ilham
What an inspiring talk by @bscholl!!
> Knowledge is changeable, but love and passion are harder. So work on what you love.
> Turns out you can change the regulations, and the way you do it is just by building the thing and communicating why it's okay.
Amazing achievement and perseverance, supersonic travel is coming back!!
@ycombinator @garrytan
Gabriele Scotto di Vettimo
Re @ycombinator Startup School day 1 just ended‼️
I got to hear from the top minds shaping the AI industry: Jensen Huang, Boris Cherny, Jeff Dean, Garry Tan, Dmitri Dolgov, Michael Kratsios, and more
If you couldn’t make it, here’s what stood out from each talk 🧵
Simone Syed
Power generation in the @ycombinator locker room
@mkratsios47 @bscholl
Simone Syed: Ayo YC Startup School!!
Sanjana Friedman
So incredibly cool
Jared Poblete: POSTCARD for YC Startup School 2026
Haiyu Wu
Recent discussions about open-sourcing make me feel that I should go back and revisit these important open-source works in representation learning that pushed the field forward and eventually made vision foundation models so useful in applications.
These are the papers that come to mind. Feel free to correct my opinions and share any important papers that I may have missed.
MoCo v1 (2019-11) – ResNet
Method: contrastive loss + a queue of samples from previous batches for negative pairing + EMA-based weight updates
# of training branches: 2
Effect: EMA-based weight updates became an important training paradigm in later representation learning algorithms, but this work used EMA only to produce stable embeddings for negative samples.
SimCLR (2020-02) – ResNet
Method: contrastive loss + more augmentations + an MLP projector to capture the effects of augmentations
# of training branches: 2
Effect: An MLP projector became the gold standard in representation learning algorithms.
BYOL (2020-06) – ResNet
Method: positive pairs only + an asymmetric architecture + stop-gradient + SimCLR augmentations + EMA
# of training branches: 2
Effect: Probably the first representation learning method without negative pairs. It inspired DINO and probably JEPA in terms of architectural design.
SwAV (2020-06) – ResNet
Method: global/local views + SimCLR augmentations + class prototypes
# of training branches: 2
Effect: Its global/local-view training became the standard setting in later representation learning methods.
SimSiam (2020-11) – ResNet
Method: BYOL without EMA + a constant learning rate for the predictor
# of training branches: 2
Barlow Twins (2021-03) – ResNet
Method: global views + SimCLR augmentations + batch-level covariance optimization
# of training branches: 2
Effect: The optimization target changed. Previous works used cosine similarity, but this work uses covariance, which explicitly regularizes the latent space.
MoCo v3 (2021-04) – ViT
Method: MoCo v1 - memory bank + SimCLR augmentations and an MLP + a frozen first layer
# of training branches: 2
Effect: The first work to show that ViT can be used for representation learning.
DINO (2021-04) – ViT
Method: BYOL - the predictor + global/local views + softmax distillation + centering + temperature sharpening
# of training branches: 2
Effect: Enabled the era of large-scale vision foundation models.
VICReg (2021-05) – ViT
Method: Barlow Twins → branch-level covariance + variance + sample-level invariance
# of training branches: 2 or 1
Effect: The first representation learning method to support stable one-branch training without EMA, centering, layer freezing, temperature sharpening, or stop-gradient.
MAE (2021-11) – ViT
Method: learning representations through pixel reconstruction
# of training branches: 1
Effect: A simple, simple, and simple training strategy that is widely used in medical imaging, from what I have heard.
I-JEPA (2023-01) – ViT
Method: learning representations through latent prediction
# of training branches: 2
Effect: The first representation learning method in the spirit of JEPA. It is also the basis of V-JEPAs, VL-JEPA, and so on.
DINOv2 (2023-04) – ViT
Method: DINO v1 + SK centering + KoLeo + LayerScale
# of training branches: 2
Effect: The first large vision foundation model to show superior performance across a wide range of downstream tasks.
DINOv3 (2025-08) – ViT
Method: DINO v2 + Gram anchoring + flattened schedules
# of training branches: 2
Effect: The first large vision foundation model to show superior performance on fine-grained tasks, such as segmentation and depth estimation, and to achieve accuracy comparable to supervised learning methods on ImageNet-1K.
LeJEPA (2025-11) – ViT + others
Method: VICReg → full-distribution regularization. It provides mathematical proofs of why an isotropic distribution is the ideal target for the latent space and proposes SIGReg, a method that is linear in the scaling factors.
# of training branches: 1
Effect: No heuristics are needed for training stability. It provides a simple solution for various tasks, including world models, such as the LeWM series; multimodal learning, such as LeVLJEPA; drone applications, such as SkyJEPA; and so on.
VISReg (2026-06) – ViT
Method: SIGReg → a faster and stronger anti-collapse method that can form an isotropic Gaussian distribution more efficiently. It decomposes the regularization into center, variance, and shape losses, allowing the model to begin introducing isotropy at an early stage of training.
# of training branches: 1
Effect: It is too new to summarize concretely.
Elena
i love how michael kratsios or more affectionately “kratsios” has been catapulted to the public stage seemingly instantly, but it also feels like we’ve always known his name, and been calling him good-natured diminutives for our entire life. the new kid golden boy on the block, fellow varsity football players affectionately slapping his ass with the towel in the locker room. “hey kratsios man, good play during practice.” “kratty boy, i love u man, ur the greatest.” “krat us up.” etc
Laura Buhler
I'm not an insider at @ycombinator , so can only comment from the outside-in, but as a longtime industry observer, it seems to me that @garrytan has breathed an entirely new life into that program. I've heard the term "re-founder" - he is fitting of it.
Agreed!
Shoshi: @mattshumer_ Crazy repo, digging into it and it genuinely contains interesting solutions. This would be next level with data rich asset library access to pull specific high quality previous generated assets from. Future of gaming will be nuts, and future game engines will be based on this
This is what good prompting looks like
I don’t make the rules 🤷♂️
Matt Shumer: Prompt:
I want you to build a first-person shooter at the level of the most recent Call of Duty games. It should be utterly perfect, visually beautiful, with every single thing done at AAA quality—from textures to physics to anything you could think of.
Fan out sub-agents and
TO
I still think about @paulg predicting in 2009 that Sam Altman would build a company as big as Apple and Google
Beff (e/acc)
OSTP Director @MichaelKratsios: AI will be the most transformational technology for the acceleration of scientific progress.
We are entering a Golden Age of Science
tfw when last nights edible finally kicks in
Bence Redmond
I attribute a lot of who I am now to last year's AI Startup School.
It was the first that I was surrounded by people every bit as ambitious as me, and it was infectious.
In fact, I went into it a complete AI skeptic! Now I'm in S26 building a cloud agent company 😅
To everyone in the bay for this year's SUS, take advantage of it! It's definitely one time of the year where losing a lot of sleep is worth it 😁
And massive props to the team at YC behind this, and the serendipity that they enable. It truly changed my life.
Ankit Gupta
Live in the future. Build what’s missing. @ycombinator
Ken Feinstein
Re @mattshumer_ Made a version where you run around a library and shoot giant book worms.
We should prioritize people and a vibrant community and more housing far far far more than some guy’s views
💥Susan Dyer Reynolds🗞️: ‘We live up on Russian Hill and our view is going to be impacted somewhat by this,’ said John Molinari.” Hate to break it to you, John, but there are no view rights in San Francisco. I wrote a story about it several years ago. https://abc7news.com/post/hundreds-voice-opposition-plan-develop-san-francisco-marina-safeway-site-twin-tower-high-rise/19566811/
Play it here: https://pages.workbench.md/p/pg_lzr8fzlTfUfh
Matt Shumer: Claude Opus 5 one-shotted this game.
EVERYTHING you see in this demo is custom code... not a single external asset was used.
AI games are going to be amazing.
(sound on)
Jake Dahn
this is pretty neat, it felt surprisingly good for an in-browser game
I'm hoping people start making new/unique/fresh ideas instead of rehashing existing games
Matt Shumer: Play it here: https://pages.workbench.md/p/pg_lzr8fzlTfUfh
First replications are coming in! Super cool to see :)
Ziwen: This quality of the game is insane with only one prompt
Rishi
This is wild… did not think it would be possible, but @mattshumer_ has once again shown us all how far these models can go
I ran an existing game I made with Kimi K3 last weekend through Opus 5 on High mode and the first-shot transformation it did in a single prompt was mindblowing
The entire map, all weapons, every interactive machine (pack-a-punch, perk machines, mystery box, teleporters, doors), sky, particles, muzzle flashes were generated with code - NOT pre-made assets
Game runs on Three.js; sound effects were fully generated with @ElevenLabs SFX + Voice Design/TTS for the characters
(sound on; transformation prompt and game link below to try it yourself)
Matt Shumer: Claude Opus 5 one-shotted this game.
EVERYTHING you see in this demo is custom code... not a single external asset was used.
AI games are going to be amazing.
(sound on)
Appreciate the kind words :)
I think there’s lots of room to push Opus much further.
I stopped the loop after a while because I didn’t want to waste my limits… but seeing the reaction to this, I’ll attempt something even crazier soon!
Rishi: This is wild… did not think it would be possible, but @mattshumer_ has once again shown us all how far these models can go
I ran an existing game I made with Kimi K3 last weekend through Opus 5 on High mode and the first-shot transformation it did in a single prompt was
true fact
Titanium: Codex is sooooo good! Damn!
Another replication! So exciting to see the community using my prompt.
Keep em coming!
Zhuo Tao: @mattshumer_ so i tried the prompt @mattshumer_ provided, and have opus 5 one shot the game for me, it took about 2 hours on my m5 macbook pro. and it's working!
Mike Solana
genuinely how the fuck is there not a snowcrash movie yet
spor
Re @Winterrose always feels odd when others demand a sort of monk-like asceticism from you once you succeed
or the “you must be what you have always been” thing. it’s strange
∀ɯɹo oɹɯ∀
I didn’t believe it at first. I even attacked the author. After digging deeper, I realized it was real.
Prompt-generated games aren’t ready yet, but this demo shows we’re close.
Players feared publishers wouldn’t let them own games. Soon, publishers may not own the games either.
Matt Shumer: Claude Opus 5 one-shotted this game.
EVERYTHING you see in this demo is custom code... not a single external asset was used.
AI games are going to be amazing.
(sound on)
dnap
Sam Altman says the most important invention of the Industrial Revolution wasn't a technology, it was the modern corporation
"Before that was invented, you had these family businesses that ran on trust, and you had to know everyone, and you didn’t have the idea of a corporation. You didn’t have the idea of stockholders really."
"Then all of a sudden, this new thing was invented. Countries, the sovereigns of the world, granted this sort of new kind of status, a thing that didn’t exist before."
"It's this very powerful thing: we’re going to let these new entities emerge and not give them the power of a state, but give them much more than the power of a person."
"Fundamentally, this was about incentive alignment, liability protection, the ability to amass capital. And all of a sudden, you could do things beyond what a family business could do."
"This was really an incredible thing. And the idea of a company has done a huge amount to align the incentives of very large groups of people. Now millions of people can be shareholders of a company."
GPT-Live for a different reading experience:
José Ocampo: Reading a book while having GPT-Live open to talk to, ask questions, write down notes and bookmarks really feels like reading in 4D
Steve Rattner
Let me see if I have this right: The Qataris give the Defense Dept a 747. Our govt puts $400mm of upgrades into it. Trump gets to keep the plane where he leaves office.
https://www.nytimes.com/2026/07/25/us/politics/trump-air-force-one-security.html?smid=nytcore-ios-share
nic carter
All political questions in America reduce to this reality. Social security, Medicare, government spending, everything - it’s this debate a thousand times over. And the highly mobilized older population keeps winning. The young will be sacrificed for the old. Over and over.
Laura Waxmann: Hundreds of #SanFrancisco Marina residents turned out this evening for townhall on controversial @Safeway housing plan, which would place what for many is an uncomfortably tall, 22 & 18 story project at the low-slung neighborhood’s waterfront.
@sfchronicle
George Conway ⚖️🇺🇸
.@realDonaldTrump still has absolutely no idea what he was trying to do, how he was going to do it, why he wasn't able to do it, or what he's going to do next to try to fix what he has no idea he broke.
He's a delusional, crackbrained, and deranged moron.
It's just that simple.
The New York Times: Breaking News: President Trump has set aside, at least for now, plans to sharply escalate the American military assault against Iran. https://nyti.ms/4vQpeCO
The @ChatGPT mobile app becomes very laggy if you’re having a long conversation - even inputting text becomes slow.
Mark Gadala-Maria
Opus 5 is a BEAST.
I tried @mattshumer_'s viral prompt and it actually one shot a real FPS zombies style game.
I tried something similar with Fable and Kimi K3 and it never got anywhere close to this.
This cost me almost $400 in tokens and took about 3 hours.
Moe Davis (U.S. Air Force, Retired)
Trump’s Golden Age . . .
Measles epidemic
Explosive diarrhea
$4.11 gas ($3.15 yr. ago)
$5.30 diesel ($3.74 yr. ago)
$6.83 ground beef ($6.12 yr. ago)
149 days of war with Iran (18 U.S. Troops killed)
$39.67 trillion dollar national debt (up $3.57 trillion in 18 months since Trump retuned)
China viewed more favorably than the United States globally
5 million Americans lost healthcare coverage since Trump returned
$55 billion in handouts to farmers mostly due to Trump’s tariffs destroying their markets
46% spike in farm bankruptcies
21% spike in mortgage foreclosures
11.9% increase in personal bankruptcies
Peter Yang
“The only job left is to understand what you don't like, put it into words, and tell the AI.”
Here’s my new episode with @jxnlco (DevEx at OpenAI) where he showed me how he uses Codex throughout his workday, including how to:
→ Set up a chief of staff for Slack and email
→ Turn past sessions into new AI skills
→ Give long-running projects verifiable goals
Some example Codex use cases from Jason:
“I want you to read the past week of Slack messages I’ve written and make a skill to figure out how to talk like me.”
“Turn this thread into a heartbeat. I want you to check my emails, my Slack, my Linear, and tell me what I need to prioritize.”
“Codex can find a booking ticket and check into the flight for me and text me the boarding pass to my phone.”
Jason wrote the official OpenAI handbook on how to get the most out of Codex, so don’t miss this episode.
📌 Watch now: https://youtu.be/ydTeb_I0b94
Thanks to our sponsors:
@RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang
Google AI Studio: Build apps with Nano Banana, and more https://aistudio.google.com/?utm_source=peteryang&utm_medium=partner&utm_campaign=26Q3-peteryang_ais
Brad Flora
> VC mistakes largest ever gathering of earnest nerds for thing that better maps to his world model.
Brother, you talk to CMU alums so I gotta imagine you’ve met your share of earnest technical founders who want to fix a thing they think is broken. Are you really befuddled that there might be thousands of them out there? https://www.paulgraham.com/earnest.html
Jay Kapoor: Looks like a megachurch for people who only worship money
Dan Gray
Something that @garrytan understands better than most:
The bottleneck on progress is the rate at which individuals with great ideas and insane determination can be converted into founders.
Venture capital has an "if we build it they will come" mentality that manifests as a zero-sum competition for attention within the pool of surfaced talent.
This assumes that the interface between capital and talent is as good as it can be; that every potentially great founder is able to engage with that opportuniy.
It's a supremely silly (and fairly arrogant) point of view that bears an unknowable cost.
(This is why I have deliberately profiled firms like @1517fund and @BoostVC, or individuals like @arian_ghashghai, @cyantist, @alexlmiller and @Besvinick, who genuinely expand the frontier of what and who is fundable.)
We have come nowhere near the ceiling of human entrepreneurial potential. Many of the impediments are regulatory, institutional, systematic. One key constraint is simply a lack of competent investors with the courage to back unusual founders and ideas, particularly in hot markets.
This is the gap that @ycombinator helps to fill, by encouraging people to engage with entrepreneurship.
On top of this, they apply a refined process which displaces many of the less worthy signals and credentials that investors lean on. It is the gold standard for screening for talent at scale and giving the market confidence to take the baton.
My only hope is that they take this on the road, to spread their cultural footprint further across the US, and beyond.
It is important that the messages reaches everyone.
Scott Stevenson: I have great respect for YC, but need to say what many are surely thinking:
Startup School in a stadium feels antithetical to the YC brand and to what makes startups work.
The YC brand felt like it was designed to filter out status chasers.
The bare Mountain View office. The
Christopher Golda
The goal of Y Combinator has never been scarcity and that's implicit in the name. We think there should be more startups in the world, and YC exists to help create them.
There are far more people who can, should, and will be in the arena. So yeah, we rented out an arena.
Scott Stevenson: I have great respect for YC, but need to say what many are surely thinking:
Startup School in a stadium feels antithetical to the YC brand and to what makes startups work.
The YC brand felt like it was designed to filter out status chasers.
The bare Mountain View office. The
chatgpt work is remarkable, and "work" undersells it.
from my phone i sent:
"use all my chat history to figure out ideas for a long weekend trip with 8 friends, plan the best three options, make a full-stack site where the 9 of us can coordinate on what we would want to do in each place and decide where to go, and then after we get to group agreement make reservations. draft an email in my gmail i can send out to my friends when the site is ready."
it...just worked.
Sam Altman
chatgpt work is remarkable, and "work" undersells it.
from my phone i sent:
"use all my chat history to figure out ideas for a long weekend trip with 8 friends, plan the best three options, make a full-stack site where the 9 of us can coordinate on what we would want to do in each place and decide where to go, and then after we get to group agreement make reservations. draft an email in my gmail i can send out to my friends when the site is ready."
it...just worked.
Naman Narula
it's a wrap for day 1 at @ycombinator startup school 2026!
> #1 takeaway from garry tan: build personal AGI. a brain that nobody curates is a garbage dump with great search
> met & chatted with @gustaf (i know!!), loved nerding out on ai native service companies. best conversation i had here.
> @bcherny gave everyone free claude max 20x for 6 months. wild.
> met hundreds of founders, and insane YC swag lol
all set for day 2 now!! :)
Republicans against Trump
CBS/YouGov poll:
Trump's policies are making you
Worse off-58%
Better off-13%
Republicans against Trump: NEW CBS News/YouGov poll:
67% say the U.S. should end the war with Iran now, including 30% of Republicans.
33% say the U.S. should keep fighting until Iran gives up more.
Anyone remember this?
We’ve come a long way.
Matt Shumer: AI games are going to be amazing
(sound on)
Gokul Rajaram
CHANGE THE DEFAULTS VIOLENTLY
@nejatian, CEO of @Opendoor (previously COO of Shopify), interviewed by @shaneparrish (@TKPPodcast)
Summary: Nejatian took over Opendoor months from bankruptcy and found a company that had drifted, one comfortable lie at a time, into being run by consultants and process. His fix was to change the defaults violently, hire an almost entirely new company, and commit to truth over feelings at every level. Opendoor now buys 6 to 7 times more homes per week than it did a year ago, and the last time it bought at this volume its OpEx was more than twice as high.
1. All Failures Rhyme. Every company that dies, dies the same way. Good people leave, control of innovation slips, G&A climbs, and the company starts picking fights with its customers and partners. Successes are each unique, so they teach you little you can reuse. Failure patterns repeat, which means you can diagnose a sick company much faster than you can copy a healthy one.
2. The Professional Leech. Opendoor's largest external vendor expense went to a big consulting firm, ahead of its cloud provider. That firm had advised offshoring every job, loading up G&A, and cutting engineering, which Nejatian calls exactly what you do if you have zero desire to create alpha. When the founders are gone and no large shareholder cares, a company becomes a host for professional leeches whose job is to drain it while making management look good. Pull the full 12-month vendor list on day one and the diagnosis is obvious.
3. Violent Defaults. Opendoor had run many 7-step, 8-month plans to bring people back to the office, and none of them worked. At 9am on his first Monday, Nejatian told the whole company everyone was in the office the following Monday or no longer employed. Change management defeats the purpose, because you want the change to feel jarring. He rewrote the careers page the same week, replacing "a happy place to work where we all care about each other" with "this will be hard," so the hiring funnel would repel anyone looking for the old company.
4. Competent But Unaligned. The most dangerous people in any company are competent and not mission-aligned. Competent people tilt the org toward themselves, and everyone has worked with the internal politician who keeps getting promoted because the boss cannot see what they are doing. Nejatian's first job was to widen that rift on purpose so those people would leave quickly. Rebuilding Opendoor meant hiring an almost entirely new company, and he says plainly that it was painful.
5. Room Size And Truth. The larger the room, the less space there is for truth. That one dynamic is why big companies are usually average, and it is the real reason two-pizza teams work. A new CEO's most urgent task is finding the speakers of truth and getting them in a room alone. Without that, Nejatian says, you die.
6. Say The Thing. Nejatian's rule, written at Shopify and carried into Opendoor: say it now, say it about the thing and not the person, and say it over and over until you have been heard. People do not have to agree with you, but they have to hear you, and staying quiet gets treated as rude. Meetings at Opendoor are louder and more ornery than average, and walking out of one you cannot contribute to is expected behavior. Meetings are a bug in the system anyway, evidence that people could not get the information they needed to decide before the meeting existed.
7. The Map Is Not The Train. Most executives run their company on dashboards, which are a first derivative of the facts. Build models on models and you drift further from reality every quarter, until you go bankrupt with nobody asking how far they are from the ground truth. Nejatian talks to customers weekly, visits homes constantly, reads the raw database, and builds his own dashboards. Watch Undercover Boss and the CEO is always surprised by their own company, which is the tell.
8. Friction Is Underestimated. Everyone has seen the supply and demand curve, and almost everyone underestimates how much friction moves where the two lines cross. The market size for Walmart, Amazon, and Google all looked far smaller before those companies existed, because friction was hiding the demand. At Opendoor, 11 people sat between a customer typing their address and getting an offer. Removing them raised demand every time, and Opendoor now buys 6 to 7 times more homes per week than it did a year ago.
9. Do Things, Tell People. Toby Lütke put "do things, tell people" on the walls at Shopify, and most companies run that order backwards. When an Opendoor home had no power, the fix ran through a Salesforce form, an email, a manager's approval, and 3 or 4 people. Nejatian gave everyone corporate cards and a Slack lookup that returns the utility phone number, the account number, and the card number for any address. The chain of command exists to reduce risk, so the fix is reducing the number of people who can say no.
10. AI Collapses Management. AI lowers the need for human management by dropping the communication barrier, which collapses the long tails of managers managing managers before anyone does real work. It also decides who wins and who loses, and there will be more winners. Opendoor has more engineers today than when Nejatian joined, because the company uses AI well. He describes it as an AI exoskeleton around every person at Opendoor, making them 3 to 4 times more efficient than the competition.
11. The First Derivative. Opendoor takes thin margins on each home and makes its real money on mortgage, insurance, title, and escrow. Shopify sells software for $1 because credit cards will not process less, and earns from merchants succeeding on it. Google makes its money showing you ads while you search. There is a word for businesses that take all your money upfront and never speak to you again, and Nejatian calls them carnies.
12. Excellence Is The Capacity To Take Pain. The person who holds their hand over the fire longest tends to win. The gap between doing something extremely well and doing it poorly is small, and people mostly give up too soon. Nejatian traces his own tolerance to fleeing Iran as a teenager and getting into no clubs, then playing rugby in North America where nobody watches, so you bleed for teammates instead of an audience. The sugar rush of easy satisfaction in modern life has under-prepared people for how much pain they can actually absorb.
put chatgpt to work
Sam Altman: chatgpt work is remarkable, and "work" undersells it.
from my phone i sent:
"use all my chat history to figure out ideas for a long weekend trip with 8 friends, plan the best three options, make a full-stack site where the 9 of us can coordinate on what we would want to do in
Dan Shamir
Used @mattshumer_ prompt with some chages and wow. Did it deliver .
Git here:
https://github.com/djtoon/sunburst_isle
And prompt is there too.
Paul Solt
Learn how Jason uses Codex at OpenAI.
@jxnlco
- Email triage
- Priorities for the day
- New skills
Follow Peter for more great interviews from AI experts. @petergyang
Peter Yang: “The only job left is to understand what you don't like, put it into words, and tell the AI.”
Here’s my new episode with @jxnlco (DevEx at OpenAI) where he showed me how he uses Codex throughout his workday, including how to:
→ Set up a chief of staff for Slack and email
→
Jack Smith
😵 After 3x 5-hour windows on the $100 Claude Code Max plan (~$600 in tokens if not on the Claude sub), this is what I have using @mattshumer_ 's prompt! The only additional prompting I did was to wrap up & write progress to memory as it got near usage limits, & to pick up from memory to get going again on a fresh limit. This was a total of about 6.5 hours of Opus 5 working on it across the 3 sessions.
Definitely not perfect in this state by any means, but also definitely impressive! I can see how feasible it would be to keep going until it's as nice as Matt's example if I was on the $200 plan. I'm tempted to move up to that plan to try a different game, but I'm actually using the tail end of my canceled Claude sub & I don't want to resub right now (enjoying Sol, & I appreciate how OpenAI treats their customers -vs- Anthropic).
Props to Matt for the killer prompt. I thought there was no way that it produced the results that he got. I was wrong!
Jack Smith: I am giving @mattshumer_ 's prompt a shot using the $100/mo Claude Max plan. So far I'm on 5-hour limit #3. Each 5-hour limit is ~$150-$200 worth of tokens. Let's see what kind of results I get!
My favorite part of all of this is seeing the hordes of people saying “there’s no way his prompt created that game”, then running it to prove me wrong, and ending up with awesome custom games :)
Raphael Schaad
Never meet your heroes, they say —
But what if your hero is Susan Kare? 😊💛
Had the most wonderful conversation after she got off stage at @ycombinator Startup School. Full interview to follow.
In the meantime, pictures of her talk in thread 🧵
Kenneth Roth
A federal appeals court rejects Trump's efforts to limit mail-in voting. Trump sees that Republicans are likely to lose big time in the midterms because of his abysmal rule, so he is trying to game the election. https://trib.al/8XfYZFX
Jared Friedman
The photos you see of Startup School are mostly the big talks. But we got feedback last year that people wanted more small-group interactions with YC partners and top founders. So, this year we ran hundreds of small-group sessions on every topic imaginable.
Bilal Zuberi // Red Glass Ventures
All I see is some of the smartest kids from around the country, and around the world, watching tech legends in person…and likely walking away thinking: “this can also be me one day, but only if I at least start building something somewhere soon.”
Just the energy we need. Scale isn’t perfect. But it’s amazing when it works.
Varunram Ganesh
Summary of people hating on YC Startup School
ChatGPT is increasingly becoming your personal AGI
Tibo: Let ChatGPT *work* for you. How many time have you wanted to negotiate your internet bill, get rid of all those spam emails you’re subscribed to, or find the perfect deal for something you wanted to do or buy. It’s quite literally just one prompt away, all from the comfort of
Brad Wardell
I use Opus 5 with Clairvoyance to make Claude of Duty using @mattshumer_ 's prompt. Single shot. 120fps at 4k.
Agustín
This is what surprised me most, got to have two small group conversations with YC visiting partners
Jared Friedman: The photos you see of Startup School are mostly the big talks. But we got feedback last year that people wanted more small-group interactions with YC partners and top founders. So, this year we ran hundreds of small-group sessions on every topic imaginable.
Eric Bahn 💛
Follow @mattshumer_, he will teach you very cool things!
Matt Shumer: This is what good prompting looks like
I don’t make the rules 🤷♂️
"Turn this thread into a heartbeat. Check my emails, Slack, and Linear at 9 am, 1 pm, 5 pm, and tell me what I need to prioritize."
From @jxnlco (DevEx at OpenAI):
"Just by doing this, [Codex] will give you a pretty good overview of your day. Then you layer on more preferences over time.
For example, when it first ran, it didn't include links, so I asked for links. Then I had it use Linear. Now it pre-drafts my Slack replies and emails for me."
📌 Watch the full episode here: https://youtu.be/ydTeb_I0b94
Peter Yang: “The only job left is to understand what you don't like, put it into words, and tell the AI.”
Here’s my new episode with @jxnlco (DevEx at OpenAI) where he showed me how he uses Codex throughout his workday, including how to:
→ Set up a chief of staff for Slack and email
→
Peter Yang
"Turn this thread into a heartbeat. Check my emails, Slack, and Linear at 9 am, 1 pm, 5 pm, and tell me what I need to prioritize."
From @jxnlco (DevEx at OpenAI):
"Just by doing this, [Codex] will give you a pretty good overview of your day. Then you layer on more preferences over time.
For example, when it first ran, it didn't include links, so I asked for links. Then I had it use Linear. Now it pre-drafts my Slack replies and emails for me."
📌 Watch the full episode here: https://youtu.be/ydTeb_I0b94
Peter Yang: “The only job left is to understand what you don't like, put it into words, and tell the AI.”
Here’s my new episode with @jxnlco (DevEx at OpenAI) where he showed me how he uses Codex throughout his workday, including how to:
→ Set up a chief of staff for Slack and email
→
Artificially Inclined™
Re @ChrisGPT @mattshumer_ I was surprised to see people doubting him/his prompt
The man prompts so hard he literally built a site dedicated to it 🤣
Paul Graham
We have multiple reasons for wanting to believe nurture is more important than nature and none for wanting to believe the opposite (unless we = the far right, which we ≠), so it's not surprising we overestimate its effect. But man are kids' personalities inborn.
Lina Colucci
My favorite reactions from the @ycombinator Office Hours Simulator:
"It was so good I had to whip out my @meetgranola note taker in the middle so I wouldn't forget"
"It felt way more natural to have a face-to-face conversation than with @ChatGPT."
"I'm actually blown away... it was so good!"
"I came in with a general startup idea and the conversation helped me clarify it into actionable next steps"
"The real-life YC Partners are always swarmed by people. I got to ask the AI Partner all of the questions that were on my mind"
Amazing work @snowmaker 👏 And we're so honored to power these live avatars at @LemonSliceAI!
eve: jared built the office hour simulator himself.
he definitely won’t want me to say it because he’s too humble. he did it while also advising yc companies and leading the software team at yc.
it’s hard to describe how deeply he cares about bringing joy to builders. we both draw a
Aaron Epstein
“At the time we thought ‘it can’t be that hard’” @patrickc on starting @stripe. 17 years later, still more to do.
Adam
Re @mattshumer_ @chrisbward Can confirm I’m trying a modified version of this prompt for an RPG and it’s been cooking for over 24 hrs and the preview screenshots look great. Subagents on loops completely changes what’s possible with one-shotting
Michelle Marcelline
Look, Ma, I’m on the Chase Center screen for @ycombinator Startup School!!!!!!!
This brings me to a question I get all the time:
𝗪𝗵𝘆 𝗱𝗶𝗱 𝗜 𝗱𝗼 𝗬𝗖 𝘁𝘄𝗶𝗰𝗲 𝗮𝗻𝗱 𝗴𝗶𝘃𝗲 𝘂𝗽 𝗮𝗻𝗼𝘁𝗵𝗲𝗿 𝟳%?
When we applied with @promptingco, I had already gone through YC with @typedreamHQ.
I already had the alumni network, Bookface, the partners, and the community. Why give up another 7% for access I already had?
Because having access to the YC community is not the same as building inside a YC batch.
The advice itself is not secret:
Talk to users.
Ship quickly.
Do things that don’t scale.
The value is doing it alongside hundreds of founders at the same stage, wrestling with the same questions:
Are users desperate for this, or just being nice?
What can we ship this week instead of perfecting for another month?
How do we get our first customers?
Is this enterprise deal real, or a 6-month distraction?
You share experiments, customer introductions, failures, and wins in real time.
Then a batchmate at your stage hits in months what you thought would take a year.
Reading about a startup reaching $1M ARR is inspiring. Watching your batchmate do it eliminates your excuses.
That is what I gave another 7% for.
Not the logo.
Not information.
Not access to a network I already had.
The environment. The accountability. The higher standard for speed and ambition.
Was YC worth it for us? Absolutely. Was it worth doing twice? Also absolutely.
𝗜𝗳 𝘆𝗼𝘂’𝗿𝗲 𝗮𝘁 𝗬𝗖 𝗦𝘁𝗮𝗿𝘁𝘂𝗽 𝗦𝗰𝗵𝗼𝗼𝗹 𝗱𝗲𝗯𝗮𝘁𝗶𝗻𝗴 𝗬𝗖 𝗼𝗿 𝗻𝗼 𝗬𝗖, 𝗱𝗼𝗻’𝘁 𝗼𝘃𝗲𝗿𝘁𝗵𝗶𝗻𝗸 𝗶𝘁. 𝗔𝗽𝗽𝗹𝘆.
The goal isn’t to own the largest percentage of your company. It’s to build a company worth owning.
Photo credit: @eve_bouff @MarchaRusli
Ahmed Karrar
Biggest takeaways from @patrickc CEO of @stripe at @ycombinator startup school
> ITS THE BEST TIME TO BUILD!!
> He dropped out of MIT twice and mentioned that people think it’s a big risk to drop out of college but he mentioned that it’s really not as bad as you think.
> It took them 2 years to launch stripe because they they wanted to PERFECT the product and make sure that user’s first impression is good since they are in the fintech industry before it really exploded but now with AI you can reduce the shipping cost and start perfect products much faster!
Thanks @ycombinator!!
I have a fucking crazy idea for a new game based on this tech.
I just kicked off an Opus 5 loop that will run on my 20x Max sub 24/7, completely autonomously, to build it out.
If you want first access, add your email here: https://shumer.dev/newsletter
Matt Shumer: Claude Opus 5 one-shotted this game.
EVERYTHING you see in this demo is custom code... not a single external asset was used.
AI games are going to be amazing.
(sound on)
brett goldstein
YC filling a stadium for startup school is good:
1. startups create jobs & products that improve people's lives
2. YC startup school helps people create startups
3. more popular YCSS = more startups
4. more startups = more jobs & products that improve people's lives
Yogi Suria
Claudepunk 2077 - Single prompt, opus 5. This is threejs but imagine claude doing its magic using unreal MCP with metahumans, i think we can just churn out games now. github - https://github.com/kalaanakonda/Claudepunk2077/
inspired by the post from @mattshumer_
Matt Shumer: Claude Opus 5 one-shotted this game.
EVERYTHING you see in this demo is custom code... not a single external asset was used.
AI games are going to be amazing.
(sound on)
max
“with agents, startups vs incumbents is no longer david vs goliath, it’s more like mecha-goliath vs goliath”
-@alexandr_wang at startup school
Y Combinator
NVIDIA started with the wrong technology, learned the right one from three textbooks bought at Fry’s, and went on to invent most of the major breakthroughs in modern computing.
At Startup School 2026, YC's @garrytan sits down with @nvidia Founder and CEO @JensenHuang to talk about confronting reality, learning your way into new domains, and why resilience — getting through one day at a time — matters more than anything else.
01:07 — NVIDIA's wrong algorithm
03:14 — Buying textbooks to save the company
05:29 — The real big idea
07:06 — The Sega story
09:37 — The $300M IPO
10:33 — Seeing AlexNet differently
12:44 — Reinventing the full stack
13:30 — How to build a first-principles org
17:01 — Build the car to fit your driving style
19:12 — Frontier algorithms
20:50 — Systems thinking is the new coding
23:26 — Should you own your own AI?
25:43 — How NVIDIA sses agents internally
30:51 — AI and job creation
34:21 — The ChatGPT moment for robots
37:24 — Where physical AI shows up first
39:18 — Why Jensen just joined X
41:01 — What to learn that still matters
44:32 — The mindset you should have - "How hard can it be?"
RK
“meta will be launching a harness and open-source models” - alexandr wang
cole
I'm a YC F25 founder and last year I got rejected from YC startup school
I honestly felt a little spiteful seeing every single one of my friends get in, and I asked myself: what was wrong with my application, what was wrong with me
Fast forward to today I spoke at the Chase Center on the future of agentic payments
I think getting rejected gave me the kick in the ass and the courage to finally apply to YC, maybe not the best source of motivation but a source regardless
So much has changed in just 1 year and would not change a single thing if I were to go back
My point: what stings the most can also be the thing that moves you
Mission accomplished?
Matt Shumer: Okay I may have been extremely wrong about Opus 5...
Holy fuck, this demo is going to blow the entirety of AI Twitter's minds tomorrow
Ayush Sharma
“Develop your own internal compass about the future”
~ @alexandr_wang
Sam siavoshian
Things you see at YC SUS
Lina Colucci
There's been a non-stop line of people waiting to talk to the AI avatars at the Office Hours Simulator @ycombinator #StartupSchool
Eric Zou
My 2 takeaways from @steipete, creator of @openclaw, and someone who held true to his soul.md and ethics.md, at @ycombinator startup school.
1. what’s most important is the joy of building. This is probably the single highest value advice amongst a room full of future founders.
Back in February, I fall into the trap of wanting to become a founder. Applied to every possible VC fellowship I could find, and nagged my entire network to try out my product.
A product for a problem that I didn’t love.
Peter was what most would count successful, and he retired. He stayed away from the laptop for months and just chilled. Then open claw came only because he found it to be fun.
2. Stay focused. hype is like the weather: you might see it coming, but you can’t control it. Adoption is climate.
I’m someone who loves to work in parallel. Tmuxing 8 agents across 2 different repos at my internship, or air frying the chicken while i shower, or learning and doing multiple side quests at the same time.
Peter is the opposite. He’d get mad if you go to him to propose a feature.
——-
Some other takeaways I can’t resonate yet
- His mistake: once you stop dogfooding what you ship everyday, it gets shit.
- what’s next for openclaw: collaboration between devs holding black boxes before the PR layer
Richard Sutton
Now more than ever.
Richard Sutton: Free Palestine.
Jake Mintz
A lot of people identify YC with the batch program, which is important. But to me the greatest power of YC is making startups accessible and a nudge to technical people that wouldn’t have otherwise have been founders. The first YC batch were mostly students who attended @paulg’s “How to Start a Startup” at the Harvard Computer Society in March 2005.
And when I look at the under appreciated impact of @jesslivingston it includes the amazing events like the first Startup School in 2007.
To me, as a summer 2009 founder, @garrytan has done an incredible job carrying this torch forward.
The rest of VC fights over the startups that would have happened anyways. YC helps create more startups more than any other institution in the world.
Scott Stevenson: I have great respect for YC, but need to say what many are surely thinking:
Startup School in a stadium feels antithetical to the YC brand and to what makes startups work.
The YC brand felt like it was designed to filter out status chasers.
The bare Mountain View office. The
Blake Scholl 🛫
Experts: jet engines are so complicated only a handful of gigantic companies can make them
Boom team:
Joe Hu
> be joe.
> always build for myself and be my own first user. even the product name has “joe” in it.
> build to solve my own problems, not to optimize for yc or fundraising.
> twitter is still on my phone dock.
> on my own track.
Thank you @steipete !
Philip Johnston
Sage words from @MaxJunestrand of @WeAreLegora at @ycombinator startup school:
“One under appreciated form of negotiation is nagging… we nagged and nagged… that’s something my dad taught me!” 😅👌
Games are arguably the type of software that humans most want to play and use themselves vs. delegate to agents.
Julian Togelius
Right now we are defending open source (or open weights) models' right to exist. The debate is not about forcing anyone to open-source their models, whatever some other Julian might imply. However, what if we did just that? Last year, I explored that scenario.
Lion’s Den is actually the jam
Sheeta Verma 👩🏽💻: “SF is dead”
no you’re just attending luma events and twitter community events
literally every Wednesday night in Chinatown there’s a Hip Hop & RnB night at Lions Den.
People are literally dancing like it’s the club on a Wednesday!!!
SF isn’t dead, it’s gatekept to avoid
agreed feels big, i want a new kind of computer
Alex Finn: I don’t think you understand what ChatGPT Voice unlocks
I went on a 4 hour hike yesterday. Through the northern California redwoods.
Got more work done in those 4 hours talking to Voice through my AirPods then I do in 8 hours at my desk
Work can be done ANYWHERE now
45 minute
I compiled 𝚟𝚎𝚛𝚌𝚎𝚕 CLI TypeScript to native with 𝚜𝚌𝚛𝚒𝚙𝚝𝚌. Incredible.
✓ Resulting binary size: 1.28mb
✓ Startup overhead: 1.5ms (mean)
✓ Compiles in 2.94s (mean)
✓ Uses 𝚗𝚘𝚍𝚎:𝚑𝚝𝚝𝚙𝚜, 𝚗𝚘𝚍𝚎:𝚏𝚜, 𝚗𝚘𝚍𝚎:𝚙𝚊𝚝𝚑, 𝚗𝚘𝚍𝚎:𝚘𝚜, 𝚗𝚘𝚍𝚎:𝚌𝚛𝚢𝚙𝚝𝚘
✓ Highly readable TypeScript
✓ Code translated by GLM 5.2 Fast
✓ Deploys like a charm!
No embedded v8 / QuickJS, fully static.
Chris Tate: New experiment: scriptc
A TypeScript-to-native compiler
→ Written in TS
→ Static by default
→ Type checked by tsc
→ Byte-for-byte Node.js compatibility
→ Falls back to QuickJS-NG for dynamic
Goal: compile existing TS/JS with no code changes
Junfan Zhu 朱俊帆
🔥Hot takes from @chelseabfinn @physical_int at @ycombinator startup school 🤓
Junfan Zhu 朱俊帆: 🤓 @ycombinator Startup School
LemonSlice
How does the YC Office Hours Simulator work?
1. Walk up to a booth and press "start simulation"
2. Get randomly matched with an AI YC partner. They know who you are and what you're working on (based on your application)
3. Talk about whatever's on your mind. Each Partner is trained on hundreds of hours of how they normally run their Office Hours (video features AI @xuster. photo features AI @garrytan)
That's it!
Built by @snowmaker and the @ycombinator Software Team for #StartupSchool 👏 @sumanyu @bryantchou @sidney_primas
Vivian Midha Shen
my crew of intense Bay Area high school kids at startup school - representing Bellarmine, Monta Vista, Lynbrook, Harker and Gunn of course!
#ycstartupschool
Krishna
Best talk of the weekend by far @maxhodak_
I was expecting a talk fueling the craze of BCI x AI interaction, especially since it was at YC SUS, but it was genuinely useful
Key takeaway: infrastructure is speed. All the boring stuff matters.
