According to a report by Business Insider on the 13th (today), Y Combinator CEO Garry Tan gave entrepreneurs worried about AI consuming too many tokens some clear advice: burn them—burn them freely. Even if the cost is high, entrepreneurs should be willing to invest heavily in AI tokens.
Speaking about AI agents, Garry Tan said: “You have to turn the intensity all the way up. For example, give it 1 million tokens at once, or 800,000 tokens. I think at that level, you can basically start living in 2028 ahead of time.”
Garry Tan believes that extensive use of AI agents can give entrepreneurs access ahead of time to AI capabilities that most people will not encounter until the future. However, this lead comes at a steep price. “I can’t say for sure, but running agents at maximum intensity might cost $50,000–$100,000 a year (Note: at the current exchange rate, approximately 338,000–676,000 yuan). The cost is astonishingly high. But for a CEO or entrepreneur, it actually makes a lot of sense.”
Garry Tan believes that high-cost experiments can gradually turn into long-term gains. After an AI agent successfully completes a task, entrepreneurs should turn the entire process into a “skill,” organizing it into a set of reusable instructions so that AI can automatically repeat the same task in the future.
Garry Tan summarized this approach in one sentence: “A Markdown file is an employee. This employee completes the work perfectly every time, and you can have it do so as many times as you want.”
Note: A Markdown file refers to a set of written instructions telling an AI agent how to complete a specific task.
Garry Tan is one of the few technology executives to publicly support “tokenmaxxing.” At the same time, some people in Silicon Valley believe that consuming large amounts of tokens is a waste of resources.
“Tokenmaxxing” literally means “maxing out tokens.” It refers to deliberately having AI agents use large numbers of tokens, rather than making token reduction and cost savings the primary goals.
Last week, Uber’s head of engineering, Praveen · Nepali · Nagar, said that Uber had observed some notable AI cost data, which was “another sign that the so-called era of ‘tokenmaxxing’ is coming to an end.” “In the next phase, whatever it is ultimately called, the key will not be who consumes the most tokens, but who can use tokens as efficiently as possible.”
Scott Wu, CEO of AI coding company Cognition, also said on a podcast episode in June that companies encouraging “tokenmaxxing” had begun to lose sight of the point. “The broad direction isn’t wrong, but people have indeed taken some aspects too far. Some might say, ‘We rank engineers based on how many tokens they use.’ It would be better to rank everyone based on how much they actually produce.”