2026-07-27

Highlights from Chelsea Finn’s (@physical_int Co-Founder) talk at YC Startup School:
> RL for robotics is bottlenecked by physical rollout cost, not just algorithm quality. 1M trajectories of a 1-minute task would take ~700 robot-days, which makes direct scaling of PPO/GRPO-style methods fundamentally different from LLM post-training.
> Memory is still a major missing piece in robot foundation model: most SOTA systems have effectively zero memory, and naive video history is too expensive (10s at 50Hz across 4 cameras ≈ 500K tokens). π solves this through short-term video memory plus long-term text summaries of the prior 10–15 minutes, which enabled a fully autonomous multi-step kitchen cleaning task.
> Generalist robot models may already be beating specialist pipelines: π0.7 mixed demos, rollouts, human video, and web data, then used rich prompting + metadata to make even lower-quality data useful — and it matched/outperformed fine-tuned specialists while showing compositional transfer to new tasks/platforms.
@chelseabfinn Thank you for the great talk!!

sam altman and garry tan (two yc presidents) are in the building! the applause felt more like a concert than a tech event

Ran a breakout session with the most cracked builders in 'Consumer' at Startup School today.
Founder from Taiwan that's building a unique new photo sharing app already with 100k's of active users on how they grew:
"In Taiwan nobody saves the phone numbers in http://Contacts.app (🤯), because we don't use iMessage, so we can't recreate the social graph from contacts, but have to reconstruct it from viral loops with sharing links."
I love the Art of Growth in Consumer.

Re @thsottiaux can we give this man the extra Codex usage? We all want to see it.
Matt Shumer: 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

sam with the last keynote of startup school!
what an epic weekend. now back to building.

“The next six months is going to be like the last 2 years of model progress”
- @sama at YC Startup School

i don’t think it’s outlandish to say that the collective economic output of this stadium over the next decade will probably exceed many countries’ GDPs
josh: the great ai psychosis hivemind of San Francisco

YC Startup School 2026, the epicenter of tech today

omg hi sam @sama
wrapping up @ycombinator startup school with @garrytan and @sama 🔥🔥

Biggest takeaways from @sama CEO of @ChatGPT at the YC summer school!
> Believe is everything and it all starts in the mind! Paul Ghram have always believed in them and told them that they would be a big company like Google or Meta so believe is everything.
> The economy would be stagnant if it was not for startups and enjoying startups would help you excel a lot faster both as a startup founder and an investor!
> YOU DONT’T HAVE TO JOIN A FRONTIER LAB in order to build a startup.
> Startups that are created today will be much more valuable than startups that are current (so start building!)
> Get fluent with the new AI tools as that is a huge competitive edge that you can take on and have!!
> Find a way to be helpful to people.
> You are going to be dealing with all of the haters all the time especially in social media but you should put ALL OF YOUR ENERGY into building stuff and ignoring the noise.
> “being a solo founder is lonely”
> The only way you can help and contribute in the economy is by building a successful startup!
> Creating super intelligence would most likely
> “YOU CAN NOW DO 3 MONTH OF WORK IN 7 MINUTES”!!!
Message he would give to himself at 20s:
> IT IS ALL GOING TO WORK OUT.
> Be happy and enjoy the journey along the way!!
Models:
> Next 6 month will be equivalent to the last 2 years in model development!
> SO IT IS THE BEST TIME TO BUILD A STARTUP!!
I hope I at least help one person get clear on what direction they want to take over the next 6 month and I would love to connect with every single person reading this and help with any shape or format 🤍

one of the best talks i’ve heard at @ycombinator sus of @garrytan and @sama.
sam altman: “It is very hard to run YC or run a startup... It is very easy to go take shots on Twitter... It will poison your soul."

Re @sama in @ycombinator Startup School:
“Startups are the thing that keeps the economy from becoming stagnant.”
“‘Permanent underclass’ is stupid. Startups created today will be much more valuable than past ones.”
“Great startups cluster when technology is moving quickly, costs are coming down, and cycles become shorter: i.e today.”
“Find things you have conviction in and think conventional wisdom is wrong, and have patience.”
“The world doesn’t intuit exponential. There are exponentials right now many don’t know about.”
“Right now is the best time to start a startup.”

Re @sama has a word to say to the haters of yc: just stop.
there was surely no better way to end this year’s edition of yc sus, and thank you @garrytan

thanks papa @bcherny

The Gang Takes Over The Chase Center

probably my favorite moment from yc sus was yesterday’s talk with @JensenHuang, especially his remarks on the enduring importance of oss:
“if not for open source, the mobile cloud industry would have never happened. if not for open source, if not for linux, if not for kubernetes, if not for all of these, platform, if not for tensorflow or more importantly pytorch…and the early versions of it, cafe, torch, theano, remember the early versions [pytorch] of all those were all open sourced. if not for all of that, how would we have modern ai?”

this is real for anyone doubting. yes, a one shot.
a lot of shit is fake on x. not too hard to detect what's real -- i still get fooled sometimes though.
if you dont believe him run the prompt urself. its not fucking rocket science.
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)

getting some interesting visual results with a variant of @mattshumer_'s ultracode prompt + Opus 5
agents looping for 9+ hours so far, will report back on the gameplay..
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

The deadline to apply for the YC Fall 2026 batch is tomorrow, July 27th, at 8pm PT!
Apply at http://ycombinator.com/apply.

You can’t watch this and say @ycombinator startup school is anything other than earnest, technical, builders 👌
STJ: I interviewed 20+ YC @startupschool attendees. Here’s what they had to say:

Re @ZavalaA has hit the nail on the head. The reason Jane Kim refused to directly answer the question about her experience is because she has none. I ran against both @JaneKim and @BenAllenCA in the primary. Ben Allen is qualified to be Insurance Commissioner. Jane Kim is not.
Ashley Zavala: Jane Kim is running to be California's next Insurance Commissioner amid the state's insurance crisis.
I asked her repeatedly what experience she has with the industry.
She wouldn't directly answer the question and instead explained what she thinks IC is supposed to do:

some takeaways from @sama today at @ycombinator startup school
• this is one of the best times in history to start a company. ai is lowering costs, shortening build cycles, and expanding what’s possible.
• the next 6 months of model progress could be comparable to the last 2 years.
• the best startups often start with ideas experts think are bad.
• develop conviction despite conventional wisdom, and be willing to wait.
• vision can be clear even when the first steps aren’t.
• if everyone believes your idea, that’s a bad sign. if no one believes it, pay attention to that too.
• be mildly helpful to lots of people. it compounds.
• the best cofounder relationships are often built years before the company starts.
• earnestness beats cynicism. spend your energy building, not taking cheap shots online.
• startups help prevent the concentration of economic and technological power.
• demand for high-quality, low-cost intelligence is still largely untapped.
• have the ambition to build, but be a little happier along the way.

Sam and Patrick both said the same thing today: be genuinely passionate about what you build and know that you’re in it for the long haul and not the status.
Thank you @ycombinator @garrytan and the whole team for an illuminating YC Startup School!

while so much is encouraging, it’s also just so sad to me
you can see in realtime a person actively rotting their soul and their humanity, all because they’ve been convinced that being an asshole on the Internet is cool
i pray for them
spor: So interesting to overhear and speak to a bunch of young talented kids these weekend. Frankly, you can tell who’s gonna make it and who’s not, very quickly.
The earnest, excited young builders? Gonna do so well. You can just feel it. For the rest, I just really really hope they

… aaand this is a wrap at Startup School 2026!
From the outside, YC might seem intimidating. I had to remember many young builders this weekend that we're just a small group of nerds that try to help other earnest builders.
Every single partner was a founder themselves. We have all been in the trenches and just want to pay it forward. There's no big machinery or gaming the system. We just want to help the next generation to build iconic companies.
We work extremely hard to read all the emails. We see your Tweets. We know what 'good' looks like. Build good things and we'll notice. We'll share the talks on YT. We hope this weekend was and will be as inspiring to all of you as it has been to us. We're looking forward to amazing applications for the Fall '26 batch and beyond.
Live in the future, then build what's missing.

Ultimately, the Twitter trolls will always lose and will never do anything of substance in the world, and it’s nice to know that, but also it’s sad. Plenty a good kid has been sucked into the cesspool and depressing nature of social media algorithms that encourages you to be a loser and hate on things and dedicate your essence to spreading negative energy. But in the end, the person you’re hurting most is yourself.
Kaya Jones: What @sama hated most about running yc lol

best part of @ycombinator startup school this year: we ran around 100 parallel small group sessions, ranging from intimate gatherings with top founders to our inaugural ML research symposium "YCML"

Ouch. Jane Kim exposed.
California, just vote for @BenAllenCA
Ashley Zavala: Jane Kim is running to be California's next Insurance Commissioner amid the state's insurance crisis.
I asked her repeatedly what experience she has with the industry.
She wouldn't directly answer the question and instead explained what she thinks IC is supposed to do:

you’re the average of the people you surround yourself with

My own version is loading...all the rumours are true!!! All built in @threejs 😱
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)

Music-JEPA: Learning a World Model of Sound from Action
"we propose to learn a world model of piano sound using JEPA by framing music as an action-conditioned system: the audio is treated as the state, and the pianoroll as the instrument action."
"Experiments show that the learned model captures the relationships between musical actions and their resulting sound. The resulting representations support downstream tasks, including beat tracking, composer identification, and key estimation, and enable piano transcription via planning, by searching for actions that best explain a target sound."

Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community.
During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion.
That’s why we created the Open Secure AI Alliance.
NVIDIA: AI security advances when the industry builds in the open, together.
We're introducing the Open Secure AI Alliance with industry leaders to develop new techniques and tools to safeguard software and agents.
By sharing models, tooling and research in the open, we can broaden the

Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community.
During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion.
That’s why we created the Open Secure AI Alliance.
NVIDIA: AI security advances when the industry builds in the open, together.
We're introducing the Open Secure AI Alliance with industry leaders to develop new techniques and tools to safeguard software and agents.
By sharing models, tooling and research in the open, we can broaden the

If multiple Mythos level models are out there in the hands of attackers, now or in a few years, then US builders and software developers need strong, open defenses. They cannot have gated, crippled defenses that refuse to help while attackers run roughshod.
We need to patch our systems not bury our heads in the sand and hope that nobody had powerful offensive AI. That's abusrd.
The answer is powerful AI patching our systems, not refusing the patch our systems.
Let developers find issues and secure them at the source.
And if you imagine that some kind of magical legislation in the US will somehow prevent the rest of the world, APTs and hackers fine tuning or training models or attacking closed models to use them in attack, then you are beyond delusional.
We must have power tools widely distributed and widely available to all, not central Soviet style controls and steering committees.
Jensen Huang: Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community.
During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion.

Today's my first day in a new role: Head of Editorial at Latent Space (@latentspacepod).
From ReadWriteWeb to The New Stack, I've always loved being at the cutting edge of tech x media. So I'm thrilled to join @swyx & his team. We have big plans! https://www.latent.space/

“stripe wouldn’t exist without yc”
- patrick collison
stripe powers 5m+ businesses and processes $1.9 trillion every year, about 1.6% of global gdp.
big things start small. yc deadline is today.

Elon’s missed promises 😳

Earnest builders win.
spor: So interesting to overhear and speak to a bunch of young talented kids these weekend. Frankly, you can tell who’s gonna make it and who’s not, very quickly.
The earnest, excited young builders? Gonna do so well. You can just feel it. For the rest, I just really really hope they

Matt Shumer was right.
I don’t game. I’m almost a boomer. And I only learned to code three years ago with AI.
But for my son’s birthday, one prompt changed my view of AI gaming completely.
My son has been a gamer his entire life. He already has everything, so I kept asking myself:
What do you buy someone like that for his birthday?
Then I saw @mattshumer_’s post about Claude Opus 5 one-shotting a playable game.
I decided to try his prompt, tweak it slightly for my son, and see what happened.
For context, I tried something similar a long time ago with Opus 4.6. It took a ton of effort, and the result was garbage.
This time?
One prompt.
A 21-hour run on Ultra.
About 20% of my weekly plan burned.
And at the end of it, my son had a playable game built specifically for him running on the 85-inch TV.
He was blown away.
This isn’t someone being impressed because a character moved across a screen. He knows games. He has played them his entire life.
Now he is playing something tailored to him, enjoying it, and already talking about how we can make it bigger and better.
Is it perfect?
Of course not.
But that isn’t the point.
The point is that an almost-boomer who doesn’t game, learned to code with AI only three years ago, and used a single prompt to create something a lifelong gamer genuinely enjoys.
There is something here.
Something big.
You can criticize the imperfections. You can argue about whether “one-shot” counts when the run takes 21 hours. You can complain about the compute or the cost.
But sticking your head in the sand won’t slow this down.
I see the same resistance every day from small businesses. People are so focused on what AI can’t do or fearful of what it might do that they never take the time to discover what it can already do for them.
The businesses working with us at Swaymak are taking the opposite approach. They are experimenting, learning, and riding a wave of extraordinary new technology.
And on this one, Matt absolutely called it:
Opus 5 and AI are going to change gaming in a very big way.
Today, though, this isn’t really about technology.
It’s about watching my son enjoy a birthday gift that didn’t exist yesterday because I was willing to try.
Screenshots attached. Video coming. I’ll put the project on GitHub soon.
Then I’m getting back to helping businesses take advantage of the biggest technological change of our lifetime.

Catalog of incorrect AI predictions.
Doomers from the rationalist cosplay community are left out, presumably because their predictions have a near 100% failure rate.
When people proudly proclaim that we need to 'act now' or make laws because something is inevitable, point them at this list and tap it loudly because it ain't.
https://boyswhocriedai.lovable.app/archive

http://x.com/i/article/2080707729870430209

i read the whole long leaked DeepSeek investor update over the weekend. it is greatly under discussed. it lays out the entire thesis & playbook for the lab down to the number of chips they have. the main idea is “restrained gratification”.
he thinks they will achieve their sole goal of AGI sooner or later. they need to restrain themselves from losing focus in the meantime, which means giving up the other things that don’t matter (like a good user experience) and maximizing profit; seems like he’s also given up his ego based on how he frames his role. no public exposure.
he thinks there’s no talent difference as all labs are full of Chinese people, just that some decided to stay in China and others immigrated to America. he thinks the only difference is resources: access to chips, but because of this constraint, they’ve been forced to optimize on local chips, which are currently a production constraint but quickly catching up.
he emphasized his biggest fear which is not chips but stability of his most tenured employees.
so based on this it should be quite obvious how to play if you were to compete with them…

My 5 biggest takeaways from @jxnlco (DevEx at OpenAI) on how to use Codex for work:
1. Ask Codex to mine your past sessions for new skills and workflows
Ask Codex to review all your threads from the last 30 days or so and suggest new skills and workflows for the tasks that you repeat the most often.
Prompt and more below 👇
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
→

thx to yc for flying me out (or back) from oxford for my first yc startup school!
what i learnt:
> don’t LARP! - @garrytan
> nowadays with coding agents allowing literally anyone to build so fast, you need eyeballs on your product. what sets you apart is traction - @steipete
> build what annoys you. and have a lot of patience because it will take a long time. be ok with people thinking your product is dumb, find a group of people who believe in it and build together @sama

I asked @jxnlco (DevEx at OpenAI): “What’s the craziest thing Codex did for you?”
Here’s his answer:
“I was on a bike ride when a coworker asked me to fix a launch video. So I connected remotely from my phone and asked Codex to use computer use to edit the video, export it, and post it back to Slack.
Twenty minutes later, there was a new video in the Slack thread. Then I told Codex:
‘Every 30 minutes, check the thread for feedback and export V2, V3, V4.’
By the time I got home, the launch video was greenlit.”
Codex had handled the revisions by repeatedly checking Slack for feedback.
📌 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
→

Jensen Huang: The narrative about AI destroying jobs is exactly backwards. AI eliminates tasks. AI is creating jobs.
Source: @ycombinator

The hard part of agent loops is not starting them!
It’s knowing when to move up a layer, and when to come back down. @swyx explains the operator skill behind useful loops: automate repeatable work, watch for failure, drop into the problem when reliability breaks, then climb back up once the system is sound.
That flexibility (not token volume for its own sake) is the key to productive loops.
Join us in the Token Billionaires Lounge, presented by
@cerebras and @aiDotEngineer

People love to hate on YC for being too AI-pilled.
Meanwhile your VC can't build an AI agent and YC GPs are out here shipping the ideas in their heads for fun

Introducing Waddle Labs: Claude Code for robots.
Connect our API to your robot and enter a prompt, then our agents write code to achieve the task in 20 minutes.
@yiding_song @theWaddleLabs

YC startup school was honestly so so much better than I expected
fully expected everyone to be shallow and self-centered, but it was so refreshing to see people genuinely curious and inquisitive and excited
taking that sf energy back to indiana
@ycombinator

SF nearly doubled homelessness nonprofit spending from 2021 to 2025, $274M to $487M, and got MORE homeless people for it (7,754 → 7,973). Funny how that sounds like a taxpayer-funded scam.

still hard for me process that we're all actually living through the singularity right now
like this is literally the single most important decade in the sum of human history and we're all alive (for now, fingers crossed on that one) at this exact moment to experience it
Matt Shumer: Really does feel like we’re entering the Singularity.
Unprecedented hacks, models solving decades-old math problems, and more, all in a week.
I created this Singularity Tracker to help everyone see what’s going on, updated as things happen:
https://somethingbig.ai/tracker

San Francisco keeps offering the same false choices—justice over safety, equity over excellence, compassion over prosperity—and Asian Americans keep getting hit hardest, with the least inherited wealth and political power to do anything about it.
Six Asian Americans are dead. An education system failed to educate. And now, people who worked hard for their money risk having it taken away with November's proposed wealth tax.

Right now is the best time in history to start a startup.
And the deadline to apply to the YC F26 batch is tonight. It only takes a few minutes, and some of the best companies in the world applied last minute too.
Apply here: http://ycombinator.com/apply
Aaron Epstein: “This is absolutely the single greatest time to start a company, I’m jealous of all of you” @JensenHuang at Startup School

i've been independently using a very similar loop over the past few weeks for frontend design / development work - using gpt-image-2 to generate a concept art northstar
can confirm that this "gauntlet loop" is extremely effective and lets you get so much more out of any model
Matt Shumer: I’m officially calling this the Gauntlet Loop.
The agent (not you!!) breaks the goal into parts, gives each part a specialist builder and a ruthless blind critic sub-agent, with a mandate to only pass if the generated artifact is better than some real-world equivalent.

today is the YC F26 deadline
357 days ago, @maebert and I applied for the 7th time
here is the actual founder video we submitted. our written application is in the replies
we had interviewed twice before. both times, we got rejected
we had raised money, built a product, and found paying customers
but there were still only two of us. I was living in a converted closet at @mission__ctrl
since then, we’ve grown revenue 20x, started serving Fortune 500 companies, and built a team of 11
YC changed the trajectory of our company and our lives
the deadline is tonight at 8 pm pt
don’t miss your shot

The deadline to apply for YC F26 is tonight at 8pm PT!
If you're an earnest builder working on something people want, we'd love to hear from you: http://ycombinator.com/apply

At Replit, @amasad and leadership have taken this approach with hiring
Hiring people who are so earnest about Replit's product and scaling Replit to become a generational company
The conviction of every team to do their best at Replit astonishes me every day
@kodylow and @nickco are the best examples of this
spor: So interesting to overhear and speak to a bunch of young talented kids these weekend. Frankly, you can tell who’s gonna make it and who’s not, very quickly.
The earnest, excited young builders? Gonna do so well. You can just feel it. For the rest, I just really really hope they

YC W10 I crushed Demo Day and got a total of 9 investor intros. Closed a $680K round that year.
YC S26 we haven't even hit Demo Day. 23 investors have already reached out. Millions committed to a round that doesn't exist yet.
Say what you want about @garrytan's @ycombinator scaling too much. Results are better for founders, not worse.
Lloyd Armbrust: I thought I was done building startups until I met @EmilyColman.
After YC W10, raising ~$100M, and building a $50M FDA-regulated manufacturing business, I wasn't planning to come back.
Then yesterday happened.
The FDA's peptide committee voted 'yes' to legalize peptides.

We applied to @ycombinator 30 seconds before the deadline, had 5 major typos we couldn't edit after, and watched people on X get interviews within days while we heard nothing (thought it was joever).
1 month later, I was booking a highly overpriced, last minute flight to be in SF for <24 hours to interview with @bosmeny and @borisjabes!
And now, we're more than 1 month into grinding out the S26 batch :)
F26 applications are due tonight. Apply, no matter how rough the application is or how late! You never know what kind of miracles happen overnight.
Attached is the 67th take of our founder video in my Berkeley dorm common room (had a resident ask why she heard the same script a million times across 4 days). Your journey can truly start anywhere :p

The deadline to apply for YC's fall batch is today at 8pm PT. Filling out an application takes ~30 minutes.
Aaron Epstein: “This is absolutely the single greatest time to start a company, I’m jealous of all of you” @JensenHuang at Startup School

Lots of folks wondering why this beautiful housing proposal in the Marina matters so much, I realize it's been a minute since we've discussed the outsized impact San Francisco NIMBYism has on US politics and economy.
It's ... not small: A significant share of 12 - 36% of GDP.