奥尔特曼谈“单人公司”如何改写创业规则

These days, there's talk about a one-person company.

Some people regard it as a new concept.

Some see it as entrepreneurial inspiration.

Others understand it as: One person plus a few models can make a company run.

But a recent statement by Sam Altman at Stanford might be closer to the truth: In the future, companies will have fewer people but manage more objects.

In the past, what was managed were people. Now, it's AI.

To manage people, you need to recruit. To manage AI, you need computing power.

So, the first question discussed by startup companies has changed from "How many people to recruit" to "How much computing power can be obtained". Some engineers are running a dozen agents simultaneously, making them complete different tasks respectively. Humans are responsible for allocation and supervision, while AI is responsible for execution.

Sam predicts that the next important moment like ChatGPT might be an agent that can work continuously for days or even weeks.

Section 1 | AI Starts to Change from a Tool to a Colleague

If AI is only regarded as a tool, many changes will go unnoticed.

In the past few years, most people used AI in a rather simple way: writing a piece of code, summarizing an article, or answering a question. It's more like an upgraded version of a search engine.

But in this conversation at Stanford, Sam mentioned a transformation several times: AI is no longer just answering questions; it has started to take over tasks.

In the programming field, it's now possible to trust an AI software engineer to complete some small tasks that last for a few hours. For example, fixing a function, modifying a program, or writing a simple module. This was hard to imagine a year or two ago.

And the following changes are:

A few hours → a few days → a few weeks.

The duration that AI can work continuously will become longer and longer, which means its role is changing. In the past, it was more like an on - call assistant, but now it's more like a colleague who can work independently.

Sam mentioned an increasingly common scenario: Some developers are running multiple agents simultaneously, making them complete different tasks respectively. Humans no longer do everything by hand at every step but are responsible for arranging work, just like a project leader assigning tasks to different members and then integrating the results.

When AI changes from "answering one question at a time" to "being able to complete tasks continuously", the way of one person working with a group of AIs starts to become possible.

Section 2 | The Structure of Startup Companies Starts to Change

In the past, when many startup companies were just established, the first thing they discussed was: How many people to recruit. How many engineers, how many product managers, and when to form the marketing team. The company's scale was almost equal to the team's scale.

But now, more and more entrepreneurs are discussing another thing: How much computing power can be obtained.

Sam said that the thinking of the new generation of startup companies is changing. Many founders are no longer in a hurry to expand the team because the larger the team, the higher the communication and management costs. AI tools, on the contrary, have magnified individual capabilities.

The amount of work a programmer can complete today is completely different from that a year ago. The reason is simple: Many tasks that originally needed to be done by humans can now be partially completed by AI. For example, writing basic code, checking program problems, and organizing instruction documents.

So, a company doesn't necessarily need such a complete team to develop a product.

In the past, engineers, designers, and operators might have had their own specific duties. Now, a small team or a founder with a few AIs can get many things done.

The basic composition of a company is changing from "a group of people" to "a few people + a group of AIs".

Section 3 | When AI Becomes an Employee, the Management Method Also Needs to Change

If AI is just a tool, things are actually quite simple. You turn it on when you need it and turn it off when you're done.

But when AI starts to take over more and more complete tasks, the situation is different.

Sam himself is an example. When he has a new idea now, he will first assign the task to AI. For example, analyzing a new business model or evaluating a product direction. He asks AI to help analyze possible paths and potential problems, and then makes a judgment based on AI's output to decide which ideas are worth discussing with the team.

This management method is relatively simple because it only involves one AI. But when multiple AIs need to be managed simultaneously, the complexity will increase significantly.

Those who are best at using AI tools often run a dozen AIs simultaneously, each responsible for a different task. They told Sam that the work is actually more difficult now.

Because managing these AIs requires dealing with many new problems.

• How to break down tasks clearly enough for them to understand,
• How to judge whether the results they give are reliable,
• When to trust the output and when to intervene manually,
• And also think about how to provide them with enough context.

The context includes the company's internal documents, product information, customer feedback, and historical project records. The more complete the information, the more an AI's performance will be like that of an experienced colleague.

So, the core of managing AI lies in clearly arranging tasks, supervising the execution process, judging the quality of the output, and retaining the final decision - making power. Humans are becoming more like project leaders, while AI is the executor.

The scope of work that one person can manage will also be much larger than before.

Section 4 | Why This Change Appears Suddenly

Many people may think that the scenario of one person working with a group of AIs sounds like a distant future.

But Sam mentioned in the conversation that this change is starting to appear because two conditions are maturing simultaneously.

First, the capabilities of AI are rapidly improving

In the past few years, many people only regarded AI as a chatting tool. But inside OpenAI, the models have started to be able to complete more and more complex tasks. Sam recalled that when they first had a conversation with the early language model, they realized that the computer was starting to do things that had never been seen before. Later, as the models were continuously upgraded, this ability became more and more obvious.

Second, the cost is rapidly decreasing

Sam mentioned a figure: From OpenAI's earliest inference model to the current new model, the cost of completing the same complex task has decreased by about 1000 times. Technologies that were only affordable for a few companies in the past can now be used by more and more individuals.

When the capabilities are improving and the cost is decreasing, a new way of working will naturally emerge.

Sam used electricity as an analogy. When electricity first appeared, only a few factories and institutions could use it. But as the infrastructure gradually improved, electricity became a part of daily life, and people no longer thought about where the electricity came from and just used it.

AI will also reach a similar stage.

When intelligence becomes easier to obtain, one person will naturally be able to accomplish more things.

Conclusion

Sam mentioned at Stanford that the last such era was when the iPhone first came out. Since then, there hasn't been one for a long time.

Now it's here again.

The basic composition of companies is changing. The threshold for entrepreneurship is lowering. The objects of management are changing from people to AI.

A one - person company is no longer just a concept.

Original Article Links:

https://www.youtube.com/watch?v=sTnl8O_BuuE&t=1066s

https://www.youtube.com/watch?v=FjlymGBt-vY&t=1345s

https://stanforddaily.com/2026/02/15/sam-altman-agi-treehacks-keynote/

https://www.rev.com/transcripts/altman-speaks-at-blackrocks-us-infrastructure-summit

Source: Official Media/Online News

This article is from the WeChat official account "AI Deep Researcher", author: AI Deep Researcher, editor: Shen Si. Republished by 36Kr with permission.