Two days at YC Startup School

YC flew a few thousand builders to San Francisco for two days at the Chase Center — July 25 and 26. I was one of them.

Walking in the first morning was surreal. An actual arena — the security line through Thrive City Plaza, Center Court split into two stages, headphones for the far one — filled with people who all build things. Day two, the surreal had worn off a little. But the whole weekend stayed the same kind of incredible, and I want to get back next year badly enough to work for it.

Here’s what stuck, talk by talk, plus the practical stuff — credits, parties, people — that made it more than a lecture series.

Day 1

Jensen Huang (NVIDIA) with Garry Tan

I’ll be honest about what survived the weekend: mostly a feeling. Jensen is relentlessly inspirational in a way that’s hard to quote afterward — the thing I keep coming back to is how hard can it be? Not as arrogance, but as a default posture toward problems that should be terrifying.

The catch, and he was upfront about it: how hard can it be is how you talk yourself into suffering. You underestimate the problem, you commit, and then it’s much harder than you thought. But that’s the point. The suffering is where the company — and the person — actually gets built. He wasn’t selling the dream without the invoice.

The part that landed practically was him talking about learning systems. I’m not even sure I absorbed it the way he meant it, but what it did for me was rehabilitate the fundamentals. System design, DSA, the LeetCode grind I’d half-written-off in the vibecoding era — Jensen made me feel like understanding how systems actually work is still the foundation everything else compounds on.

The full conversation is up now: Jensen Huang: The Mindset That Built NVIDIA.

Boris Cherny (Head of Claude Code, Anthropic) with Diana Hu

This was the one I was most primed for, since most of what I build these days runs through Claude Code — usually two to four terminals at a time.

Two things stuck. First: exploit long-running agents. Stop treating Claude like a chat you babysit turn by turn — let it explore, let it run loops, give it work that takes a while and get out of the way. I do some of this already, but he framed it as the whole point, not a power-user trick.

Second, a discipline with a name I initially lost and then recovered: unhobbling. Every time a new model ships, strip your code and prompts back down and start clean. All the scaffolding and workarounds you wrote for the last model’s weaknesses don’t transfer — they hobble the new model, a ceiling you built yourself. That one stung a little, because I know exactly which of my prompts are load-bearing crutches.

Also: attendees got six months of Claude Max 20x. Which is insane.

This session’s up too: Boris Cherny: Building Claude Code.

Jeff Dean (Chief Scientist, Google DeepMind)

Full honesty: I fell asleep.

Not because it was bad — because it was over my head. This is the guy behind “Latency Numbers Every Programmer Should Know,” a genuine legend of systems engineering, and the talk was pitched at a depth my current self couldn’t hold onto after a morning at the arena. Somewhere between the technical weeds, I faded.

What I actually took from it wasn’t content — it was the gap. There’s a level of technical depth where people like Jeff Dean operate, and I felt clearly, physically, that I’m not there yet. I want to close that distance, not just nod along at the keynote version of it. Which loops back to what Jensen accidentally convinced me of: the fundamentals are still the way in.

Dmitri Dolgov (Co-CEO, Waymo)

The schedule had me at Blake Scholl’s Boom Supersonic session, but the Waymo talk pulled me in instead. Partly this was just relevance — I don’t think much about supersonic flight, but I had literally taken a Waymo to the arena that morning. Hard to pass up hearing how the thing that just drove you works.

The technical part I enjoyed most was the sensors — how lidar, cameras, and radar each cover the others’ blind spots, and the system comes from combining them rather than betting on any one. But the bigger thing was the framing. Dolgov talks in nines: not “does it work,” but 9.999-something reliability. And listening to him, it hit me that self-driving is in some sense solved — the surreal part is that the remaining work is logistics. Rollout, regulation, scale, trust. The demo was the easy part.

The other thing that stuck: everything he said felt grounded in accumulated experience — miles driven, edge cases hit, data collected. Not vision-deck confidence. Earned confidence. There’s a difference, and you can hear it.

Susan Kare (Iconographer and Designer)

I’ll admit I walked in not knowing who she was. It took about five minutes of her talk to understand I was listening to one of the most legendary designers alive — she drew the original Mac icons, the ones that taught a generation what a computer interface could feel like.

What made the talk great wasn’t a framework or a lesson. It was the way she told it: lighthearted, funny, full of old stories and quotes from the people she worked with in those early days. When your career is that legendary, you’ve earned the right to just tell stories — and the stories were the point.

And her work looked like her. Whimsical, warm, human. Watching the designs go by while she talked, you could see her personality in every pixel. I don’t think I’d ever felt that so directly before — that the best design isn’t neutral. It’s someone.

Day 2

Patrick Collison (Stripe) with Harj Taggar

I’ll be honest: I’ve forgotten most of the specifics. What I remember is that it was just solid — and that Patrick was strikingly humble for someone running Stripe.

The story that stayed with me was the early days: nobody believed them. Two random guys trying to convince banks and payment networks — institutions with every reason to ignore them — to take a chance on their thing. Hearing how hard that actually was, from the person it happened to, hits different than reading it in a retrospective where the ending is already known.

Oh, and they gave all of us Stripe Atlas for free. Amazing.

Alexandr Wang (Meta) with Garry Tan

I walked into this one with baggage — you hear a lot about what’s going on at Meta, and most of it isn’t flattering. But I’d also heard Alexandr’s story, and the story is incredible, so I was excited anyway.

Part of what got me was surface-level and I don’t care: the style, the hair, the age. He felt like us. And then you remember what he’s actually doing — leading AI at Meta at that age — and the gap between “feels like one of us” and “operates at that level” is its own kind of aura.

There was also a point made that weekend — I think it was Alexandr, though I’ll admit it might have been Patrick or even Jensen; the talks blur — about the AI-taking-jobs framing being backwards. When AI absorbs work, it frees humans to work on better ideas. More people on harder problems, not fewer people, period. What got me is that this is literally what I already believed while everyone around me fixates on replacement. The move isn’t to mourn the assembly line — it’s to be the one building what replaces it. And the flip side of that, which Alexandr kept pressing: the models are here whether you like it or not, so the actual skill is figuring out how to leverage them. Not whether. How.

The takeaway that survived: first-principles thinking, and conviction. You have to stand in your idea even when people don’t believe you — especially when people don’t believe you, because disbelief means less competition. The same theme Patrick hit from the other side: nobody believing you isn’t a signal to stop. Sometimes it’s the moat.

He also mentioned attendees are getting credits for Muse Spark, Meta’s new model — the pitch being something like 6–7x cheaper than Opus on output for near-parity performance on agentic work. Haven’t received mine yet, but at that price it’s worth trying.

Peter Steinberger (“ClawFather”)

I’d heard so much about OpenClaw that seeing the ClawFather himself on stage felt slightly unreal.

What struck me first was that he seemed aged — and he said as much himself. AI time moves so fast that OpenClaw, the thing that made him famous, already feels like it kind of died. Months, not decades. He talked about it the way you’d talk about a past life.

But the part I loved was his motivation. He works on things that annoy him. That’s the whole system. “Why can’t I just prompt this?” — and then he goes and builds the thing so he can be lazy about it. I love that, because it’s exactly what I want: to just talk to Jarvis and have everything happen. Laziness as a design goal is underrated.

The other filter he uses: it has to be fun, and it has to be something you actually use. Not a market thesis, not a TAM slide. Annoyance in, fun out, and you’re your own first user. That might be the most honest product framework I heard all weekend.

And then there’s his stack. He’s probably one of the heaviest AI users alive — maybe the heaviest — so hearing how he actually runs things was worth the session alone. The core of it: an always-on machine he remotes into, so his agents never stop when he steps away. No closing-the-laptop-kills-the-session problem; the work just keeps going and he checks in from wherever.

I’m now fully talked into buying a Mac mini for this — leave it running at home, SSH in from my Air, and let Claude keep working while the laptop’s closed in my bag. It’s such an obvious upgrade once someone shows you they live that way.

Chelsea Finn (Stanford / Physical Intelligence)

The takeaway here was simple and loud: robotics is the next big thing.

Watching the demos — a robot folding laundry, making a latte, washing dishes — something clicked that no amount of reading about “embodied AI” had done. These are the chores. The actual, daily, nobody-wants-to-do-them chores. Watching a robot just do them, competently, made the value so concrete it was almost funny. So valuable. Not in the abstract-TAM sense — in the “I would pay for this tomorrow” sense.

Software ate the world; the models ate software. Watching Physical Intelligence’s demos, it’s hard not to conclude the hands are next.

Honestly, it made me want to go work for her lab. And she was clear that a robotics background isn’t required — pure software engineers are super relevant to what they’re building. But I know myself: I have loops to close first. Master the software side, get deep with the models — the dream is still working on frontier models at Anthropic. Robotics goes on the list as a real interest now, though, not a spectator sport. That’s new as of this weekend.

Sam Altman closing it out

The closer, and honestly, the one that landed least for me.

Which isn’t to say it wasn’t something to see. The personality is real — he’s not rigid or formal at all, just speaks freely, radiates optimism, tells stories from his YC partner days like he’s catching up with friends. “Stop larping” got a real laugh out of me. The aura is there.

And one idea of his did stick: startups have never had this much leverage. A couple of people with AI can now do what used to take a team, and he framed that as raw economic growth waiting to be claimed — the cost of trying something has never been lower, so the case for starting has never been stronger.

But something felt off, and I’ve been trying to be honest with myself about what. Where Patrick read as humble and Susan read as generous, Sam read — to me, that day — as a little self-regarding. Maybe that’s unfair. Maybe I walked in carrying too much of what I’d already seen online and just found what I was looking for. I genuinely can’t tell how much was the talk and how much was my bias. But of the ten sessions, the biggest name left the smallest mark, and I think that’s worth writing down.

The credits

Attendees got access to over $25K in compute credits from OpenAI, Anthropic, AWS, and Azure — plus six months of Claude Max 20x, free Stripe Atlas for incorporating, and Muse Spark credits from Meta (still waiting on those).

The one that actually changes my day-to-day is the Claude Max 20x. I live in Claude Code; six months of the top tier is genuinely insane, and it removes the last excuse for not running agents the way Boris and Peter described. Atlas is the sleeper — incorporation is one of those things you put off until it’s free and sitting in front of you. The raw compute credits I’ll burn as side projects demand.

The people

The talks were half the event. The other half happened at night.

Friday, before Startup School even started, there was a fireside-chat party with Arlan — someone I’d read about and heard about for years, so just being in the room felt insane. My flight got me there embarrassingly late, maybe an hour before it wrapped. Still worth it, because I met Surjan, a researcher at Microsoft working on computer use and red teaming. He told me something that stuck harder than half the keynotes: always work on a problem worth solving — one that actually demands critical thinking. It landed at the right moment, because I’d just made the hard call to leave Nabi, and this was the exact reasoning behind it. He’s working on economic AI algorithms now — the kind of hard problem that can’t be vibecoded — and talking to him was genuinely inspiring. Met another builder, Chenny, working on something in the same orbit.

Saturday was the LemonLime “Calling All Crazy Builders” party. Met a guy who watches anime and was just easy to talk to — we follow each other now. Also got time with one of the founders, who told Stanford frat stories including what he claimed was the real origin story of Snapchat, which I will not be repeating here. Kind of inappropriate? Definitely memorable. The founder himself was chill in a way that reminded me of Revin’s CEO — an esports background he actually built on, clearly inspiring as a person, even if the product isn’t something I’d want to work on myself.

Sunday was the AWS party at the Exploratorium, and honestly, a wash. Uber voucher problems, a massive line, too much money on mid food (partly my own ordering mistake, plus whatever that SF health surcharge is). No real perks. But I got to see the Exploratorium, and I forced myself out of my comfort zone to talk to strangers, which is its own kind of rep.

The real return on the whole weekend, though: Caden and Neevas. Caden’s at Palantir — super funny, instantly easy to be around. Neevas is intensely ambitious and a little awkward, which I mean as a compliment. I bonded with both of them for real. Two days of celebrity keynotes, and the thing I’d protect most from the weekend is two new friends.

What actually changed

Ten talks, three parties, two days. Here’s what I’m actually doing differently:

Back to fundamentals. Jensen made the case by accident and Jeff Dean made it by putting me to sleep: there’s a depth of systems understanding I don’t have yet, and the vibecoding era doesn’t excuse me from it. System design and DSA are back on the schedule — not for interviews, for the gap.

Agents that don’t stop. Between Boris’s long-running loops and Peter’s always-on machine, my setup is officially embarrassing. A Mac mini is getting bought. Claude keeps working while my laptop’s closed in my bag.

Conviction as a strategy. Patrick’s nobody-believed-us and Alexandr’s disbelief-means-less-competition are the same lesson from opposite ends of the journey. I’m taking it personally, because the thing I’m building has plenty of doubters — starting with me on bad days.

Problems worth solving. The line from Surjan that reframed my whole month: work on problems that demand critical thinking. It’s why I left my job. The weekend confirmed the bet instead of making me doubt it.

And a new item on the long list. Robotics went from “cool demos online” to something I genuinely want to explore — after I close the loops in front of me. The hands are next. I want to be ready when they are.

See you next year, hopefully. I plan to earn it.