In this post
This post summarizes a Y Combinator Lightcone episode, AI Revolution: What Nobody Else Is Seeing (January 2025), recorded in Sonoma at a retreat of about 300 YC alumni AI founders. The guest is Paul Buchheit, creator of Gmail and a YC partner. The figures below are the hosts’ own statements about Y Combinator companies. We have not checked them, and the episode gives no public data behind them.
Key Takeaways #
- The hosts say demand for AI products is unusually high, mostly for agents that do a person’s work for businesses.
- They say YC batches have averaged 10% week-on-week growth since summer 2024.
- Founders at the retreat talked about evals and testing far more than at past YC events.
- The hosts argue the hard part is no longer convincing buyers they want the product but making the product work.
The Current Landscape of AI Startups #
One host quotes a talk at the retreat by Aaron Levie, CEO of Box. In past shifts, to cloud and to mobile, there were always enterprise decision-makers saying “never”. With AI, Levie’s point, in the host’s retelling, is that nobody is saying no: everyone is saying yes, and more.
The hosts add that the fast-growing companies they mention are all the same kind: they sell AI agents to businesses, riding enterprise pressure to adopt AI. That is a selection of YC companies the hosts know, not a sample of all startups.
Growth Rates Like Never Before #
The episode’s numbers:
- 10% week-on-week growth used to be an “amazing” target, hit by perhaps the top one or two companies in a batch.
- Since summer 2024, one host says, the summer and fall YC batches averaged 10% week-on-week growth over the 12 weeks of the batch. That is the batch average, not only the best company.
- One company, the host says, went from zero to $12 million in 12 months.
- Another host says the old advice was to reach $1 million in annual recurring revenue 12 to 18 months after the batch. Some of his companies now do it in about six months.
These are claims from YC partners about companies they back. Survivorship is built in: they talk about the companies that are doing well.
The Power of AI Tools #
The hosts say software that does the work of a person, such as customer support, sales or phone calls, is in demand, and that it is “very, very hard” to build well. Their takeaway is that a technical CEO who is not a strong salesperson can win big enterprise contracts if the product works.
The Importance of Testing and Evaluation #
One thing the hosts noticed: many founder talks were about evals and testing, which they say would not have happened at an earlier YC event. One founder building an AI agent said the most valuable thing his company has built is not the code base but the eval set, a gold-standard labeled set of what the correct answer is. The hosts add that few other companies have this, and that prompting skill (agency and taste) is where much of the opportunity lies.
The Shift in Hiring Practices #
The hosts say that at the retreat they heard nobody brag about blitz scaling or hiring as many people as possible. They also describe founders who will not hire an engineer who does not use code-generation tools such as Cursor. That is a few founders’ hiring stance, relayed by the hosts, not a rule.
The Future of Work #
On job loss, one host uses a story about Milton Friedman and workers digging a canal with shovels. If it is a jobs program, Friedman said, give them spoons, not shovels. The host reads this as the useful model for job-loss fear: AI lets people create far more wealth. The hosts say they do not know how fast AI might displace work, and that no one at the retreat is sure. They are optimistic and say history shows people invent new work.
Conclusion #
This is a snapshot of how Y Combinator partners saw AI startups in early 2025, from companies they work with. The part you can use: the demand they describe is for agents that work, and the asset they say matters is a good eval set. The growth numbers are the hosts’ account, not data you can plan around.
