Garry's List
You paid $51 billion for California government this year.
There’s no data showing what you got for it.
So we reviewed the closest thing that exists: state auditor reports—independent investigations the executive branch didn’t ask for and can’t suppress.
We found child abuse cases open 18x longer than the legal limit, $10.4B in unverified unemployment payments, and wage theft claims unresolved for five years.
How did we get here?
TBPN
YC Managing Partner @harjtaggar says many young founders are still too conservative with token spend and advises them to not skimp on token budget.
"Token spend is more like rent. You don't say, 'I'm not moving to San Francisco to build my tech startup because the rent's expensive.' Yes, it's expensive, but it's the place to be. You don't conserve on that. You actually want to spend money on that. There's other things you save money on."
"Many of the founders, especially if they're younger, just don't have the budgets to use the latest models and be cost insensitive and token-max. They're just used to conserving and trying to do more with less."
"Internally, we wonder if there's some training to push the founders to spend more on tokens because maybe they're thinking about tokens as something you should conserve spend on."
TBPN
.@paultoo's "$100B question" challenges YC founders to think beyond their first $100M in revenue and envision what it would take to build a $100B company.
YC Managing Partner @harjtaggar says the question reveals how much ambition and long-term vision founders in the batch really have.
"Paul always had this great prompt which was, 'Don't just think about how you're going to be a $1B company with $100M in revenue. Think about the next act after that. How do you get to $100B? Google started off with search and then became so much more.'"
"It's useful to do at this stage because you get a sense for how much ambition and vision the founders actually have. Sometimes it helps them get more excited about their own company."
software engineering is so different now. hard to remember what it was like even 6 months ago.
Honestly I don't want to be a hater but I feel like this is what happens when nobody internally wants to tell the CEO no.
Maybe this thing will blow our socks off?
Nathie @ AWE: Evan Spiegel showing off the new Specs AR glasses to the public for the first time.
Publishing a new tutorial to make Codex or Claude Code your personal advisor using a skill with 4 files. Plus, I managed to save Fable's advice too before it got restricted 🥲
📌 Subscribe to get the tutorial tmr: https://www.youtube.com/@PeterYangYT?subscribe
Peter Yang: For me, the best use case for Fable isn't coding.
It’s giving it a ton of context about my business and life, then asking:
→ What opportunities am I missing?
→ What should I stop doing?
→ Where am I lying to myself?
Having a sharper second brain for business strategy has
Seeing so many people saying Fable wasn’t that much better than other models.
Sure, if you’re just doing normal frontend/backend work that Opus/GPT 5.5 can already do, it’s going to feel basically the same.
The magic comes when you’re asking it to do things other models can’t.
Nicolas Dessaigne
I was 37 when I did YC and it transformed my company
Brad Flora: YC is absolutely YC for 35+ founders. We fund many founders in that bracket every batch and will continue to. If you want to build a huge fast-growing company alongside a peer group of the most ambitious people in the planet, you should apply to YC regardless of age. If your
Diana
woah first time oss models are leading
Design Arena: BREAKING: GLM-5.2 is now 1st on Design Arena.
With an Elo of 1360, GLM-5.2 has jumped ahead of the now unavailable Claude Fable 5.
And it's open weights.
This is an improvement of 4 positions and 27 Elo points to achieve one of the highest Elo scores in our code categories
This aged well 😂 iykyk
𝐕𝐚𝐫𝐭𝐢𝐜𝐨: When is Argentina playing their first World Cup game? Tired of watching these amateurs.
GPT-Realtime-2 is something new
Pat Simmons: GPT-Realtime 2 is the future of the operating system.
I've been experimenting with it for a couple weeks now, and I gotta say, it's pretty gosh darn incredible.
Opening apps, searching the web, even editing in Premiere. All with just my voice.
And it only takes a few prompts
You’ll never achieve anything if you are afraid of being cringe
Then again this is from someone trying to use looksmaxxing to achieve outsize return, good luck with that
Looksmaxx VC: @omooretweets This is so cringe
9Mothers is one of those technologies whose impact is immediately obvious to warfighters
Camilo Acosta: Just an anti-drone artillery gun at YC demo day. Nothing to see here…
It’s time to ship
Han Xiao
zhipu和deepseek在25年春都曾经是jina reader的数一数二的大客户，也都是由我直接founder support。二者给我留下的印象就是非常精，对技术指标要求非常苛刻，动不动就p99 latency要降到多少。沟通基本就是一个微信群要随叫随到，从他们手里赚点钱非常难。zhipu这边买了用了我们服务两个月后，来一个清华来的实习工程师发现我们jina 官网的api token价格比给他们2b的价格便宜不少，觉得我是奸商欺骗他们，我解释说2b价格本来就不一样，其次我直接为你服务调优debug这些什么都不算么？他说不算。我这人ego比较大当时是一肚子火，一方面我觉得zhipu太小气，面对我们一个成长中的startup完全没有任何的支持。另一方面，第一性原则角度来讲他们也没做错什么，无非是想要最低价。他们也没有义务帮生态中的任何玩家。我觉得b2b销售本来就非我所长，自己ego太高放不下，即便放下我做founder sale极度缺乏耐心。再加上jina本身除reader服务之外需要我做大量技术管理，来美国后又忙并购的事情，zhipu这个2b的deal就凉了下来。
Nikita Bier
Re @garrytan https://x.com/nikitabier/status/1695947720547094568
Nikita Bier: My advice to every founder:
Quick one to set the stage:
☑️ 30 minute function invocations
🆕 24 hour sandbox lifetimes
Vercel Developers: Vercel Sandboxes can now run for 24 hours.
▪︎ Run agents without mid-task timeouts
▪︎ Pair with persistence for infinite state
https://vercel.com/changelog/vercel-sandbox-can-now-run-for-up-to-24-hours
Philip Johnston
We did @ycombinator 2 years ago with an average founding team age of 41 and became YC’s fastest unicorn 💪🏻👴🏻💪🏻
YC is the YC for 35+ founder 👌
Brad Flora: YC is absolutely YC for 35+ founders. We fund many founders in that bracket every batch and will continue to. If you want to build a huge fast-growing company alongside a peer group of the most ambitious people in the planet, you should apply to YC regardless of age. If your
AI for helping crack a health mystery. So many stories like this, and a clear motivation to be excited about AI:
Amy Deng: I’m an AI researcher turned brain tumor patient, and recently I used the models to crack my mystery fatigue faster than my PCP could.
I believe everyone can do the same with their own symptoms. Here’s how:
New http://vercel.com/home is live, highlighting the present and future of our platform
It was so inspiring to chat with and learn from Will Shu, founder of Deliveroo.
Few people have been in the arena day-in-and-day-out like him. Founder Mode before it had a name.
Vercel
Every agent needs streaming, models, durability, isolation, channels, and integrations.
The Agent Stack has them all:
▪︎ AI SDK
▪︎ AI Gateway
▪︎ Workflow SDK
▪︎ Sandbox
▪︎ Chat SDK
▪︎ Vercel Connect
𝚎𝚟𝚎 implements them with one command ↓https://vercel.com/blog/agent-stack
The hardest part of building an agent is not building the agent. It’s the data.
It’s, ironically, figuring out OAuth, tokens, credentials, scopes… We have AGI sitting there, waiting to be unleashed.
Vercel Connect solves for both security and ease of use. It’s delightful
Vercel: Vercel Connect makes accessing external data and systems simple and secure.
It gives your apps and agents short-lived tokens with precise scopes. https://vercel.com/blog/introducing-vercel-connect
http://Eve.dev is Next.js for agents.
I built Next with a simple premise: 𝚙𝚊𝚐𝚎𝚜/𝚒𝚗𝚍𝚎𝚡.𝚓𝚜 is all you need. Put some React in there and you’re good to go.
Eve asks for even less. 𝚊𝚐𝚎𝚗𝚝/𝚒𝚗𝚜𝚝𝚛𝚞𝚌𝚝𝚒𝚘𝚗𝚜.𝚖𝚍. Put some English in there and you’re good to go.
Like Next, it embraces the filesystem. You can guess what 𝚝𝚘𝚘𝚕𝚜/𝚚𝚞𝚎𝚛𝚢-𝚍𝚋.𝚝𝚜 does. An agent is just a directory, whose entire spec fits in the tweet below.
And like Next on Vercel, it’s seamless to deploy. The infra, like Sandbox, Gateway, Workflow… is the output of your creation.
Vercel: Introducing eve, an agent framework.
𝚊𝚐𝚎𝚗𝚝/
𝚊𝚐𝚎𝚗𝚝.𝚝𝚜
𝚒𝚗𝚜𝚝𝚛𝚞𝚌𝚝𝚒𝚘𝚗𝚜.𝚖𝚍
𝚝𝚘𝚘𝚕𝚜/
𝚜𝚔𝚒𝚕𝚕𝚜/
𝚜𝚊𝚗𝚍𝚋𝚘𝚡/
𝚜𝚌𝚑𝚎𝚍𝚞𝚕𝚎𝚜/
Like Next.js, for agents.
https://vercel.com/blog/introducing-eve
Hugo
Introducing V, a personal agent template.
Built on Eve. Works on iMessage, Slack, and web. GitHub and Linear tools with long-term memory.
http://github.com/vercel-labs/personal-agent-template
Harshita Arora was 17 when she cofounded AtoB and got into YC.
I was 27 when I cofounded Posterous and got into YC.
Nicolas Dessaigne was 37 when he cofounded Algolia and got into YC
It doesn’t matter what age you are. YC is the community to join to go fast when starting.
React → http://ai-sdk.dev
Next.js → http://eve.dev
@aisdk is more relevant than ever, given the intense model competition landscape. Just today, GLM 5.2, an open model, surpassed Opus 4.8 in our Next.js Evals (http://nextjs.org/evals) 🤯
But the world needs a practical solution for how to build and deploy agents. Just like React needed Next.js to solve the task of building an actual web application. And that's eve.
Rahim Nathwani
Mission High's average SAT score is under 1000.
And that's with only about half of graduating seniors taking the test.
Compare that with Lowell (average score over 1300, with over 80% taking the test).
Data here: https://tools.encona.com/sat-explorer#schools=383340%2C383408&year=2019&county=San+Francisco+County
Rahim Nathwani: For most California public schools, the number of students admitted to UC Berkeley is a small fraction of those who met or exceeded state math standards in when in 11th grade.
The left chart shows a weak, positive relationship.
The right chart shows an exception.
@minilek
If you need a website, this is the best website creator on the planet: Ploy by Bryant Chou
Drop everything and use this to improve your website now. A total redesign to your taste is at your fingertips. Try it with a side project and you'll bring it to your main project in 2 weeks or less, I predict.
brryant: AI is making marketers lazy.
So we made the website do the work instead. Today, we're launching @ployai: the all-in-one marketing platform that turns your website into your hardest working employee.
And we're coming out of stealth today with a $27M seed led by @ycombinator and
We don’t care what your age is
We care if you can build with craft and care
And if you can make something people want
Raphael Schaad: I was 33 when I did YC. No excuse.
I was solo founder.
I was international.
I was pre-revenue.
But I had ambition, a big vision, and could build.
Peter Yang
Here’s my new tutorial on how to turn Codex or Claude Code into an AI advisor for life and career decisions.
It involves setting up an /advisor skill with 4 files:
→ SKILL md: How the advisor should behave
→ plan md: Your goals, principles, energy, etc
→ learnings md: Insights from past chats
→ eval md: A checklist AI runs before giving advice
This is my favorite skill because it knows my goals, gives me useful feedback, and gets better the more I talk to it.
📌 Watch now: https://youtu.be/H29laIK8q7M
(Part of) the wonderful team that made Vercel SHIP London possible 🫶
Y Combinator
Many founders get stuck trying to find the perfect startup idea before they commit. But the perfect idea doesn't exist in the abstract. The only way to find what works is to pick one, go deep, and get feedback from real customers.
In this episode of Startup School, YC's @xuster breaks down how to choose what to build, "burn the other boats," and go deep enough to practically run your customer's business— and why that depth is what surfaces the better idea underneath.
00:59 — The "Perfect Idea" trap
02:42 — Why working on multiple ideas fails
03:21 — How to actually go deep
04:51 — Could you run your customer's business?
06:18 — Build at the edge of what AI can do
08:37 — Aim at the most ambitious version
09:33 — What happens when the idea fails
10:27 — Walk fast in one direction
Jesse Genet
Doomers gonna doom, be the light ✨
Our AI future is as positive as we make it, so use AI in positive ways for your family
Trying to do my part to share how I do just that and @NYMag decided to help share it 💕
Jay Owen
Time for VibeCon with the @businessbldrsfl team and @Replit
Also, I don't know if anyone noticed, but Fable ~solved computer use.
Matt Shumer: Seeing so many people saying Fable wasn’t that much better than other models.
Sure, if you’re just doing normal frontend/backend work that Opus/GPT 5.5 can already do, it’s going to feel basically the same.
The magic comes when you’re asking it to do things other models can’t.
dex
At HumanLayer, we’re on a mission to solve the AI slop code problem.
In 2025 we open-sourced our Research, Plan, Implement framework, now deployed inside fortune 500s like Block and Uber - places where shipping slop is just not an option
And that was just the beginning.
Today, we’re opening access to HumanLayer - an Agentic IDE, collaboration platform, and building blocks for your software factory.
HumanLayer enables engineers solving hard problems in complex codebases to:
> move 2-3x faster across the entire SDLC (not just coding)
> maintain rigorous standards for system architecture and program design
Hundreds of engineers at companies of all sizes are already using HumanLayer to ship fast without sacrificing quality.
I'm excited to invite you to try humanlayer today at http://humanlayer.com, and I'm even more excited to see what you build.
@0xblacklight and I are deeply grateful to our team, our customers who give us so much incredible energy and feedback, our investors who have always been in our corner, and our friends and family who have supported us along this crazy journey
if you're a staff or principal engineer trying to make AI coding work at scale for your team, we'd love to hear from you
as @swyx likes to say - let's make this the year of no more slop
Vibecon open doors
Alykhan Kara: Vibecon is beyond the coolest conference ever.
Well done @Replit @amasad @pirroh @Franciscocrz !!
Snap should re-position Specs as the "ultimate 24/7 agentic interface" for people suffering AI psychosis to:
- Monitor 10 AI agent workstreams at once
- See notifications when an agent is done working
- Use voice to prompt and talk to the agents
Then it will sell like hot cakes 😂
David Hsu
93% of technical execs we surveyed are alarmed about vibe-coded apps in their orgs.
Today we’re shipping the biggest update in @retool's history. One governed runtime for all of it.
What "if you know what you're doing" actually looks like:
The model already has incredible general smarts... but what it doesn't have is what you specifically know about your field (it may know a LOT, but an expert can usually push it past what it knows pretty fast), or the exact process you'd follow (or want IT to follow, sometimes this is different, as humans and AI have different strengths).
Prompting is how you get that in. Give it your strategy, your details, your process in the right format and you go from a "good" generic answer to "wow, this is how I would've done it".
It also lets you do things neither you nor the model could've done alone.
For example, last week I worked with @PSkinnerTech on a frontend task, creating really precise/animation-heavy/complex components that we had no idea how to build on our own. But we could describe an approach: create a video of the ideal end-state, have the model attempt to build the same thing in code, write a diffing algorithm that compares its own version with the video, scores how close it is, and hill-climbs to improve.
My only reference was Levenshtein distance for text (I don't know the visual domain as well as text), and it decided on a derivative of heatmaps as a good visual equivalent. We supplied the direction and loose strategy; it supplied the hard ideas and execution.
And this isn't a one-off. It's how I built a few of my ThreeJS demos too. Being able to point at something and say "this is the quality bar," then have it loop until it hits that bar, is incredibly powerful. The model's not going to do that on its own.
You don't have to know how to do the thing, you have to know how to point at the shape of a solution.
That's prompting too, and it's more powerful than it looks.
Matt Shumer: I'm not sure if it'll last (I go back and forth), but today, there's massive capability gains you can pull out of the models if you know what you're doing.
My calendar is pretty great now ngl.
This is what PMs dream of.
Peter Yang
My calendar is pretty great now ngl.
This is what PMs dream of.
Gabriel Jarrosson
Every recent YC batch has outperformed the one before it on revenue. Companies are hitting $500K ARR before demo day.
That's the ‘YC is indecline’ investors are warning everyone about 😂
Investors have been saying YC has declined since 2010. First it was "they're not very good." When the results made that impossible to say, it became "they used to be great."
The goalpost moves every time the results speak for themselves.
The investor telling you YC has declined is usually the same one offering you a check. If you believe them, you skip YC and take their money instead. On worse terms.
Ask what they're offering instead... is it $2M at 20% with no network, Bookface, community, or portfolio sales opportunities?
I've watched every recent batch from the floor... Every investment I've made in recent batches is evidence of what I think. If YC had declined we'd have stopped writing checks!
Talk your ideas to life
jordwalke: Now you can talk to @replit Agent using your voice. It's the most natural way to collaborate. Now available everywhere: mobile app, mobile web, and desktop web.
I think you still need both, but the main lede is: technical founders now have access to business thinking
Business founders now have access to technical thinking
Net net: more startups that actually work, period
Romàn: No one wants to admit this, but the Steve Wozniak / Steve Jobs era of "technical founder + business founder" is over.
For 40 years the model was the same.
One founder builds. One founder sells.
That split made sense when writing code took a CS degree and distribution took a
Brian Bourgerie
Customized education like we’ve never had is on the doorstep.
Matt Shumer: While we wait for Claude Fable 5 to come back, here's another wild demo I built for Alpha School.
A 3D game that generates itself on the fly, fusing what a kid has to learn with what they're actually into.
Just the first prototype, done in two days with Fable.
(sound on!)
Ethan Clark
Working in robotics right now is what I imagine working with language models felt like in 2023. Everyone throwing things at the wall to see what sticks
Pixel prediction (Cosmos), action prediction (VLA), reward prediction (TD-MPC), and representation prediction (JEPA). Different paths for the same problem
The recipe that won in language was self-supervised pretraining at internet scale then light finetune on top. Only representation prediction runs that playbook. It learns from action-free video data so you can pretrain on YouTube and egocentric data then add a control layer. Everything else needs action-labeled data that doesn't scale
As an RL maximalist, I used to hate LeCun's cake. Turns out he was right all along which is how I ended up a JEPA truther
This is just one example that proves that Fable is in a dramatically different league than Opus, 5.5, etc.
Try doing this with Opus. It's nearly impossible. You could spend hundreds of hours and not get close.
Fable did this by looping over several prompts over two days.
Matt Shumer: While we wait for Claude Fable 5 to come back, here's another wild demo I built for Alpha School.
A 3D game that generates itself on the fly, fusing what a kid has to learn with what they're actually into.
Just the first prototype, done in two days with Fable.
(sound on!)
Garry's List
San Francisco has lost 17 Walgreens, 3 Targets, a Whole Foods, and its last Bayview grocery store—all in three years, all driven out by 42,000 retail thefts a year, lax prosecution, and operating costs that made a 1-3% margin impossible to sustain.
Now, Supervisor Bilal Mahmood is proposing a taxpayer-funded subsidy to lure stores back, without addressing any of the reasons they left.
GPT-5.4 for improving a challenging reaction in medicinal chemistry:
OpenAI: GPT-5.4 helped drive a medicinal chemistry project from literature review to a validated experimental result.
Paired with http://Molecule.one’s Maria AI and specialized lab, the model proposed an unexpected way to improve a widely used reaction in drug discovery.
Jay Owen
Excited to hear from @amasad from @replit about AI and the future of business at VibeCon in NYC
One of the best designers I’ve ever worked with is now a principal engineer.
He wants to stay anonymous, but he now designs and builds 95%+ in coding harnesses and the terminal.
His workflow is basically:
→ Get AI to create a design md first
→ Ask AI to generate the components
→ Give AI feedback until it feels right (taste!)
He thinks doing this is now basically a core skill for designers (and frankly any builder).
Not saying you shouldn't design in Figma, but I agree with him that it's important to learn the above too.
Peter Yang
One of the best designers I’ve ever worked with is now a principal engineer.
He wants to stay anonymous, but he now designs and builds 95%+ in coding harnesses and the terminal.
His workflow is basically:
→ Get AI to create a design md first
→ Ask AI to generate the components
→ Give AI feedback until it feels right (taste!)
He thinks doing this is now basically a core skill for designers (and frankly any builder).
Not saying you shouldn't design in Figma, but I agree with him that it's important to learn the above too.
Manny Bernabe
LIVE TOMORROW: Design in Claude. Build in Replit.
Replit is now available inside Claude. Going deep on the new integrations:
→ Claude Design → Replit. Design on-brand apps in Claude Design, then send them straight to Replit for building, refining, and shipping.
→ Replit Connector for Claude. Delegate any general task to Replit straight from Claude.
→ Walking through it all with Jean-Luc Thumm and Pablo Vargas from Replit's technical staff.
→ Builder Spotlight: Carlos from NeuraAgro joining from Medellín, Colombia (Race to Revenue alum).
→ Plus This Week in Replit and live Q&A throughout.
📆 Thu, June 18 · 9AM PT
📺 Live on YouTube, LinkedIn, and X
RSVP: https://luma.com/hmdnrinb
Jam-packed show! See you there! 🙌
Vibecon
Design with Claude, Ship with Replit
Replit ⠕: Design in Claude. Build in Replit
You can now send your design from Claude Design to Replit to turn it into a working app
Rough estimate on $ productivity lost by Fable 5 ban: $12M per hour
Frontier AI-coding daily actives, mid-2026: 5M devs
Fully-loaded cost: $90/hr
Work routed to Fable in 48 hours: 17.8%
Fable is on average ~15% more productive
Effective throughput loss per dev = 17.8% × 15% ≈ 2.7% of output
2.7% × $90/hr = ~$2.40/dev/hr × 5M devs = $12M per working hour
Ankit Gupta
Dude, just try it for like 5 minutes. You live 20 mins from my house you can try mine
Ed Markey: Tesla wants Americans to believe its so-called Full Self-Driving technology is safer than human drivers. Then it should prove it.
@SenBlumenthal and I are calling on NHTSA to demand the data, test Tesla’s claims, and strengthen AV safety data reporting.
Give the people the model they paid for.
WIRED: Trump administration officials tell WIRED that if Anthropic wants to rerelease Fable 5, it will need to ensure the model's guardrails can't be circumvented. Security experts say that can't be done. https://www.wired.com/story/the-white-house-wants-anthropic-to-block-all-jailbreaks-that-may-not-be-possible/?utm_brand=wired&utm_social-type=owned&utm_source=twitter&utm_medium=social&utm_campaign=aud-dev
Was so fun interviewing Spike Jones!
Francisco Cruz Mendoza: Closing out Day 1 of @Replit Vibecon with @amasad and Spike Jonze on the main stage 🎬
Samuel Spitz
Replit Animation is still 🔥 at motion videos (1-shot btw)
Weird that Codex imagegen can generate images but can't save those images into a local folder?
Paul Graham
I just came across this essay I wrote in 2012 predicting that hard tech startups would become a big thing. Not for the last time, YC applications turned out to be a good predictor of future trends.
The Hardware Renaissance: https://paulgraham.com/hw.html
So I have Codex running on a /goal and it's been working for 2 hours but the problem is it's making alot of wrong assumptions so I have to monitor and steer it constantly.
Is this expected? Perhaps I should've had it make a detailed plan first?
