out an initial draft, then adjusts the result using their own judgement.
By evening, A is still wrestling with details. B has already completed two additional iterations and even has the energy to prepare a presentation script and Q&A notes for the report.
The gap does not arise from a flash of inspiration at one moment, but from the dozens of daily decisions about whether to let AI try first.
Ma Liang's assessment is candid: some lower- and mid-level roles will inevitably face severe disruption from AI.
As an individual, the only thing you can do is place yourself on the advanced side of that divide—learning to use AI to unlock creativity and productivity you previously would not have imagined possible.
VI. When a project partner moves into your computer
Returning to AIdo, consider several concrete scenarios to see how change unfolds in the individual era.
You are preparing to upload an image to a website, but the upload limit is 1MB. Previously, you might have opened PS, Preview or an online compression tool and worked through several steps.
With AIdo, you simply drag the image onto it and say, “Compress this to below 1MB.” Compression, resampling and saving again—actions previously scattered across several applications—are bundled into one natural-language task.
You are editing a talking-head video, and one of the most time-consuming tasks is removing each silent gap individually.
Using AIdo is straightforward: drag in the video and say, “Remove the silent sections.” It handles everything else.
You have written a long Word document and want to add an appropriate image to every paragraph to strengthen the presentation.
In the past, you either searched for images one paragraph at a time or abandoned the idea.
Now the action is still simple: drag in the document and add one sentence—“Give each paragraph a suitable image.” Behind the instruction is an end-to-end chain of text understanding, image generation and layout editing, but for you it is a single conversation.
These apparently minor tasks are, in essence, a continuous transfer of repetitive mechanical work from the human mind to an Agent.
When operations like these occur dozens of times throughout the day, you suddenly realise that the computer is no longer merely a cold toolbox. It has become a project partner to which you can assign an outcome.
VII. The opportunities and anxieties facing AI creators
In Ma Liang's field—desktop Agents and AI assistants—opportunity and anxiety are two sides of the same coin.
The opportunity is obvious: vast numbers of traditional software products, hardware products and industries have yet to be seriously reconstructed with AI.
Excel still operates in ways shaped decades ago. Business systems described in the source as being “in ROM” still require users to click through layer after layer of menus. Computers at many workstations run workflows inherited from an earlier era every day. Countless parts of this environment could be rewritten by Agents.
But the challenge is equally clear: foundation-model providers could absorb an application category almost as an afterthought.
Today they provide an API; tomorrow they may also release an official desktop assistant, workflow platform or operating-system plug-in. A startup may spend two years building in a vertical category, only to have its visibility overwhelmed by the name of an “official” product.
This is the real tension every AI founder must confront:
On one side is the natural pressure cast by the shadow of giants; on the other is the determination to keep solving real users' problems.
Ma Liang's approach is to move as close as possible to nuanced, sticky scenarios that cannot easily be solved by a uniform cloud template, and to make the integration of software and hardware, the accumulation of personal preferences and the human-computer interaction experience robust enough to endure.
VIII. Several small steps individuals can take over the next 1–3 years
Shifting the perspective from companies back to people, how should ordinary individuals position themselves over the next 1–3 years so they are not left behind?
Ma Liang's advice can be reduced to a few simple points:
Actively bring AI into your everyday work and learning processes.
When you encounter a problem, do not immediately follow your old habit of opening software, looking for someone or searching through materials. First ask yourself: is there an existing AI tool I could try?
If there is no ready-made tool, could you use AI to help write a small script or simple automation that solves the problem?
Try a wide range of new products. Even if they initially feel clumsy or slow, do not dismiss them too quickly—you are training your own “AI-use muscle.”
Once you have genuinely tried every tool you can find and concluded that none works as well as your old workflow, return confidently to the original method. At that point, you are not blindly rejecting something new; you are making a rational choice after comparison.
Sometimes, whether you qualify as an advanced individual is determined not by how many model concepts you understand, but by whether you are willing to take one additional step:
Let AI try first.
What is being reinvented is not just software, but the relationship between people + computers + AI
What Silicon Geek is building is not a simple desktop utility.
It is rewriting our relationship with computers: moving from “I operate you” to “we complete something together.”
AIdo and Desktop Companion are only the first generation of this form. As Agentic AI capabilities continue to improve, familiar devices around us—PC systems, phones, tablets and even small pieces of hardware on a desk—may quietly become partners with initiative of their own.
Viewed closely, the phrase “Pioneering Intelligence | The Individual Era” is a common question that this era poses to everyone:
Will you treat AI as another new software product, or as infrastructure for rearranging your life and reorganising how you work?
Ma Liang chose to begin again at the operating-system layer.
What about you? Which small habit will you start with, allowing AI to come a little closer?
Selected Interview Q&A
Q1: Please briefly introduce Silicon Geek and what you are building.
Ma Liang: Silicon Geek was founded less than six months ago. We define our mission as “using AI to reinvent familiar software and hardware,” transforming them into forms of human-computer interaction better suited to the AI era and finding new market opportunities in the process.
Q2: What is your flagship product, and how would you describe it in one sentence?
Ma Liang: Our flagship product is AIdo, a PC desktop Agent capable of calling hundreds of tools. You can think of it as “the AI-powered super-evolution of PC productivity software.” We hope it becomes users' productivity partner in their everyday computer use.
Q3: Which specific pain points does AIdo solve for users?
Ma Liang: There are two core points. First, it makes “scheduling decisions across a vast range of tools” on the user's behalf, eliminating the need to decide which software to use. Second, it builds a runtime environment on the PC that is “imperceptible to the user but controllable by AI,” allowing the intelligent agent to operate the computer and perform real work.
Q4: You also have companion hardware. What does it do?
Ma Liang: We have a desktop intelligent-assistant device called Desktop Companion. It can control a computer by voice to perform many kinds of tasks, making the Agent more natural to use and genuinely enabling work to begin as soon as you speak.
Q5: What personally motivated you to start this company?
Ma Liang: It is simple. In this great era of AI, I hope to create something new that genuinely changes how the world operates—allowing AI truly to benefit knowledge workers by reducing cognitive load and improving efficiency.
Q6: What is your long-term vision for the company?
Ma Liang: We hope to become a defining force in next-generation AIPC human-computer interaction, enabling everyone to have a dedicated and almost all-powerful AI partner.
Q7: In your view, what has been the most important recent breakthrough in generative AI?
Ma Liang: The greatest breakthrough is the improvement in the LLM's own capabilities and function call capabilities, together with the industry's broad adoption of protocols such as MCP and A2A. These developments give large models increasingly powerful capabilities for invoking external tools and move the field from simple AIGC towards Agentic AI.
Q8: How have you implemented Agentic AI in your product?
Ma Liang: We focus Agent capabilities at the PC operating-system layer and turn them into a productivity assistant such as AIdo. We then add a “personalised algorithm for choosing among a vast range of Tools,” helping users select the optimal approach automatically in multitasking and highly concurrent scenarios. With use, it gradually becomes an intelligent Agent assistant that “understands you better than you understand yourself.”
Q9: At such an early stage, how do you balance technical iteration with commercial execution?
Ma Liang: At this stage, we are more focused on refining the product. We iterate rapidly on a day-by-day basis while opening internal and public beta tests, identifying and solving problems through real use and gradually digging into the central pain points of vertical scenarios.
Q10: How do you view the current wave of Agentic AI? Is it the final form?
Ma Liang: I do not believe LLM-based Agentic AI is the final form. Today, the field is primarily exploring text-based logic and the ability to invoke code. A more complete Agent may model the lower-level “mechanisms that generate logic,” integrating complex neuronal mechanisms across multimodal signals including vision, hearing, taste, smell and touch.
Q11: What opportunities and challenges will AIdo encounter as it expands abroad?
Ma Liang: The opportunity is that China's supply chain can support an integrated software-and-hardware productivity assistant, a route where we have a relative advantage. The challenge is that many MCP-based tools differ greatly across overseas markets, especially in local-life and online-service use cases, so the system must be rebuilt for each context.
Q12: What role do you think Chinese AI companies currently play in the global AI market?
Ma Liang: Today, they mainly act as “open-source leaders, pioneers in application innovation and cost optimisers.” In the future, they will move deeper into original innovation and become originators of new paradigms. DeepSeek's extensive use of reinforcement learning during the RLHF stage and its proposal of contextual optical compression in OCR are examples of paradigm-level experiments.
Q13: The conference theme is “Pioneering Intelligence | The Individual Era.” How do you interpret it?
Ma Liang: AI will reduce the barrier to using many tools, but the productivity gap between “advanced individuals” and “lagging individuals” will actually widen, while lower- and mid-level roles will face greater disruption. As individuals, the only response is to embrace the change proactively, become “advanced individuals,” and use AI to amplify our creativity, productivity and commercial value.
Q14: Can you give several concrete examples of AIdo empowering individuals and small teams?
Ma Liang: If an image exceeds an upload limit, drag it to AIdo and say, “Compress this to below 1MB,” and it compresses it automatically. While editing a talking-head video, say, “Remove the silent sections,” and it cuts out the gaps automatically. When writing a Word document, drag it in and say, “Give each paragraph a suitable image,” and it automatically generates, adds and edits images in batches.
Q15: For AI creators in your field, what are the greatest opportunity and challenge?
Ma Liang: The opportunity is that vast amounts of traditional software, hardware and industry workflows have yet to be reinvented with AI. The challenge is whether foundation-model providers might directly absorb your category almost as an afterthought.
Q16: Over the next 1–3 years, how do you think individuals and small teams should position themselves?
Ma Liang: In everyday work and study, people should embrace AI as fully as possible and step outside their comfort zones. When a problem arises, first ask, “Is there an AI tool that can solve this?” Try more new products; if they truly do not work, return to the old workflow. Make AI part of your daily process rather than a new toy you use only occasionally.