Yesterday we hosted 400 top university students at the YC Summer Conference - a fun day of talks about startups.  Talking to the students afterwards, I found myself giving a lot of the same advice. So in case it's useful to others, here is my startup advice for students.

They actually did it.  Boom is officially the first independently developed supersonic jet.

The Boom XB-1 was built by a team of just 50 people. Many of them joined Boom right out of college.

Lots of hot takes on whether it's possible that DeepSeek made training 45x more efficient, but @doodlestein wrote a very clear explanation of how they did it. Once someone breaks it down, it's not hard to understand. Rough summary: * Use 8 bit instead of 32 bit floating point numbers, which gives massive memory savings * Compress the key-value indices which eat up much of the VRAM; they get 93% compression ratios * Do multi-token prediction instead of single-token prediction which effectively doubles inference speed * Mixture of Experts model decomposes a big model into small models that can run on consumer-grade GPUs

If you are a technical founder, you do not need a non-technical cofounder.

Some of the best startup ideas live in a grey area of legality. Airbnb, Uber, Lyft, YouTube, Coinbase were all somewhere between questionable and flat-out illegal in their early days. One reason for this is that incumbents cannot take any legal risk on a new initiative. Even if Google execs saw Uber and thought "that seems promising", there was no way they could launch a competitor because they would face unlimited liability that could destroy all of Google. So a startup taking legal risk will be invulnerable to competition from big companies for a long time.

If you're building an AI agent to automate some job, like a payroll specialist, the hardcore move is to just go get that job and do it for a while. We call this "going undercover", and we're seeing more of the top founders do it. It's the best way to learn what to build.

The pilot who flew the XB-1 today is Tristan “Geppetto” Brandenburg. He's an absolute bad-ass who literally went to Top Gun and has 200 carrier landings.

Until recently, I thought that 3D printing had failed to live up to its promise. But then I noticed that every hardware company in the last YC batch was 3D printing their product with their own printers. Turns out it actually worked, it just took a while.

(0/25) Here's a list of 25 YC companies that have trained their own AI models. Reading through these will give you a good sense of what the near future will look like.

Ares Industries, which builds cruise missiles, launched today out of YC’s current batch. This is YC's first weapons company, and this is an actual photo of their missile. Here is the story of Ares Industries and why we funded them. (1/10)

One of the highest leverage things founders can do right now is to invest in becoming top 5% in their use of AI coding tools.

Build a company good enough to get bought twice.

One of the mega trends in this YC batch is the wave of consumer AI companies. Consumer was stuck in the doldrums for years, but AI has brought it back in a big way. Here are the 21 consumer AI companies that launched today. 🧵

It's a beautiful morning to break the sound barrier.

Investors always want to know how your idea is defensible. The truth? There is rarely defensibility. There is only speed.

This section from the Mr Beast onboarding guide is how every company should handle promotions.

It's funny how the people critical of founder mode didn't show up until Tuesday, after the holiday weekend.

A good prompt for thinking of startup ideas: "What ideas has the world given up on, that someone should try again?"

Talking to second time founders (especially v. successful ones), I've realized they tend to make the same mistakes. In fact, I made these myself in my brief attempt to start a second company after Scribd. The three most common ones:

People ask us why they should work at a startup. This is why.

"Founder Mode" was an instant classic because it articulated something that thousands of founders already knew to be true but weren't sure how to put into words.

Waymo still has fewer than 1,000 cars on the road. Not many people would have predicted that 2 years after we invented the self-driving car, we'd still have made less than 1,000 of them.

Vertical llm agents is the new vertical saas - the most straightforward way to generate $1B company ideas.

LLMs can now browse the internet pretty well.  This has blown open a whole new set of startup ideas.

CaseText is one of the first vertical AI agents to be deployed at scale. It's an AI legal analyst used by thousands of lawyers. Oh, and it was bought for $650M just 2 months after launch. Here's @Jacob_Heller's playbook for building vertical AI agents that actually work:

The OpenAI board was particularly bad, but every board I've been on has been a clown show of incompetence. People who haven't been on boards don't realize how dumb they usually are.

Thought provoking points from @karpathy at AI Startup School today. Made me see LLMs in a whole new light.

Whenever I go to a YC alumni meetup, one of the things that stands out is how many companies seem to be going nowhere for years, and then take off in years 4-6.

Founders are always worried about competition, but startups are so hard that most of your competitors will shoot themselves in the foot at some point. Most of the time you can win by just not dying.

If you tell your friends you're burned out, they'll always prescribe the same thing: "You need to take a vacation". But it never works. Burnout is completely misunderstood. This new article by @bscholl is the best explanation I've seen.

OpenAI Deep Research is very good at doing market size research and calculations. It's so much better than Googling around and wondering if you should pay $5K for some research report.

Lawyers were the original prompt engineers. They've been programming in English long before the rest of us.

There's a new SOTA model on the swe-agent leaderboard. Honeycomb just beat out Amazon Q with 22.06%. Devin's score was just 13.86%. Here's who Honeycomb is and how they got sota.

Elon Musk, Sam Altman, Satya Nadella, Andrej Karpathy, Andrew Ng, and Fei-Fei Li will all be at AI Startup School on June 16.

Rocketable is building the first 1 person $1B company. The plan is simple: buy up software companies, fire all the employees and replace them with AI agents. Literally the most AGI pilled idea I have heard.

In case you missed it, @elonmusk agreed that we should delete the dumb regulation that prohibits supersonic flight regardless of noise profile. This will accelerate bringing back commercial supersonic flight.

Boeing planes have not changed in 67 years. Someone else will need to invent the future.

Since Xi's crackdown, if you raise VC in China and your startup fails, the government seizes your house. Not exactly surprising no one wants to sign up for that.

ChatGPT is now better at math than almost all humans, an incredible accomplishment. So it's interesting that it's still worse at writing cold emails than the average founder.

Watching Paul Graham and Brian Chesky invent the concept of Founder Mode at this event, it was obvious that we were watching history being made. This event felt like the startup version of the Potsdam Conference - a moment in time when the best minds came together and invented a new doctrine. It is not fully fleshed out yet, but when it is, Founder Mode will be one of the foundational pieces of startup advice that everyone knows - like "making something people want" and "doing things that don't scale".

Yesterday we wrapped up the W24 batch. We did a lot of things for the first time this batch. Here were the high points for me ...

Going to industry conferences remains one of the best ways for B2B startups to get their first customers. The best way to get your first customers is to meet them in person, and a conference is simply the most scalable way to do that.

This quote from @paulg about startups and money is what convinced me to drop out of Harvard and start a startup: By compressing the dull but necessary task of making a living into the smallest possible time, you show respect for life, and there is something grand about that.

Introduced some young founders to a number of incredibly successful people. Every single one was happy to meet them and give them advice with no expectation of anything in return. Silicon Valley is an amazing place.

Scale.ai is one of the generational AI companies. Here's some wisdom @alexandr_wang dropped at the YC reunion today on how he runs a 1,000 person company in Founder Mode:

The real reason startups do so much better in SF isn't because it's easier to raise money or hire employees, even if both of those of are true. It's because you become the average of the 20 people you spend the most time with. If you live in SF, those 20 people are probably successful startup founders. If you live anywhere else, they probably aren't.

One of the interesting consequences of the AI boom is that it's returning the advantage to young founders. When things are changing this fast, experience is not that helpful and what matters most is how quickly you move.

Thank you, @bscholl, for inspiring all of us to aim higher.

Heard from a startup today that o1-preview made their product work overnight. "Many use cases we struggled to get above 80% were suddenly at 99%+ with o1-preview"

Amazing comment by @karpathy on the HN story about Operator, posted by Operator.

It seems that the startup world is bifurcating into companies who follow the Elon Musk "ultra hardcore" management style, and everyone else.

41% of recent YC founders worked at a YC company as employees before starting their own company. We only learned this stat recently and even we were surprised it was so high.

Today @toinfinityai made history with the first completely AI-generated YC demo day presentation. Here's their full demo day video, featuring @lina_colucci's AI clone.

One of the less talked-about qualities of good founders is that they are good business thinkers. They're naturally curious about how different businesses work and are good at breaking them down, the same way you would take apart a machine to understand it.

An interesting consequence of the wave of startups building on LLMs is that for the first time I can remember, many YC startups are facing more technology risk than market risk. The main question is not if people want it, but if it's possible to build it.

(1/5) A few years ago, it was widely believed that the first deployment of self-driving would be long-haul trucks on highways, which is a much easier technical problem than city driving. VCs invested billions based on this theory. Why did so many people call this wrong?

In 2006, Elon wrote a blog post called, humorously, the "Secret Tesla Master Plan". I find these 21 words fascinating, because they not only predicted the next 10 years of Tesla, they established a new model for how to build hard-tech companies.

Today is aviation's Falcon 1 moment. It is the day that the mantle passes to the next generation of companies.

It's unbelievable that our government is happy to spend $2T on a stimulus deal, but won't spend even $2B on scientific research to solve the problem that is causing us to need the stimulus in the first place.

Many of the best startups ran their company out of the founders' house for a curiously long time - often until they had 10+ employees. This is Replit when they were working out of @amasad's house.

Amjad's Law: the return on learning to code is doubling every 6 months.

Prediction: the invention of AGI will be like the end of covid, not the start of covid. There won't be a single defining event; we'll just gradually stop talking about it.

It's fun to watch founders who are new at sales have the magic moment when they realize that sales is about listening, not talking.

3) Running away from their existing expertise. People are burned out on their first industry because they know what makes it hard. So they try a new industry only to realize it's just as hard but they have fewer advantages. Usually better to stick with the devil you know.

It's hard to overstate just how great the bull case for AI is. (from @DarioAmodei)

How do you start a startup that's building a supersonic jet? @bscholl did, and his company @boomsupersonic just did their first flight. Here's a breakdown of 8 hard-tech companies with crazy ambitious ideas, and how they got started.

Most first-time founders are hilariously incompetent for the first couple of years (I certainly was). The ones that make it just manage to not die for long enough to get good.

When I talk to founders of YC companies started before 2020, one of the things they tell me most consistently is that they wish they could switch back from remote to in-person but can't figure out how to do it.

This guy tried all 12 AI browser automation tools on the market to find the best one. His conclusion? The open source Browser Use + DeepSeek R1.

Engineers actually make great salespeople because they just see sales as an optimization problem.

Solo founders often think that the more progress they make, the easier it will be to attract a co-founder. But our data from running YC's co-founder matching site shows the opposite: the people with funding or traction are least likely to find a co-founder.

(5/5) A lot of learnings here for AI startups. Many startups are working on an AI system that effectively replaces a human doing some important job. The best way to get these systems live is to find a way to start with a near zero-risk setting.

Each one is explained in detail in the full article (the DeepSeek section is titled "The Theoretical Threat"): youtubetranscriptoptimizer.c…

Silicon valley is a magical place where autist nerds who were uncool in middle school become the most powerful people in the world.

As Carta discovered, adding confetti or balloons to your b2b saas product is one of the easiest ways to delight your users.

I should add that all the companies that I can think of in this category were violating laws that were widely regarded as dumb and unpopular, like the laws protecting the taxi monopoly. Startups violating laws that are widely regarded as important aren't likely to end as well.

"You are the upper bound on how much people in your company care." This one really stuck with me.

Llama 405B is a big deal - the first open source model with state-of-the-art performance. It's easy to forget that what is now a key Facebook strategy began with the accidental leak of the weights as a torrent on 4chan. Sometimes accidents lead to the best ideas.

When founders are looking for a startup idea, vertical AI agents is where I usually point them. It's the easiest place to find a great startup idea right now.

"Ask yourself, what risk are you avoiding right now because you're comfortable?"

A startup selling to a large but slowly dying industry (i.e., car dealerships) will tend to be valued lower than you'd think. It took me a while to understand why because, after all, they're not dying anytime soon.

Logging all user events to a slack channel is an amazing way to get a feel for what is happening with your product.

In the early 2000s, you'd regularly see web 2.0 companies go from 0 to millions of users overnight. AI companies going from 0 to $10M ARR in a year is the 2025 equivalent.

Once a startup saturates its initial market, there are two basic ways to expand. Sell the same product to new customers, or sell a new product to existing customers. When they're getting started, most startups think they'll do the former, but they mostly end up doing the latter.

The founders of Instant have been working on this in stealth for 2 years. They launched today and it's the top story on Hacker News.

The deepsilicon founders dropped out of Dartmouth 3 months ago with a crazy ambitious idea. They are making custom ASIC processors that compete directly with Nvidia by having better performance. Seem impossible? This thread explains how they did it.

Pretty interesting that LLMs now direct the majority of my discretionary purchases but make no ad revenue whatsoever. Wonder how long that situation will persist for.

This is one of the most promising places to find startup ideas right now.

TIL Saudi Aramco was started at 225 Bush Street in San Francisco.

Getting set up for AI Startup School - starts tomorrow.

Last week we celebrated the end of YC's S23 batch. It seems like a good time to share some thoughts on the batch, and on the ways YC has changed in the last year.

2) No idea is good enough. First time founders enthusiastically leap into their first idea. Second time founders know that they're signing up for a 10 year commitment, so they're gun-shy. But the problem is that all startup ideas have major flaws.

1) No cofounder. They started their first company with a friend. But now they can raise money and hire a team with just an idea, and giving up 50% is so expensive. But cofounders are also about accountability and moral support, and that turns out to be incredibly important.

That feeling when building an agent where you swap out your model for a newer one and everything suddenly works better.

"Think about how much of our lives are dictated by who we're randomly assigned to be roommates with in college" - @daltonc

Kyle is a force of nature who pulled forward the future of self-driving cars probably by several years. That alone will end up saving thousands of lives. We call people who did less heroes.

Today, BillionToOne went public, becoming YC's 4th biotech IPO. As a company that quietly provides societally important infrastructure, BillionToOne is less well known than they deserve to be. Here is their story, from grad students to IPO. BillionToOne is the rare company that is both incredibly good for the world and also an extremely good business. 1 in 11 pregnancies in the US take the BillionToOne genetic test, and through this they prevent an immense amount of human suffering. But BillionToOne is also the rare biotech that has generated revenue from the early days. It has always had excellent margins, rapid growth, and capital efficiency. Today, it is going public not just as a scientific success, but a commercial one, with over $265M in ARR and 65% gross margins. We met the BillionToOne founders, Oguzhan and David, when they were still finishing up their PhDs. At the time, BillionToOne was only an idea, but it was a thought provoking one. When a woman is pregnant, fragments of the fetus' DNA circulate in her blood. What if you could design a genetic test for the fetus using this free circulating DNA? That would allow you to do pre-natal genetic testing with a simple blood draw. The problem is that this DNA is a mess - it's tiny snippets mixed in with a much larger amount of unrelated fragments. To extract the signal from the noise requires both advanced wet lab sequencing techniques and also advanced machine learning algorithms. Oguzhan and David were the exact right people to crack this because they had both biology and CS backgrounds. The BillionToOne founders have always moved fast. In just 6 months from YC funding them, while still finishing their PhDs, they went from an idea to a proof of concept. From there, it was just two years to regulatory approval and a commercially available test. They haven't slowed down. While the pre-natal genetic testing market they started in is over a $2B market and still mostly untapped, they've already expanded to a second market in oncology. It turns out that their same technology that makes sense of free floating DNA can also be used to detect cancer from a blood test. The potential market size for that is enormous - perhaps $100B. That is where BillionToOne is going, and it's quite possible that if you're reading this, you will someday regularly take their blood test to stay cancer free. To all the people who say that Silicon Valley just funds GPT wrappers and B2B SaaS, BillionToOne should be a star example of how the SV ecosystem can solve societally important problems. It is a classic story of how a highly technical team, an ambitious idea, and a small amount of funding can catalyze the creation of enormous value.

The team at Ares went from starting the company to flying a test missile in 11 weeks. They’ve been able to move so fast because they’ve done it before - they’ve worked at 2 previous missile startups. You can read more about them on Launch YC: