Microsoft AI CEO: When the marginal cost of AI approaches zero, the distribution of power changes.
Recently, in Trevor Noah's podcast "What Now?", Mustafa Suleyman, the CEO of Microsoft AI, underwent a nearly two - hour in - depth interview and put forward a key judgment:
The marginal cost of AI is approaching zero, and the way power is distributed has changed.
This is not about the progress of AI technology, but about making AI accessible to everyone.
In the past, taking action required organization, processes, and teamwork. Now, an individual plus AI can complete the work that used to be done by an entire department. It can send emails on your behalf, generate contracts, build websites, modify code, and even negotiate with others on your behalf. What you gain is no longer just a tool, but a "right to act".
Mustafa's view is clear:
AI is evolving from a "prediction system" to an "action agent".
This means the decentralization of power: it's not that technology replaces humans, but that technology re - allocates the permissions of "who can do what".
Every prompt is quietly reshaping the boundaries of power.
The question about AI is no longer what it can do, but who can use it to achieve what.
Section 1 | AI, More Than Just a Simple Tool
Many AI experts like to talk about technological breakthroughs, but Mustafa Suleyman is different. This Microsoft AI CEO and former co - founder of DeepMind is more concerned about: who can use AI to actually accomplish something.
This isn't something he just figured out recently.
In 2011, Mustafa was still starting a business in London. He founded DeepMind with Demis Hassabis and Shane Legg. At the beginning, there were only three of them. They didn't want to create a tool, but an intelligence that could "understand the world".
At that time, mobile phones had just become popular, and the App Store had only been launched for three years. People thought what they were doing was too far - fetched. But Mustafa said he just believed:
"If we can create a technology that can truly understand human behavior, it could be one of the most powerful ways to improve the human condition."
More than a decade later, he hasn't changed.
In this interview, he repeatedly emphasized that the deepest change brought by AI is not being smarter, but truly having the ability to understand and create.
He gave an intuitive example:
"If it has seen a million images and knows where one eye is and where the other is, it won't draw the mouth on the forehead, but under the nose."
It sounds basic, but the meaning behind it is not simple. AI doesn't just recognize data; it can infer reasonable new results from existing data and generate them in its own way.
The biggest difference from traditional software is that it doesn't need to write specific programs for each situation, but can flexibly handle various new scenarios.
In his words:
"What we've created is neither a tool nor a human. It's a fourth kind of thing. It's not a button you control, nor an employee you hire."
This characteristic is precisely the core of 'AI power'.
AI doesn't just give people greater influence. It turns the 'ability to execute a task' into a digital resource that can be used and replicated at any time.
Whether it's sending emails, scheduling, preparing financial reports, or handling customer inquiries, AI no longer just helps you save time; it can directly 'take action'.
From the founding of DeepMind to the platform - based promotion of Microsoft AI, his view has remained the same:
AI is not a smarter advisor; it is an executor that can take actual action.
When AI starts to take actual action, we need to judge it just like we judge people. Humans judge each other's credibility through actions. We can't see a person's heart, but we can see what they say and do.
Mustafa said the same applies to AI.
When it is put into applications, connected to email, web pages, and interfaces, it is no longer just a simple information assistant, but an action agent.
Section 2 | One Person Can Do What a Team Used to Do
Mustafa mentioned a change that sounds like a technical topic but is actually very practical:
The marginal cost of AI is approaching zero.
What does it mean? It means that using AI once hardly costs any extra money.
In the past, if you wanted to send an email, you had to type it yourself; if you wanted to make a spreadsheet, you had to use Excel to type slowly; if you wanted someone to do these things for you, you had to hire a team or add a position.
But now, using AI to do these things costs so little that it can almost be ignored. Even, if you write a sentence, AI can automatically perform ten steps of operations.
Mustafa's exact words were very straightforward:
Any individual can have the influence of a large - scale company.
It used to sound unbelievable, but not anymore.
AI can search for information, reply to customers, schedule, write copy, and package processes on your behalf. It doesn't just assist you; it helps you complete a whole set of tasks.
Mustafa's example was also very clear: you can tell AI an idea, and it can turn that idea into a piece of code, generate a web page, publish it, and even promote it automatically.
In the past, this set of actions required a product manager, a designer, two engineers, and market operations to cooperate. Now, one person plus AI can complete the process.
You no longer wait for others to help you; you directly tell AI to do it.
The decrease in marginal cost means: the limitations of manpower and funds originally required from idea to implementation are disappearing.
Just like in the era of mobile Internet, you no longer need to find a newspaper or create a column to publish an article; you just open a social platform. In the AI era, you no longer need to apply for resources, start a process, or find a team. With just one sentence, AI can execute.
Mustafa emphasized the meaning behind this:
"It's not that organizations are getting stronger, but that individuals' practical abilities are increasing. AI is turning 'executive power' into a resource that everyone can obtain."
This is the new starting point:
You no longer need to rely on a company or bear high costs. AI gives you the ability of a team.
Section 3 | Those Who Can Use AI Have Executive Power
During this conversation, what Mustafa talked about the most was not how powerful AI is, but what you can use it to do.
He said:
Prediction is power.
The ability to predict accurately and execute is the essence of power.
So, who held this power in the past?
In the past, who could take action, who could publish, and who could execute were controlled by large companies and platforms. But the arrival of AI has simplified the originally complex operations, making it easier for ordinary people to start doing things.
For example:
You type a sentence on the AI interface: "Help me translate this paragraph into English and organize it into a customer email." AI can automatically recognize, generate, format, and even consider the tone and wording carefully.
He described it like this:
"Without complex operations, with just one sentence, it can immediately execute the task."
Moreover, this kind of action is not single - sided; it can be carried out in batches and infinitely replicated.
It suddenly increases what one person can do:
If you want to run a marketing campaign, AI can write advertising slogans, generate posters, and analyze the audience;
If you want to launch a new product, AI can complete the registration process, generate graphics and texts, and arrange a promotion plan;
Even if you want an AI to "help you arrange AI", it's possible.
For all these tasks, you just need to type a few lines of text and click a few times.
As this ability becomes more widespread, the power structure that used to rely on resource integration, process approval, and organizational deployment to get things done is loosening.
The new power belongs to those who can correctly use AI to turn ideas into reality.
More importantly, this power will continue to increase. Mustafa predicted in the interview:
"AI will soon have perfect memory and be able to make continuous and accurate predictions for you over a long time sequence based on the interaction experience with you. What you see today is a one - time prediction engine. In the future, you will see an intelligent system that can truly remember you for a long time, understand your rhythm, and take the initiative to execute."
This intelligent system is not a robot, but the AI in your phone, email, and web pages.
It no longer waits for you to click to take a step. Before you even speak, it has already prepared the plan, execution, and feedback.
When AI has this ability, the definition of power is being reshaped:
In the past, to see if a person had power, we looked at how many resources they could control;
In the future, to see if a person has power, we will look at how many AIs they can mobilize to complete actions for them.
This is not just a conceptual change, but a complete rewrite of the way of working: the boundaries between positions are blurring, individual functions are expanding, and traditional divisions of labor are being shuffled.
Section 4 | Where Is the Uncontrollable Point of AI?
In the previous three sections, we talked about how the capabilities of AI have changed, the cost has decreased, and the power distribution has also changed. From platforms to organizations to individuals, the threshold for using AI is constantly decreasing.
But the question arises:
When everyone can make AI do things, where are the boundaries of AI itself?
Mustafa Suleyman put forward a core concept:
What we need to build is Humanist Superintelligence. This kind of AI cannot have motives, cannot persuade humans, and cannot make autonomous decisions.
In other words, the real risk doesn't lie in how smart AI is, but in whether it has "autonomous ideas" and whether it will act on its own without authorization.
These are the four technical red lines proposed by Mustafa:
- Can it modify its own code? —— If it can, it may bypass the design constraints of developers.
- Can it set its own goals? —— If it can, it will no longer just execute human intentions.
- Can it actively obtain external resources? —— If it can, it has the ability to act independently.
- Can it take autonomous actions? —— If it can, it is close to an independent behavioral subject.
Mustafa emphasized: Once these red lines are crossed, what we face is no longer a model or a tool, but an action agent with behavioral intentions.
To deal with this potential risk, he mentioned a key governance concept: Containment.
"Containment doesn't mean making AI go back or stop. It means ensuring that it always operates within clear boundaries."
This means that every time an AI ability is released, the following need to be established simultaneously:
- Permission mechanism: Who can call it and to what extent?
- Accountability mechanism: Who is responsible if something goes wrong?
- Technical encapsulation: Is the high - level control interface of the model sealed?
- Behavior isolation: Can the model be prevented from overstepping its authority to connect to other systems?
Mustafa clearly pointed out that the risk of AI, different from any previous technology, doesn't occur at the "moment of production", but at the "moment of activation".
What's more troublesome is that when these models are continuously integrated into office software, browsers, mobile apps, and even cloud APIs, many people don't even know what - level capabilities the model they are using has.
This is not a "model risk", but an "activation risk".
You think you're using a writing tool, but actually you're calling an AI agent with great decision - making power.
The real control point doesn't lie in the model parameters, but in the design of the technical boundaries. If these basic boundaries are not clearly established now, they are likely to be broken through by disordered use in the future.
This is not a governance problem to be discussed in the future.
It is a boundary that must be clearly defined before AI enters our lives.
Conclusion | Who Gets the Ability? AI Gives You an Option
In the past 20 years, the main theme of technological progress has been "centralization": centralizing computing power, data, and resources to achieve great things through scale advantages.
Now, the marginal cost of AI is approaching zero, and everyone can obtain unprecedented action ability.
Mustafa's words revealed this turning point:
"It's not just a tool, but a callable executive power."
This is the redistribution of power. AI is returning the ability to do things from platforms to organizations and from organizations to individuals.
When the ability is returned to individuals, more people get room to play, rather than making a few people more powerful.
Perhaps, this is the most anticipated future of AI:
Not to replace people, but to enable ordinary people to accomplish things.
📮Reference Materials
https://www.youtube.com/watch?v=tQ5wO1lznCQ
https://podcasts.apple.com/us/podcast/will - ai - save - humanity - or - end - it - with - mustafa - suleyman/id1710609544?i = 1000727247919
https://www.businessinsider.com/microsoft - ai - ceo - mustafa - suleyman - economics - of - information - radically - change - 2024 - 6
https://som.yale.edu/sites/default/files/2023 - 06/Yale%20June%202023%20CEO%20Summit_Key%20Themes_v7.pdf?utm
https://www.madrona.com/ia - summit - 2024 - mustafa - suleyman - from - saas - to - agents/?utm_source = chatgpt.com
This article is from the WeChat public account "AI Deep Researcher", author: AI Deep Researcher, published by 36Kr with authorization.