YC Startup School markets itself as a weekend of talks. Jensen Huang, Boris Cherny from the Claude Code team, the founder of Clever, a stack of YC partners. That’s what’s on the page when you apply and that’s what I was there for.
It’s free, but it isn’t open. You apply, and if they want you there they fly you out. I got in, flew to San Francisco, and the talks were genuinely good.
Boris Cherny’s was the one that actually moved something for me. The gist was that we spend way too much effort prompting and steering these models, and most of that effort is really us hobbling them. The direction is clearly toward less steering, not more. I’ve been building agents for a year and my instinct the whole time has been to add more scaffolding, more instructions, more guardrails. Hearing someone who works on this full time say the opposite is one of those moments where you realize your mental model is a version behind.
Jensen’s point was simpler and probably more useful commercially: AI replaces tasks, not people. That sounds like a throwaway line until you point it at your own product. Every AI company that’s actually working right now is just taking some annoying task and doing it well. Not replacing a job, not being an everything-agent, just killing one specific annoying thing. It made me go back and ask which exact annoying task Tolus kills, which is a better question than the one I’d been asking.
So I left with two good ideas. Here’s the thing though. Ideas are the cheapest thing at an event like that. Both of those talks will end up online and you’ll get most of what I got for free, from your couch. What I couldn’t have gotten anywhere else was who was standing next to me.
I met a guy heading to Harvard who had just raised $2 million for half a percent of his company. That’s a $400 million valuation on someone who hasn’t started freshman year, for non-invasive brain-computer interfaces, so basically Neuralink without the surgery. I met a girl who deferred Stanford to build surveillance robotic birds, which I have not stopped thinking about since. And I met an absurd number of incoming Berkeley freshmen, my exact year, all of them building something real.
That’s the part that actually changed me, and the interesting thing is it took a few days to work out why. Your sense of what’s normal gets set by whoever you can see. Mine had been set by my high school, where building a startup at all made you the weird one. Spend a weekend around people your exact age raising rounds and turning down Stanford and that baseline moves. Not in a motivational-poster way. In a way where a bunch of things I had been quietly filing under ambitious got recategorized as ordinary.
Which brings me to the most useful advice of the weekend, which also didn’t come off a stage. Gray Baker invited my cofounder Hao to a semi-private dinner, 15 people at the table, and spent a good part of it pushing him to just apply to YC. The argument is almost insultingly simple. There’s no cost to applying now versus later. You apply now. Then next batch. Then the one after that. It’s a 0 loss function.
Obvious, right? I had known that for months and still hadn’t applied, because I’d decided I wasn’t ready yet. The thing is, “not ready” was never an assessment. It’s a comparison, and I had been running it against the wrong set of people. Worth noticing where that landed too. A table of 15, not an auditorium.
The other half of this happened after the event was already over.
The day after, I had a meeting with Ethan Harrison, an incoming senior at Berkeley who went to my high school 4 years ahead of me. Nothing to do with YC, we’d set it up weeks before. But it’s the same lesson wearing different clothes, so it belongs in the same post.
Two things came out of it. The first is a connection. Ethan is close with John DeNero, who teaches CS61A and apparently loves startup people. So: Ethan went to my high school -> Ethan knows DeNero -> DeNero likes startups -> I now have a pitch to write. That’s a door I had no path to a week ago.
The second is worse and more useful. Ethan told me about Gradescope, and I had never heard of it, which in hindsight is a little embarrassing.
Gradescope was built at Berkeley in 2014 by a CS professor named Pieter Abbeel along with a few of his students and TAs who were drowning in grading. It started as a prototype called Pandagrader, went through Berkeley’s SkyDeck accelerator, and by 2018 it had been acquired by Turnitin. It now runs most of the CS, math, and chemistry classes on campus. So what I’m walking into isn’t really a competitor, it’s infrastructure. Students there have been submitting to it since before I started high school.
And here is the part that actually made me laugh. A month ago I wrote a whole post about how Tolus exists because we don’t believe in Turnitin-style verification in the age of AI. The incumbent at the school I’m about to attend is owned by Turnitin. I did not see that one coming.
Then there’s Pensieve, which I assumed was some kind of Gradescope spinoff and is actually a completely separate company. Launched February 2025, LLM-native from day one. It transcribes handwritten STEM work, grades it against the rubric, and attaches a confidence score so the professor only reviews what the model isn’t sure about. 100+ universities in about 18 months, and it’s already running in Data 8 at Berkeley.
So that’s the real landscape. An entrenched incumbent owned by the exact company my thesis is against, and a fast LLM-native challenger that got to my campus before I did. Neither of them is a direct competitor. They grade the work, we check whether the student can actually defend it, which is a different layer of the same problem. But they own the workflow and they own the campus, so the question stops being how do we beat them and turns into what do we build on top of, and inside of, what’s already there. That’s not a pivot. It’s just the thing I should have known before I had opinions about the market. I learned all of it from one guy over coffee, the day after the conference I flew across the state for.
So in one week I picked up two ideas, a recalibrated sense of what people my age are actually doing, a professor connection, and an incumbent I didn’t know existed. Exactly one of those came from the thing that was advertised.
Tyler Bosmeny, who founded Clever, said the thing that explains why the rest of it worked. Edtech is hard, long sales cycles and broke customers and procurement, but people will go out of their way to help you if you’re building something real for a good reason. The caveat being it can’t be slop. Nobody helps you build slop. Every useful thing on that list came from somebody deciding I was worth 20 minutes of their time, and that’s the filter they were running.
So, what I actually took away:
- The talks are the advertisement. The people are the product, and that holds whether it’s a room YC screened or a guy from your high school.
- “I’m not ready yet” is a comparison, not a fact. Change who you’re standing next to and it tends to evaporate.
- Go find out what already exists in your space before a senior tells you about it over coffee.
This goes further than YC. When you’re deciding whether to apply to something, the speakers and the agenda are the least important part of the pitch. That’s marketing. What you’re actually getting is proximity to whoever else got in, and the harder the filter, the more that’s worth. The uncomfortable version of this: if you get into a room like that and spend the whole time sitting in the audience, you used maybe 20% of it.
It’s also why I’m a lot more excited about Berkeley than I was in June. I met a lot of people my year headed there and the density of talent was genuinely startling. That was the first time I’ve seen the Berkeley network do anything for me and I don’t even go there yet, which seems like a good sign for the next 4 years.
Anyway. I’m applying to YC this batch, and then the one after that, and the one after that. 0 loss function and all. But first I’ve got a DeNero pitch to write and a competitive landscape to actually go read, because of a tool I found out about the day after the conference ended :).