Garry Tan Declares the Era of Personal AGI, Urging Founders to “Own Your Intelligence”

Garry Tan began his lecture by quoting the 17th-century philosopher Spinoza. Spinoza earned a living by grinding lenses and devoted himself to philosophical studies, but his ideas clashed with the existing religious authorities. The Dutch government offered Spinoza a pension and a professorship in exchange for stopping his writings, but he firmly refused, choosing to pursue truth. Tan compared this to the modern entrepreneurial spirit, emphasizing that when founders are offered salaries to stop developing, they should pursue truth instead of comfort.

Tan pointed out that 400 years later, humanity is repeating a similar mistake regarding intelligence. Many people are waiting for AGI as a single, massive event, like a god in a data center, a specific announcement, or the day the sky changes color. He argued that this perspective is wrong, declaring an updated Spinozist heresy 400 years later, and emphasized the importance of personal AGI running on individual infrastructure.

Garry Tan criticized the term 'personal AI' as being misdefined by marketing departments and clarified the true meaning of what he calls 'Personal AGI.' He explained that it is not a $20/month chatbot, a simple autocomplete feature, or an assistant that only knows a user's schedule. He pointed out that such services are merely subscription models that users borrow, which reset when a tab is closed and can easily change if the company shifts direction, categorizing them as corporate AGI. The Personal AGI Tan refers to means an agent that the user directly owns, accumulates knowledge over time, and grows according to the user's intentions.

Garry Tan illustrated the productivity revolution brought about by AI agents with a personal example. He runs YC full-time and stated that his output has increased by approximately 400 times compared to 2013, under the same conditions of brainpower, time, and needing to pick up his child at 5 PM. He added that this dramatic increase in productivity is the result of integrating AI agents into his personal infrastructure.

YC's portfolio companies are also achieving remarkable results through the use of AI agents. Tan revealed that a year and a half ago, in the Winter 2025 batch, a quarter of companies used 95% AI-generated codebases, and these companies are using AI agents for all their work, not just code. He predicted that this batch would be one of the fastest-growing and most profitable batches in YC history.

Garry Tan stated that while he cannot definitively say AI caused the growth, there is a clear correlation among the fastest-growing founders YC invests in. These founders are actively utilizing AI not just as a simple autocomplete tool, but as 'personnel.' He explained that this approach plays a crucial role in enabling founders to handle work at a scale previously impossible and develop innovative solutions. He conveyed the message that treating AI not just as a tool, but as a team member to maximize its potential, is key to success.

In the past, one piece of startup advice was to 'do things that don't scale.' However, Garry Tan explained that with the advent of AI agents, the meaning of this advice has fundamentally changed. Now, agents provide a multiplier effect, allowing a single founder to handle unscalable tasks at a large scale. He emphasized that while the advice itself hasn't changed, the physical aspects of every startup and the scope of what a founder can actually do have changed dramatically.

Tan explained that while human working memory can handle about 7 pieces of information at a time, an AI agent can simultaneously process a million tokens, or about a thousand pages of information. He analogized this to having three Harry Potter books open in one's mind at once, with the ability to find and synthesize necessary information within seconds. He pointed out that while one might argue it's not AGI yet, it is already a different operating system, and almost everyone on Earth is still living with organizational charts and methods designed for a 7-digit brain, emphasizing the gap between AI's overwhelming information processing capability and human's limited cognitive capacity.

Garry Tan unveiled his personal OpenClaw system, which he calls the 'G Brain.' This system is a Karpathy-style knowledge wiki consisting of approximately 220,000 markdown pages, chronicling 25 years of his life in diary format. It includes all emails, meeting minutes, personal notes, photos, drafts, and even experiences where he was wrong. Most of this information is compiled, curated, and searched by the agent. Tan explained that thanks to this system, he no longer asks questions he's already answered, and if a founder emails him about a crisis, his agent can pull up all previous conversations to provide context before he even finishes reading the email. He emphasized that this accumulated experience is the core of the system.

Tan declared that markdown files are now actual code, and anyone who can write clear English instructions is a programmer. He emphasized that language models act as compilers, so it is no longer solely the domain of engineers. At Y Combinator, people who have never opened a terminal in their lives—such as media staff, event staff, and finance teams—are building skill files and scheduled tasks. He mentioned that one finance person compiled about a hundred Excel spreadsheets into a single app built with an internal agent, and she became an 'agent manager' rather than a programmer, predicting that everyone will become one.

Garry Tan pointed out that the most significant cause of AI agent failure is confusing the 'latent space' where calculations are made with the 'deterministic space.' Latent space involves areas like taste, judgment, and interpreting what humans truly want from ambiguous requests, existing within the model and controlled by markdown files. Deterministic space, on the other hand, includes precise and specific calculations such as arithmetic, SQL queries, and seating charts, all of which should be stored in the SQL database that markdown files use. He explained that agents malfunction when the roles of these two spaces are not clearly distinguished.

Garry Tan cited his work on a Spinoza biography as an example of AI agents' powerful information processing and summarizing capabilities. Five days before his lecture, when he decided he needed Spinoza, he instructed his agent to acquire three biographies of Spinoza by Nadler, Goldstein, and Stewart—about 1,500 pages in total. Overnight, the agent read all three books and summarized them, created a chronology of Spinoza’s life, noted differences in opinion among the three biographers, and compiled chapter-by-chapter quotes and the best quotes. Furthermore, knowing Tan needed presentation materials, it ranked the 10 most narrative-worthy moments of Spinoza's life with accompanying presentation notes. All the content that opened the lecture—the knife incident, the bribe, the desk—came from this work. Tan called this a 'compendium skill' that he uses daily, emphasizing it as a personal mega-version of in-depth research, more profound than anything offered by corporate AI products.

Garry Tan revealed that AI agents are giving rise to companies that are breaking existing equations for corporate growth. Emergent, one of YC's Summer 2024 batch companies, achieved a 9-figure (hundreds of millions) revenue within just 8 months of public launch. He added that when the company exceeded $15 million in annual revenue, it had only 15 employees. This is a concrete example demonstrating that small teams can achieve immense scale with the help of AI agents.

Tan asserted that utilizing AI agents is no longer a distant future prospect but the current standard for competition. He stated that current batch companies are already operating in this manner, and if founders do not leverage AI agents, their competitors will politely take their lunch. This implies that the adoption of AI agents has become a prerequisite for startup success and an immediate competitive advantage.

That is not the future, that is the bar right now with this batch. If you're not doing it, your competitor is and they will eat your lunch politely and thank you for it.

Garry Tan emphasized that the knowledge base, which can be thought of as the 'brain' of an AI agent, requires continuous management. He analogized an uncurated brain to a garbage dump with excellent search capabilities. Search functions can present outdated facts with perfect confidence, and bad skill files can perpetuate bad processes indefinitely. Therefore, he explained, roles like a librarian, responsible for memory and hygiene, verifying the source of all facts, checking for inconsistencies when new information conflicts with existing information, and pruning, are essential.

Garry Tan presented a 5-stage roadmap for building and utilizing AI agents, along with changes to expect over the next 90 days. In the first week, the agent will still be immature with a thin library, and more time will be spent correcting than saving. By the fourth week, the flywheel will start turning: the agent will answer questions tailored to personal context, morning tasks will produce actually readable output, and the first two skill files will work successfully, leading to writing the third and fourth. By week 12, one will have a library that answers questions before they are even finished, dozens of skill files handling weekly tasks that were usually dreaded, and one or two tools that others will ask to borrow. He defined this state as a 'startup' and encouraged listeners to build their own personal AGI startup within 90 days.

Garry Tan strongly advocated for the principle that 'skill files must be yours.' He warned that if you do not own your skill files, your job itself will become a skill file. He analogized this to artisans in the past who owned their tools and were free, whereas in the factory era, that freedom was broken when looms became factory property. Knowledge workers believed they were safe because their tools were in their minds, but the emergence of skill files ends this perception, he explained. He emphasized that for the first time in history, an individual's cognitive abilities can be extracted, stored, version-controlled, and owned, making the question of who owns them the most important one.

Garry Tan shared the most moving real-life example of personal AGI: the story of a friend with a son suffering from a rare disease. This friend, without any institutional support or permission, built a repository of 80,000 markdown files through individual effort. This was a 'brain' for one young boy, pushing himself to the absolute limit of human knowledge about his son's exact condition. All specialist visit notes, papers, seizure logs, and drug interactions were indexed, interconnected, and readily available. As a result, when a new doctor suggested an idea, the father could know within minutes if it had already been tried. Tan emphasized that the story of one father with a laptop and a library illustrates the essence of 'Personal AGI,' explaining it as a powerful example of an individual owning and utilizing their intelligence to tackle seemingly impossible problems. This story dramatically showed how an individual's capabilities can be extended through AI.