We (almost) met Sam Altman

Good morning!

Notice things looking a bit different this week in our weekly email? Will share more on that soon. But let us know what you think of the redesign at the end of today’s article. 👀 

More pertinent, we had a pretty wild week! Monday and Tuesday were spent at Y Combinator’s AI Startup School, a two-day student conference featuring insights from everyone you probably know in AI.

So, naturally, we’re pushing our planned programming to bring you everything we learned from our two days alongside 3,000 of the brightest college students and recent grads from around the world.

Expect deep dives into the below, and more:

• Sam Altman’s belief that this is the best time, ever, to build a startup.

• Andrej Karpathy on entering the decade of agents.

• Elon Musk’s thoughts on superintelligence (hint: it might be coming way sooner than you think).

P.S. We’re in SF this week! if you want to meet up and say hi, reply and let us know.

Arek and Ethan 🦄

Arek here. I’ll be your proverbial newsletter narrator for this week. Ethan applied a bit late to Startup School, so he had to FOMO watch the conference from the videos I sent him.

Let’s start by setting the scene: 3,000 students flocked into Mission Bay, San Francisco, at 8 am, forming a line wrapping over a half mile long from the entrance of YC’s venue at Pier 48 past the nearby Visa headquarters.

ICYMI, Ethan and I were just in New York hosting a slew of Tech Week events. So when I share that I woke up at 5:30 am to make it to the front of the line, I attribute it not just to an innate tech bro instinct to be 50 feet away from Sam Altman in the flesh, but rather a standard case of jet lag.

And so I stood there, in this ever-growing, endless line, waiting to be admitted to one of the most anticipated tech events of the year.

And finally, I made it in!

Not exactly front row seats, but close enough to get a view without using the max zoom on my phone.

YC’s CEO, Garry Tan, welcomed us to the conference, beginning with our formal introduction to YC AI Startup School. He was followed by Daniel Lurie, mayor of SF. The stage was now ready for the first distinguished speaker:

Sam Altman, CEO of OpenAI.

Before I dive into the content of each talk given on the first day, let me share how day 1 was organized.

The day included seven events, half of which were composed of fireside chats with YC partners like Garry and the other half comprising lecture-style talks, each about 45 minutes. Between sessions, students had breaks, lunch, and the chance to attend “Mini Office Hours” for feedback on startup ideas, similar to standard YC office hours.

So, without further ado, our insights from all the talks we could attend on day 1!

• One of the hardest things is to hear someone you trust completely say your idea is going to fully fail. It takes an incredible amount of strength to continue to build from there.

• Continued persistence is wholly underrated and a super undervalued skill in building a company. This generally looks like building for years without seeing success.

• You can win by iterating fast with low cost.

• Hire people who are scrappy and motivated, rather than those with a loaded resume. There are cases where loaded resumes are needed, but generally not for startups!

• Model capability scales with time spent. AI’s ability to handle longer tasks is doubling every seven months.

• The frontier is shifting from copilots to full end-to-end AI-native workflows. Founders are starting to build full replacements for jobs, especially in areas like law, finance, and internal business ops (where some error tolerance exists).

• Persistent memory systems are emerging as a major unlock. Long-term context, like files and past interactions, will be key to enabling more sophisticated and useful AI.

• Human data remains critical. Pre-training teaches language structure, but human feedback via reinforcement learning teaches usefulness, especially for longer, nuanced tasks.

For added context, John leads work on AlphaFold, an AI platform pioneering protein folding prediction. He is one of the recipients of the 2024 Nobel Prize in Chemistry for his work on AlphaFold.

• Breakthroughs often come from architecture and design choices, not just more data. In AlphaFold’s case, research decisions had 100x the impact of dataset size.

• You can’t generalize breakthroughs across domains. A model like AlphaFold may solve protein folding, but won’t necessarily help with broader tasks like simulating molecular dynamics. ML applications need clearly scoped goals.

• Fast iteration wins. Lower the cost of failed ideas so you can test more of them. Run small experiments often, and periodically zoom out to check for real progress.

• Talk to users. Don’t solve elegant but irrelevant problems. Build something that people need to drive real-world impact.

• Real traction in AI startups can take 2-3 years. Early product-market fit might look like slow growth, but sustained usage is the best signal you’re on the right track.

• Don’t just vibe code. Understand systems. Knowing distributed systems is becoming essential, especially as AI products get more complex. Vibe coding is fine for frontend engineering, but backend engineering needs rigor.

• AI makes development cycles faster, but sneakier. Bugs can be harder to detect as code gets quickly iterated on, so reliability and strong testing loops matter more than ever.

• Network effects aren’t sticky in AI; brand and experience are. Most AI tools are interchangeable, so retention comes from real utility and thoughtful design, not just distribution.

• AI’s evolution mirrors the brain’s. The next big leap for AI is spatial understanding, which will enable models to better navigate and reason about the physical world.

• Language is generative, but the world is not. Bridging that gap is a core challenge for embodied AI and real-world applications.

• Great researchers are intellectually fearless. The most promising minds aren’t afraid to tackle foundational problems or break with orthodoxy.

• Startups often begin when no one else will hire you. Elon didn’t set out to be a founder. He just wanted to build for the internet, and no one at Netscape called him back, so he built Zip2, a service that provided online city guides and business directories to newspapers.

• Avoid ego inflation in organizations. When ego outpaces ability, the feedback loops that drive innovation break down. Focus on the work, not the glory.

• First principles thinking helps clarify what truly matters. This mental model shaped how he approached building companies, from rockets to AI.

• AI is a tsunami. He believes superintelligence could arrive as soon as this year, with multiple competitive intelligences emerging globally. Don’t get distracted by surface-level noise.

After a long day of talks, I made my way to a small dinner hosted by YC partner Tyler Bosmeny. Tyler was the founder of Clever, one of the largest unified identity platforms for education, and remained at the company until not long after its purchase by Kahoot in 2021.

It was a wonderful dinner with some great conversations! I really enjoyed talking with everyone there and hearing Tyler’s journey and insights.

At this point, I tried to muster up some energy to attend some afterparties, but after heading over to one, I opted to go back to my hotel and crash for the night. After all, we had a stacked day 2 ahead, and I needed time to process all the great insights from earlier.

I was up bright and early once again, this time starting the day with a fireside between Garry Tan and Microsoft CEO Satya Nadella.

• AI sits atop the cloud, which sits atop decades of compounded infrastructure. Foundation models are becoming an infrastructure layer, like SQL, and products now live in tight feedback loops and smart toolchains built on top.

• Economic surplus is the benchmark for AI's success. We need real, useful products that justify the massive compute and energy use, earning social permission through tangible impact (like helping Indian farmers via AI chatbots).

• The AI bottleneck isn’t tech; it’s change management. Just as PCs reshaped workflows, AI will require new artifacts and behaviors. Most cost in sectors like healthcare lies in processes, not tech.

• Software engineering won’t disappear, it’ll shift. Developers will become more like managers of meta-systems: enforcing working code, quality, and trust, especially as AI agents take more action on our behalf.

This talk was my personal favorite!

• LLMs are becoming the new operating system. They're not just utilities that run on usage-based pricing. They're evolving into platforms that can take actions and serve as the foundation for new types of software.

• Vibe coding enables rapid building, but has limits. Much of the modern stack is human-oriented. Many actions can't be fully automated yet because the documentation wasn’t written with machines in mind.

• We need documentation for LLMs, not just people. For agents to work, docs must be prescriptive. Vercel’s move to include cURL examples in its documentation is a step toward machine-readable documentation and better agent integration.

• The ecosystem is still early. We’re in the “mainframe” phase of AI compute (think the big, clunky computers of the 60s). Chat is still like a terminal, and the true GUI layer for AI hasn't emerged yet. Internet 3.0 isn't about code or weights; it’s about prompting. Build with agents in mind.

• Execution speed is one of the strongest predictors of startup success. Fast iteration, rapid prototyping, and willingness to pivot based on data are key. Build 20 prototypes, not one perfect product.

• Agentic AI, with feedback loops, outperforms zero-shot models. An emerging orchestration layer is forming between applications and foundation models, enabling more interactive and adaptive systems.

• Concrete ideas beat vague ones. Detailed, actionable ideas (especially from domain experts) enable faster execution, clearer direction, and better interpretation of product signals.

• The application layer is where the most opportunity lies. Tools like AI are becoming easier to use, and coding is now accessible to more roles. Engineers with product sense are increasingly valuable, and even non-technical people should be coding more.

• Great talent is drawn to bold, world-fixing ideas. Leaving a stable job to build something ambitious like “building a bank” isn’t just risky, it’s magnetic. Amazing people join companies with missions that feel a little insane but deeply meaningful.

• The next generation of banking will involve stablecoins. In his view, this shift represents the frontier of financial infrastructure: ripe for builders willing to rethink it from first principles (I asked him this question personally, BTW).

• Start small, move fast, and seek feedback relentlessly. From cold emailing designers to charging early, feedback loops were essential to Figma. He converted big ideas into manageable chunks to accelerate learning and iteration.

• AI supercharges prototyping, but it’s still early. It increases iteration velocity, and public content (like Midjourney prompts) helps showcase capabilities. Still, applied research must involve user feedback and design thinking to stay grounded.

• Stay human. Designers and PMs should help shape AI tools, but replacing human relationships with AI is a “societal self-own.” His guiding philosophy? Explore consciousness, learn from others, and keep the simple things simple.

• AI is the new Space Race, and America must lead. The release of DeepSeek was a “Sputnik moment.” Maintaining U.S. supremacy in AI across GPUs, models, and infrastructure is viewed as essential for national competitiveness.

• Policy is shaped by builders, not just politicians. He sees himself as a spokesperson in DC, not a politician. Anyone with a clear opinion can influence policy.

• Fast feedback loops win. Tech and DC share a lesson that small, focused teams with tight iteration cycles can make an outsized impact. He borrowed this idea from military and startup playbooks, such as the OODA loop: Observe, Orient, Decide, and Act.

• Build products and relationships. The advice is simple: stay in touch, follow up, and be authentic. Whether in tech or policy, long-term influence comes from trust and shared missions, not just ideas.

And with that, Startup School was a wrap!

I made my way over to an afterparty hosted by our friends at Google Cloud before running to my flight back to LA.

All in all, it was an amazing two days of events. Kudos to the Y Combinator team for putting on such an incredible series of speakers and bringing together students from quite literally every corner of the world.

Attended and had a favorite talk? Would love to hear your thoughts. Reply and let us know!

We’ll be back with our usual startup of the week programming next week. Otherwise, hope you enjoyed this rundown on Startup School!