Why I wanted this hour
A camp ends. A habit compounds.
I want to use this hour on something specific, not a pep talk: the difference between a certificate that ends up in a drawer and a portfolio that gets you into rooms. Breathe. You already did the hard part.
Agents, everywhere
Every job gets a co-founder like the one you had this week.
A doctor’s agent reads the scan first. A lawyer’s agent drafts the first version of the contract. An engineer’s agent writes the first version of the code, the way Kimi Code did for your product. Your classroom gets one too: a math tutor that shows its work, a writing partner that drafts and critiques, not instead of your teachers, next to them.
Your first job, five years out
Whatever job you start with will already have an agent built in.
Word processors, spreadsheets, codebases, hospital charts: every tool you open on day one already has an assistant living inside it, the way Kimi Code lived inside your terminal all week. It won’t be a separate class, it’ll just be how the tool works.
You first
Name one job you think agents can never touch.
▸ click to reveal after they answer
Five years ago, most people’s answer was “writing code.” The honest answer today: we don’t fully know yet. Which is exactly why taste, judgment, and relationships matter more, not less.
The honest part
Nobody knows the timeline. Not me, not the experts, not the labs building this.
I could stand up here and pretend to know exactly when. I don’t, and anyone who tells you they do is usually selling something. What I do know: that uncertainty is exactly why builders win. You don’t need the forecast to be right, you need to already be moving when it arrives.
The honest tension
Does AI spread opportunity or hoard it? Both.
It distributes
A teenager in Beijing, on a laptop, just shipped a real product with the same model a funded Silicon Valley team uses. Ten years ago that took that team months and millions. You did it in five days, with no funding at all.
It concentrates
That model is owned, trained, and priced by a handful of labs sitting on almost all the world’s compute. They decide what it can and can’t do, who gets it first, and what it costs. That’s the power and alignment question from Day 1, with real stakes behind it now.
Neither side cancels the other. Both are true at the same time.
Table talk 1 · callback to Day 1OFFLINE · screens closed
You debated who should hold the power over AI. Does building something real change your answer?
3:00 on the clock
1This week you directed a lab’s AI to build something real. Who actually held the power: you, or the lab?
2If everyone in this room can now build like this, does that make the labs bigger, or smaller?
3As a table, decide: does your Day 1 answer still hold? One person reports back.
3:00
Destination: table talk, one voice reports back
The script, updated
The ladder used to be the whole plan.
The old script
grades
→
brand job
→
promotion
→
repeat
one straight line, same rungs, every two years
The new script
build
→
prove judgment
→
own it
loops back, no fixed order, you choose the next rung
The barbell of AI-era careers
Two ends of the bar are safe. The middle is risky.
Direct
the AI
the AI
the risky middle
Be
irreplaceable
irreplaceable
Routine, repeatable work with no judgment layered on top. That’s the part agents are already best at.
Which paths compress, which expand
The path you list is not the path you show.
Compressing
One well-defined task, repeated: a first-pass draft, a routine analysis, a basic support reply. AI already writes a fast first version of all of these. That queue just got shorter.
Expanding
Judgment plus ownership: deciding what to build, standing in front of an unhappy customer, being the name on it when it breaks. AI hands you ten drafts. You are still the one who picks, and answers for it.
That’s why “list of things you studied” now loses to “portfolio of things you shipped,” starting with the live URL you built this week.
How you’d describe yourself
Same skill, two ways to prove it
The claim
The proof
If college feels far away
Then you have the biggest advantage in this room: time.
Build one small thing a month from now until then and you will arrive at those applications with a portfolio nobody can compete with. The older students are slightly jealous of you right now. They should be.
How admissions actually reads this
An officer can click a link in ten seconds. A claim takes longer to trust.
A live URL, a launch video, a first-revenue story: that’s not a better essay, it’s a different category of evidence entirely, verifiable in ten seconds, in a stack of thousands of essays that all say “passionate” and “hardworking.”
Two applications
Same student, two ways to fill the “activities” line.
The résumé line
Vice President, Investment Club. Member, three other clubs. Attended a summer program.
The proof line
Built [product], live at [URL]. 12 users, first yuan earned in five days. Video attached.
Say it like this
“I found a problem, interviewed users, built the product with AI, and earned our first revenue in five days.”
That sentence works in an essay, an interview, a dinner with relatives. Make it yours: name the pivot when your first idea didn’t land, quote the exact thing a stranger said about your product, and say the number, the first yuan you earned. Specific beats impressive, every time.
Two ways to use AI on your essay
Which one actually gets you in?
The copy-paste
asked it to write the essay, changed a few words, never read it against my own voice
The sharpen
wrote the true draft myself, asked what’s weakest, rewrote that part, can defend every sentence out loud
Table talk 2 · callback to Day 1OFFLINE · screens closed
You debated whether AI tells the truth, and whether school still makes sense. Do those answers hold up now?
3:00 on the clock
1On Day 1, what did your table decide about whether you can trust what AI tells you?
2This week you decided how honest to be in your own essay and your own product. Harder or easier than trusting AI to be honest with you?
3Recall Day 1’s school question. You just spent five days building instead of sitting in class. Does that change what you think school is for?
3:00
Destination: table talk, one voice reports back
What stays human: taste
Taste is knowing which problem is worth solving.
On Day 1, a hundred sticky notes went on the wall. You didn’t pick the loudest one, you picked the one that made a real person’s day worse and decided to fix it. That choice was yours, not the AI’s.
What stays human: judgment
Judgment is knowing what good looks like.
When strangers used your product and got confused, you had to decide which fix actually mattered and which was just noise. AI can suggest ten fixes. Only you can tell the important one from the vanity one.
What stays human: relationships
Relationships are knowing real people.
Every real insight this week came from a conversation: a parent, a neighbor, a stranger on the street who told you the truth about your idea. No model can have that conversation for you.
What stays human: agency
Agency is deciding to start before anyone tells you to.
Nobody assigned you a company this week. You picked the problem, picked the team, picked the price, and when a stranger got confused by what you built, you were the one who had to answer for it, not the model. That’s the part AI can’t do for you: deciding what’s worth building, and owning what happens after you ship it. You already did both once. Do them again next month.
The deeper one
AI amplifies direction. It does not supply it.
The best tools in the world can’t tell you what you actually care about. That part is still entirely yours to figure out, and it’s worth more time than any of us give it. What did you notice this week that you actually wanted to fix, not just the assignment you were given?
Five things, starting today
Keep it going, in five moves
1 · Keep building
One weekend a month, one small thing shipped. Not a company, just the loop again.
2 · Ship small, show it
Small and public beats big and private, every time.
3 · Find your people
The group chat doesn’t end when camp does. We keep every card.
4 · Read the source
Paul Graham’s essays, the YC Startup Library. Free, where we got half of this week.
5 · Teach someone else
The fastest way to keep this week is to explain it to somebody else.
Go deeper · on the portal
Keep learning from the source
Paul Graham’s essays
Start with two: “What You’ll Wish You’d Known” (~20 min, he wrote it for high schoolers, literally) and “How to Do What You Love” (~15 min, not what sounds impressive).
paulgraham.com/articles.html
YC Startup Library
Free essays and videos, we borrowed half of this week from here. Same people run the YC YouTube channel, 2.3 million subscribers deep.
ycombinator.com/library
Steve Jobs, Stanford, 2005
15 min · watch tonight. The most-watched graduation speech ever given. Not about AI, about deciding what’s worth doing at all.
youtube.com/watch?v=Hd_ptbiPoXM
6 · Before you go
Write down the next thing, right now, while it’s still real.
Ideas fade fast once the adrenaline of demo day wears off. This card is how we make sure this doesn’t end at 4:30 today.
Exercise 1 · do it nowOFFLINE · screens closed
The next-project card
5:00 on the clock
1Write the ONE thing you will build next month. Specific.
2Add the first step you will take this week.
3Sign it. We collect them, and the alumni community holds you to it.
5:00
Destination: TEAM FOLDER, cards collected
You first
What will your generation build that ours cannot?
There is no answer on this slide. That’s the point: you are going to write it. Let the room talk.