In this video, Gary Tan explains that AI is going beyond a simple coding aid and is becoming a new workforce that is changing the way the entire company works and its organizational structure. The key is not to choose a better model, but to connect tasks well with skill files, organization charts, processes, and memory systems so that AI can execute them repeatedly. He emphasizes that the winners in the future will not be companies that use AI a little, but companies that were designed natively in AI from the beginning and organizations that have a 'corporate brain'.

1. After the AI revolution, an era where one person can do the work of a thousand people

The presentation begins by asking the audience, "What should we make now?" Gary Tan describes himself as a founder, investor, and person who runs Y Combinator, a 20-year-old organization. And now YC itself is changing into an AI native organization, and the process of an old organization changing in a new way is "strange and surprising," he said.

The company he wants to create at YC is completely different from the past. The goal is to create an organization where one person can now accomplish tasks that previously had to be done by hundreds or thousands of people.

"One person is building a company that does what used to take a thousand people. This is not a metaphor."

He urges the audience, who will soon enter the entrepreneurship battlefield, to start with the knowledge that what is possible today is much greater than people think. This means that the moment you view AI simply as a convenient tool, you miss out on the scale of productivity that is actually possible.

2. The secret to 400x productivity is not a model, but a 'way of connecting work'

Gary Tan explains the explosive increase in AI productivity based on his own experience. When he was creating YC's internal social network in 2013, he worked almost like a full-time engineer, writing about 14 lines of actually usable code per day, he says. It was normal at the time.

However, as of 2026, he runs YC full-time, works fewer hours than before, and has to pick up his child at 5 p.m. Nevertheless, when he calculated his output, he found that it was about 400 times compared to the past.

"It's the same person and same day, but the actual output is about 400 times higher."

Of course, he acknowledges that these figures may be considered exaggerated. Let's assume that the AI ​​agent may have written a lot of code unnecessarily, half of which may be skeleton or auxiliary code. He explains that even if his calculations are as unfavorable as possible, the increase in productivity is at least 8 times, and the median is 80 times.

"No matter how hard you try to calculate it, the number is big."

But what he really wants to emphasize is not the number 400 itself. The important thing is that productivity is not all about which model you use.

"The person who is twice as productive and the person who is 100 times more productive use the exact same Claude: same weights, same context window, same API."

So where does the difference come from?

"Leverage is not in the weights of the model. It's in how you connect things."

YC's case also supports this. By the winter 2025 batch, about a quarter of participating companies were already generating 95% of their codebase with AI. And that deployment, he says, became the fastest-growing and most profitable deployment in YC history. Although we cannot say for sure that AI is the direct cause of growth, we emphasize that it is clear that the fastest-growing startups do not use AI as an autocomplete feature.

"Fast-growing startups don't treat AI as autocomplete. They treat AI as a workforce."

3. Blueprint for an organization that operates AI like employees

Gary Tan explains what it specifically means to "connect things." His core argument is that the way software is created with AI agents is almost identical to the structure of a business organization.

Skill files are employees

A skill file organized by an AI agent to perform a specific task is equivalent to one employee. If one ability and one role are clearly recorded, AI can perform that task repeatedly.

"One skills file equals one employee."

For example, if you need to modify a test file, you can write down the rules and procedures necessary for that task in a file like tests.md. When a task comes in, AI determines which files should be loaded and carries out the work according to established procedures.

The resolver table is an organizational chart

The resolver table, which finds and connects work-related files, plays a similar role to an organization chart. When a task comes in, we decide who should be responsible for it and what knowledge and procedures should be brought in.

"When work comes in, the resolver decides who will handle it and where to send it."

Filing rules are an internal process

Filing rules, which determine where files are stored, how they are linked, and what information is prioritized, correspond to a company's internal operating procedures. This means that the path through which work flows in an organization and the information management rules are implemented within the AI ​​system.

Trigger evaluation and testing are personnel evaluations

Automated testing to ensure that the AI actually retrieves the necessary knowledge and procedures is like a performance evaluation. For example, when you need to modify the test file, you can check whether tests.md is indeed loaded automatically.

"Performance evaluation is a test to check whether test.md is actually loaded when modifying the test file."

Ultimately, the roles, organization chart, work processes, and evaluation system that make up a company can all be expressed in Markdown files and code.

"Every element of an organization that used to employ a thousand people can now be expressed in Markdown files and other types of Markdown files."

Gary Tan says that until now, people have created organizations, but there has been no operational/management layer to manage AI. Now, using Claude Code or Codex is not just about writing code; it's about hiring, training, and managing a workforce comprised of Markdown.

"When you sit down in front of Claude Code or Codex, you are not writing software. You are hiring, training, and managing a workforce of Markdown."

4. AI native company that creates explosive sales with a small number of people

This method is not limited to theory. In YC, there are already companies that have made AI the basic operating method of their organizations.

Gary Tan introduces that Emergence, an AI app builder company, has reached 9-digit annual recurring revenue, or ARR, within 8 months of its public launch. When it achieved $15 million in sales, it had only 15 employees. Another company, Retail, says it has about 40 employees and $60 million in sales.

"This level of revenue per employee has never existed before – not in software, not in oil, not in railways."

He emphasizes that these companies are not naturally occurring freak success stories. These are companies built from the ground up based on new rules of productivity.

AI native companies do not deploy hundreds of people in sales, customer support, operations, or finance departments. Instead, each task is recorded as skills and procedures, and an agent executes them. And engineers handle tasks that need to be done directly by humans, while maintaining and improving AI skills so that they can do more work.

"AI-native companies record sales, support, operations, and finance tasks as skills, and then have agents execute those skills."

This method can also be applied to complex and specialized tasks such as tax reporting. If the necessary procedures and rules are created as a skill file, AI can perform actual tasks.

"If you have the skills file to file your taxes, AI can actually file your taxes too."

Even within YC, non-engineer employees are creating skill files and automation tasks. Media representatives, event operators, and finance team members also manage AI agents. One finance manager consolidated about 100 Excel files into one internal application she created herself. Although she is not a programmer, she is now an agent manager.

"Now everyone at YC is an agent's manager."

Therefore, the reason why YC can handle a large scale of work with fewer people than other organizations is not because people simply work harder. This is because the form of the organization itself is different.

"The key is not just the engineers being 400 times faster. It's about the entire company operating at a level 400 times higher than everyone else."

5. Separate what AI is good at and what code is good at

Gary Tan explains that many of the recurring problems in AI engineering are caused by placing calculations in the wrong space. He divides it into latent space and deterministic space.

Latent Space: Human Judgment and Ambiguous Problems

Latent space is the area where real large-scale language models operate. The following is relevant here:

• Understanding people's intentions
• Interpreting ambiguous requirements
• Work that requires taste and judgment
• Understanding problems expressed in natural language
• Non-deterministic tasks with no fixed answer

The AI model handles these issues by referring to the rules and context written in the Markdown file.

Deterministic Spaces: Precise Computations and Iterative Operations

On the other hand, deterministic spaces are an area where traditional software and code excel. Things that require consistent results, such as data storage, numerical calculations, sorting according to rules, and accurate seating arrangements, are better handled with TypeScript or another programming language.

Gary Tan gives an example of the problem of matching the thousands of participants at YC's Startup School. AI can understand each person's tendencies and the meaning of each meeting, but deterministic calculations are needed to determine which seats to actually place 800 people in.

In the past, the information of 800 people had to be printed out on paper and placed manually in a large space, but now, with a combination of AI and code, it can be processed in about 10 minutes and with hundreds of dollars worth of tokens.

"What previously would have taken a month may now be possible with a few hundred dollars in tokens and 10 minutes."

The key is not to leave everything to AI. Separating the judgments that AI must make and the calculations that code must make. Making this distinction incorrectly can slow down the system, increase errors, or fill the model's context window with unnecessary data.

6. Seven memories of humans and one million tokens of AI

Gary Tan explains the strengths of AI by comparing them to human working memory. Humans can keep about seven pieces of information in their minds at a time. There are usually seven to nine at the most, which is why phone numbers usually have seven digits or why you forget the eighth item when you go grocery shopping.

Checklists, organizational charts, filing cabinets, and databases created by mankind were all devices to compensate for this limited memory.

"All the checklists, organizational charts, and filing cabinets created by mankind are like prosthetic legs to compensate for the limitations of human memory."

However, an AI agent can handle about one million tokens, or roughly 1,000 pages of information, at once. Gary Tan explains this to his 10-year-old, saying that the AI ​​is like opening three Harry Potter books and reading them at the same time.

"An AI agent can put three Harry Potter books open in its head, find the needle in them, and put together the contents of the three books in a matter of seconds."

Human seven-digit memory and AI's three books create completely different operating environments. However, AI's memory is not infinite. This is because the information on the entire company is much more than the three Harry Potter books.

A company is a huge library where a lot of knowledge is accumulated, including emails, meetings, decisions, customer conversations, and post-mortem analysis. Therefore, whether the AI ​​behaves like a smart colleague or like a goldfish that only remembers what it has just heard depends on what information is shown to the AI.

"Your company is not three books. Your company is a library."

"The question that determines whether an agent becomes a genius or a goldfish is who decides which three books will be open on that desk."

This is what Gary Tan calls context engineering.

7. The brain of the company is its library and librarian

Gary Tan says that a company's AI system should not be understood as just a search function or RAG. RAG is just a basic technology for finding the documents you need. The real challenge is deciding what information to store, how to link it, and which information to elevate to high-priority memory.

"Search is easy. The product is creating knowledge worth searching."

The brain of a company is not just a data repository. It must have the following features:

• Decide what knowledge to record from the beginning
• Connecting and strengthening relationships between knowledge
• Manage frequently needed information with 'hot memory'
• Classify old reference materials as 'cold references'
• Decide which side is correct when there is conflicting information
• Select the most appropriate materials needed for a specific task and provide them to agents

So the brain of the company is a combination of a library and a librarian.

"The brain of the company is both a library and a librarian."

Gary Tan introduces GBrain, a system he is building publicly. This can be used with a variety of agent environments, and is described as a search and memory layer that allows agents to select the "three books" needed for a specific task.

His personal system stores approximately 220,000 pages including emails, meeting details, 20 years of records, and knowledge written by agents. When an entrepreneur emails a crisis, an agent comes looking for the next piece of information before he can even finish reading the message.

• Past conversations with the founder
• Other portfolio companies that experienced similar problems
• Solutions that actually worked at the time
• Relevant context that Gary Tan already knows

"Whatever an agent does, she starts by knowing what I already know. That's the difference between an assistant and a colleague."

However, there are risks to the company brain. An unmanaged knowledge repository can become a dump with only search capabilities. Old information can be retrieved as if it were up-to-date, and a bad skills file can cause a bad process to repeat itself forever.

Therefore, the corporate brain needs:

• Provenance, which records the source and time of creation of all information
• Contradiction check to check when new information conflicts with existing information
• Human and AI joint librarians that delete and organize old knowledge
• A system that is continuously managed like an actual operating infrastructure

"If you treat your brain like an operational infrastructure, knowledge builds like compound interest. If you treat it like a garbage dump, you get an agent who is certain of the wrong answer in a way that no one can trace."

8. Eliminate 'one-off tasks' and make everything a skill

The practice principle that Gary Tan most strongly emphasizes is Never do one-off work. If you ask an AI agent to do something and you like the results, you shouldn't stop there. The work method must be turned into a reusable skill file.

At first, the agent may be clumsy and the results may not be perfect. However, users can improve the results by telling them, "I don't like this, so change it." And if you are satisfied with the final result, you should save the process as a skill.

"If you are satisfied with the work once, you must not stop there and turn it into a skill."

"If you have to ask for something twice, you've already failed."

This principle changes not only individual productivity but the learning ability of the entire company. This is because, instead of explaining the same tasks from scratch to new employees every time, the procedures once learned can be accumulated as company knowledge.

"When organizations capture what they learn in this way, they get smarter every day. Organizations that don't, no matter how good their models are, wake up every morning with amnesia."

The performance of the AI ​​model is closer to a continuously borrowed resource, but the skills and knowledge accumulated directly by the company remain as an asset to the organization.

"The quality of the model is borrowed, but if you build the brain, it's yours."

9. What we need to create now is not a company that uses AI, but an AI native company.

Gary Tan clearly answers the question, "What should we create now?" Rather than adding a few AI functions to an existing company, we need to create a company where AI becomes the basic structure of the organization from the beginning.

The AI ​​native company he proposes is as follows.

• Thin manpower structure
• Skills file that defines all repetitive tasks
• Company Library to find knowledge suitable for your job
• Company brain that accumulates information and experience
• Founder who still has a deep understanding of the product and code
• All members who manage and improve AI agents

"Rather than a company that just uses AI, create a company designed natively to AI from the beginning."

He says you can use GBrain, which he is building, but you don't have to rely on any specific tools. The analogy is that Open Claw is a powerful tool like a Ferrari, and Codex is like a great Honda that can accomplish most tasks.

What is important is not the specific repository or product, but the following principles:

1. Identify where calculations should occur.
2. Use skill files like employees.
3. Create a knowledge management layer that acts as a librarian.
4. Transform one-time tasks into repeatable capabilities.
5. Accumulate company knowledge from the first week.

He predicts that every company will soon need a corporate brain. Personal AI that understands individuals, a knowledge layer that manages company memories, and a librarian that selects necessary information could become a huge new market.

"Every company in the world will soon need a brain."

He says he wants to keep GBrain's core layer open source like Linux, and hopes to see a founder emerge who will create a decisive company in this space.

10. Abundance is not a policy document, but actually released software

In the last part, Gary Tan addresses people's fears that jobs may be lost to AI. He understands the anxiety, but says the problem is the inability to imagine the possibilities ahead.

"It's a failure of imagination."

The role he expects from his audience is to use AI to double, tens, or hundreds of times the capabilities of themselves and their organizations. And he says that companies created in that way should become lighthouses that show society how to use AI.

"Affluence is not a policy paper. Abundance is actually released software."

He describes the case of a father who created a corporate brain consisting of 80,000 Markdown files for his son, who suffers from a rare form of epilepsy. The story is that without a research institute, government funding, or permission, one person pushed the limits of human knowledge about his son's disease with just a laptop and a library of knowledge.

"There was no research institute, no grants, no permits. There was just one father, a laptop, and a library."

To him, this case is not just a touching anecdote. Libraries and librarians who retrieve three volumes of correct information at the moment of need are examples of AI native structures that actually change people's lives.

Now, there is no need to think like before, "I want to know that person, but I can't meet him." Codebases that were too complex to be touched can be repaired, and archives that were too large to be read can be analyzed. It can also handle datasets that were too messy to clean up.

"Every record that was too big to read, every data that was too complex to organize, every ocean that we were told to 'never boil,' we can now boil."

Conclusion: Create a company, and create the brain of the company underneath it.

Gary Tan's final message is simple. What entrepreneurs need to create in the future is not simply products that utilize AI. It is the foundation of the AI ​​native company itself and the memory, knowledge, and skills that continue to make the company smarter.

"Build an AI native company, and build what's underneath: the brain, the memory, the library that builds like compound interest."

He urges the audience to log their work, turn it into a skill, write tests, and release it as real software.

"Boil the ocean. Write a test. Release the skill."

Ultimately, competitiveness in the new era does not lie in monopolizing the best AI model. It depends on how well the organization stores its knowledge, retrieves it when needed, and converts what has been done once into the ability to use it again. Thin, fast organizations with well-connected AI will operate on a completely different physics of productivity than the giant corporations of the past.