The Battlefield Moved: Notes from YC AI Camp
The battlefield moved.
For the last year, everyone chased chatbots. That race is ending. The next wave is agents that finish whole jobs while you watch. Not a chat window that asks for instructions every minute. A helper that books the trip, files the forms, runs the store, closes the loop, and only taps you when a human should decide.
Sam Altman is already talking about an agent store. That tells you where the platform is headed.
If you are building, the obvious move is also the losing move. Do not clone ChatGPT. Do not fight OpenAI on their turf. If you build the same thing as everyone else, the best people will not join you, and the giants will run you over.
Aim for the idea that sounds slightly crazy today, but becomes obvious if it works. Build the narrow tool the big platforms will need to send traffic to later. The moat shows up after, not before.
The Playbook for This Era
This is the founder shape now.
Stubborn belief in the goal. Small ego about the path.
Talk to users. Ship early. Ship often. Hire for slope, not titles. Be fine looking wrong for a long time.
There are two ways to win in AI.
One is brute force. More data. More chips. Bigger runs. That is how frontier models got here.
The other is insight. One clean design change that beats raw size by a hundred times. John Jumper put it plainly: data, compute, ideas. Everyone can buy the first two. The third is the multiplier.
Know which path you are on. If you are not building the next foundation model, do not pretend you are. You are building product. Workflow. Trust. Interfaces. Guardrails. Distribution. Humans in the loop. Real outcomes.
The biggest opening is hiding in plain sight.
Models can already do far more than the products around them show. That unused space is the gold mine. Stop waiting for the next model. Build agents that take over real, valuable workflows now. That is the whole plan.
The YC AI Camp Experience
Y Combinator ran a two-day AI camp for about two and a half thousand young builders. The message came through in every talk, hallway chat, and half-baked notebook demo: if you are hungry and curious, start a company now. The cost of trying is low. AI opened new rooms to explore. Even if you fail, you will learn faster than any other path.
AI keeps stretching what code can do. The span of code a model can write is doubling every seven months, and the pace is speeding up. Ideas that felt too big last year might be easy next year. So aim high, but stay concrete. Big ambition. Small steps. Fast loops.
When you build, pick problems that are still rough. Use tools that break. Add AI where it feels awkward and brittle. That is how you find the sharp edge early. Then watch for places where adoption could explode once one switch flips. Those are the gold veins.
The event felt like a pressure cooker.
Two thousand five hundred people packed into a hall in San Francisco, buzzing about what to build next. You could feel the charge. If you like making things, this is your moment.
What the Builders Said
Sam Altman
Sam Altman's story is still the simplest proof that the start can look wrong. OpenAI began as eight people and a wild guess that AGI mattered. No plan. No revenue. Just belief. That kind of belief pulls in talent that has nowhere else to go.
He kept coming back to one phrase: the product overhang. Models can do more than the apps wrapped around them. That gap is your opening.
His favorite next step is memory. An AI that knows you well enough to help before you ask. Not in a creepy way. In a useful way. The direction is clear: an assistant that runs in the background, connects to your stuff, and does work. Agents are coming. Software that acts like a junior worker and fades into the background. He even hinted at a new device for that world.
His warning matched the thesis. Do not clone ChatGPT. If you build the same assistant, OpenAI will beat you. If you build what is still missing, the moat can show up later.
His hiring rules were simple. Look for smart, driven people who finish things and work well with others. Judge them by what they built, not their resumes. Choose slope over starting point. Someone who improves fast beats someone polished but slow.
Conviction is the hard part. Elon once wrote to Sam that OpenAI had a zero percent chance. That stings. You keep going anyway. It gets easier after each hit, but it never gets easy.
Satya Nadella
Satya Nadella framed AI as the fourth platform shift after client-server, the web, and mobile. But he said the rollout is less about chips and more about habits. Dropping an agent into a messy workflow does not fix it. You have to redraw the workflow.
He kept returning to three missing pieces: memory, safe tool use, and permission systems. Build those and you get a reliable helper, not a loose cannon.
Trust rests on three fronts: privacy for users, security for firms, sovereignty for nations. Ignore any one and adoption stalls. The world will only accept the energy cost if the value is obvious.
He reminded everyone that AI will demand a lot more energy. We need to earn the right to burn it by making things that matter. The real bottleneck is people changing how they work. Most knowledge jobs still run on copy-paste, approvals, and invisible busywork. Memory, tool use, permission systems, and agents that take action are the next frontiers. If you want a target, build productivity tools that remove drudgery and prove clear surplus.
Andrej Karpathy
Karpathy's talk was the one I replayed most.
Software history feels like three long chapters.
Chapter one was hand-written code. People wrote exact instructions.
Chapter two was neural nets. Instead of writing rules, we trained on piles of data. The rules became weights.
Now we are in chapter three. Large language models act like new kinds of computers, and we "program" them in plain language. A prompt is a program. English is a coding language.
That flips the world.
Anyone who can write a clear sentence can, in theory, build software. Andrej once built a simple iPhone app in a day this way even though he did not know Swift. I felt a smaller version of the same shock. This site was vibe-coded in two days even though I know limited JavaScript. The hard part was not the code. It was setup and deployment.
He compared the era to the 1960s. Back then, big mainframes sat in locked rooms. People used them through time-sharing terminals. Today, big models sit in cloud data centers. We send prompts over the net. When the service halts, we feel an "overall intelligence drop." The whole planet gets a little dumber until the lights come back on.
Then he cooled the hype around flashy autonomy demos. Full robot autonomy is farther off than it looks. Andrej rode in a Waymo car in 2013 that drove perfectly for thirty minutes. Twelve years later we are still tuning self-driving. Agents will be similar. Think of an Iron Man suit. At first, the human is in control. Over time, you slide toward autonomy, but only when each step is safe.
So the best AI products today aim for partial autonomy.
Cursor for coding. Perplexity for research. They do four jobs: gather the right context, call several models, show the result in a clear UI, and give you an autonomy slider. You stay in the loop, checking work with diffs or cited sources.
His point was simple. Speed in the loop matters more than raw model IQ. A two-line change you can scan beats a giant block you have to sift through. Keep the model on a leash. Verify fast. Ask for the next step.
He also argued we should build software the models can read. Put a plain text llm.txt at the top of a project. Write docs in clean markdown with curl examples instead of click paths. Tools that digest a repo into one text blob or auto-summarize a codebase save time for humans and models. Track how much old code gets eaten by 2.0 and 3.0 replacements. Deleting old code can be a progress metric.
Andrew Ng
Andrew Ng stayed on the ground, where founders live.
Speed decides life or death. AI tools make you faster, so learn them. The money is not in chips or huge models. It is in the apps people use every day.
He pushed agentic AI as the next step. Not one long answer, but a loop that plans, searches, drafts, checks, and fixes. That creates an orchestration layer, glue that calls many models and tools. He said this glue makes writing an AI app about as easy as writing a web page was in 1999.
His advice kept returning to clarity. "AI for health care" is fog. "Let patients book MRI time online" is clear. Clear ideas let teams move.
He drew a hard line between prototypes and production. Prototypes are about ten times faster now. Production code is maybe twice as fast. Write ugly code first on your laptop. Make it safe when you ship.
He likes fast feedback ladders. First, look at it yourself. Then ask three friends. Then ask strangers in a coffee shop. Then send it to a hundred testers. Run big A/B tests last. After each round, study what happened and sharpen your gut.
He called AI parts Lego bricks: prompts, evals, guardrails, RAG, voice, embeddings, and more. Every new brick you learn lets you build more. Do not worry about token bills at the start. If you reach the point where you must optimize, you are already winning. But design your code so you can swap models easily. Run evals each week and switch to the best one.
He also said domain skill still matters. An art student writes better image prompts. A doctor guides a triage bot better than both of us. The moat comes later. First build something users love. Momentum is hard to copy.
François Chollet
François Chollet went deeper than most speakers dared.
He argues that making models bigger will not give us real thinking machines. Compute gets cheaper, which is why deep learning took off once GPUs were cheap and data was plentiful. Big models got good at set tasks. People guessed scaling would turn that into general intelligence. He says that guess was wrong.
Knowing a lot is not the same as thinking on your feet. Huge language models mostly replay skills they picked up in training. Real intelligence means handling something you have never seen before.
He built ARC to show the gap. Humans breeze through it. Scaled-up language models stay near zero unless you let them learn during the test. Last year, researchers let models change themselves at run time. They fine-tuned, wrote small programs, revised their own thoughts while solving the task. Scores jumped, but still lag humans. His conclusion: pre-training is not enough. You need test-time adaptation.
He contrasted two views of AGI. Minsky: a machine that can do most jobs people do. McCarthy: a machine that can figure out new problems on its own. The second is harder and more useful. Skill is the product. Intelligence is the factory that makes new skills.
His view is almost poetic. The world looks endlessly varied, but it is built from repeating patterns, like a kaleidoscope. Being smart is spotting those patterns and reusing them fast.
He split patterns in two.
Type 1 patterns are fuzzy, like seeing one tree resembles another. Deep nets are great at these.
Type 2 patterns are exact, like noticing two bits of code do the same thing. Humans reason with these. Deep nets struggle.
Search over programs can uncover type 2 ideas, but brute-force search explodes. The trick is to guide search with type 1 intuition so you only explore promising branches.
He painted a future system that works like a hacker. Faced with a new task, it pulls building blocks from a growing library, plugs them together, tweaks them until the test passes. Each success adds new blocks and sharpens its hunches, so the next task is easier.
His new lab, Ndea, is betting this mix of learning and search will speed up science. First milestone: beat ARC from scratch with a self-improving programmer. His line, in my head: scaling automates the past. To invent the future, models must write new code while they run.
John Jumper
John Jumper's talk was a lesson in trading hardware for ideas.
He said his goal was always to shorten the wait between having an idea and proving it. Proteins are the tiny machines that run every cell. DNA lists their parts like beads on a string. The string folds into a working gadget. Figuring out a shape used to take a year, a crystal the size of dust, and about a hundred thousand dollars. Meanwhile, DNA sequences piled up three thousand times faster than structures. The backlog was huge.
At DeepMind, he described AlphaFold through the simplest lens: data, compute, ideas. Many people had the same protein data and enough chips. What changed the game was a handful of design moves. Swap out one and accuracy drops. Put them together and error falls to a third of the old best.
A careful test showed AlphaFold 2 trained on one percent of the data beats AlphaFold 1 trained on all of it. One unit of insight can be worth a hundred units of data.
They open-sourced code and posted predictions for almost every known protein, about two hundred million. Biologists typed in proteins and saw shapes that matched hidden lab results. Trust spread by word of mouth. Work that took months now started with a download.
His point for founders: good AI is an amplifier. It turns scattered facts into clear guesses so experimenters can test the right thing first. Pick a goal that would thrill real users. Run many cheap tests. Do not obsess over elegance. Share tools when it helps. Other people will push them further than you can.
Fei-Fei Li
Fei-Fei Li kept returning to one phrase: spatial intelligence.
Words are one-dimensional strings. The world is three-dimensional and moves through time. If we want useful robots and whatever the metaverse becomes, we need models that understand 3-D space the way we do.
Her origin story is still one of the cleanest examples of seeing the missing piece. In 2007 she and her students pulled a billion images off the web to build ImageNet because vision algorithms had no big pile of pictures to learn from. In 2012 a team ran it on two cheap gaming cards with an old neural-net idea and cut error rates in half. The lesson is still alive: data plus compute plus the right model can move a whole field at once.
She hires for one thing: fearlessness. People who see a blank page or a huge problem and start anyway.
She stayed practical about open source. Open some things when it helps the ecosystem. Keep others closed if that keeps the business alive. A healthy ecosystem needs both. Clean 3-D data is hard, so you mix real scans, synthetic scenes, and geometry hints. Quality beats raw volume.
Elon Musk
Elon Musk's talk was pure first principles.
He kept saying: work on things that help lots of people. Forget glory.
He told the old stories again, and they still land because they start so ordinary. He kept trying ideas because no one would hire him. He and his brother slept in the office and showered at the YMCA. The work felt worth doing because it was useful.
His control lesson was blunt. If someone pays you, they also pay you with advice. Keep the money. Treat the advice with care.
Then he returned to his favorite move: strip problems down to core facts and reason up. Rockets are metal plus fuel plus math. If raw parts cost only a few percent of the final price, the bottleneck is how people build. Start from the floor, not from history.
He applied the same lens to data centers. When suppliers said it would take two years to build a giant AI cluster, he asked what steps actually take time: a building, power, cooling, cables. He rented an empty factory, stacked generators, set chillers, and slept on the floor while crews pulled wire day and night. Break the "impossible" into parts and you can sometimes bend the schedule.
He warned about ego. The ego-to-ability ratio is a silent killer. When ego grows faster than skill, feedback stops. Startups die there.
He worries the same dynamic hits AI. If we force an AI to repeat polite lies instead of truth, we break its link to the real world. He says truth seeking is non-negotiable. On the long view, he sees three arcs: smart truthful AI, robots that use it to do physical work, and spreading life to another planet so one disaster cannot end it.
Aravind Srinivas
Aravind Srinivas talked like someone living inside a scaling storm.
Perplexity's servers groan daily because users keep arriving. The team rebuilds everything to handle ten times more traffic. His moat is speed: build, launch, fix, launch again.
Big firms copy anything that makes money. Assume OpenAI, Anthropic, and Google will also build a browser. Your edge is to move faster and care more about small details. He still stops meetings to fix bugs himself.
Perplexity's next big swing is a browser called Comet. One box does three jobs: search, chat, and tasks. It keeps tabs in sync and runs jobs in the background like a small private cloud on your laptop. It pulls data from email, calendar, Amazon, social sites, housing sites, and markets. It can find flights, book rooms, fill forms, pay bills. A browser agent can walk the web like a person, so it does not need perfect APIs.
He talked about trust the way a search company must: be right, be quick, show your work. For facts like weather or game scores, aim for no mistakes. For questions with many sides, show several views rather than pick one. Political bias is handled the same way: surface the main views and let readers decide.
He laid out a business ladder. Subscriptions first. Pay per task next. A cut of what users buy later. You do not need Google-level margins to win.
His view of the web was blunt. AI summaries will push more traffic to big trusted sites and less to thin SEO pages. The long tail gets thinner. Real brands and real data survive.
Michael Truell
Michael Truell's Cursor talk was about the end of normal coding.
The goal is to stop typing code as the main act. In the future, you tell the computer what you want and it builds it. Cursor today feels like a supercharged editor. Tomorrow it should feel like talking to an assistant that changes a codebase.
He said a big jump will come when people can trust the AI to change large codebases without staring at every line. Today, pros still read everything. Huge codebases break current models. Context windows are too small. Continual learning is weak. Agents lose track over long jobs. Those limits have to move before AI can match whole engineering teams.
He also said humans will still supply taste. Someone must decide what the program should feel like and why it matters. Taste covers looks and deep logic.
Their story was a reminder that a wrong first idea is normal. They started with a co-pilot for mechanical engineers in 3-D CAD. It flopped. Not enough data, and they did not love the field. They learned how to train big models and run them cheaply, and that skill later powered Cursor.
They chose to build their own editor, not just an extension, because real change needed UI control at the root. Three months after writing the first line they shipped a public beta. Then they tuned it in public for about a year before it took off.
Their north star metric is paid power users who use it four or five days a week and pay for it. Growth came mostly by word of mouth and constant product work, not marketing. Scale becomes a moat: more users, more feedback, better models, more users.
He described their hiring style too. Stay tiny early. The first ten hires matter. Those people become the immune system. In interviews, they do not allow AI help on the first screen. Raw problem solving still predicts talent. Later they teach tool use.
Dylan Field
Dylan Field offered a quiet warning and a quiet hope.
When software gets easier to build, design matters more. AI lowers the floor so new people can join and raises the ceiling so experts can try wilder ideas. But it does not remove judgment. Prompt-to-app feels magical until you look closely. It still needs a designer's eye.
He argued designers stay important because they add empathy and taste, and he expects roles to blend. Some people who call themselves developers will start calling themselves designers.
He made a point about flow. The faster the tool answers you, the faster you enter flow. Figma tries to feel like play. When the loop is tight, work feels like a game, and new paths appear. That is why real-time multiplayer was worth the pain.
His founder advice was simple. Notice what eats your time, then hire or automate it. Ship sooner than feels safe. If a roadmap says nine months, ask how to cut it to three. Two-year stealth builds are rare exceptions. Product-market fit can take years. Figma needed five.
Jared Kaplan
Jared Kaplan's notes felt like the math behind the hype.
Scaling laws are a hidden map. If you grow data size, model size, and compute together, error drops along the same smooth curve. OpenAI, Anthropic, and Google keep aiming at that curve, and it keeps rewarding them. He said the same rule seems to work in reinforcement learning now too. If it ever breaks, it will be because we created a new bottleneck, not because the curve stopped being real.
He told everyone to watch task length. The number of steps an AI can finish by itself is doubling about every seven months, and speeding up. As models get more general, they hand people larger blocks of free time. We can pass off longer chains of messy work and trust the system to close the loop.
He listed the gaps plainly: deeper real-world knowledge, longer memory that sticks, stronger self-correction, one brain that handles text, sound, pictures, and video, clean scaling to giant jobs, and training data that stays aligned with what we ask. Human data is still the rocket fuel, especially for teaching models how to handle long-lived tasks and remember past actions.
His founder advice matched the others, just with more math behind it. Ship before everything works. Find spots where growth can shoot up once a switch flips. Do not bolt AI onto the side. Bury it inside the real task so the user just sees the job get done.
A Few Personal Moments
A few moments made the whole thing feel real.
I got to visit Anthropic and snag Claude credits.
An AWS party handed out cloud credits like candy.
I met other Amherst alums and traded notes late into the night.
I left with a notebook full of ideas and the itch to start building tomorrow.
The One-Page Takeaway
If I had to compress the whole week into one page, it would be this.
The battlefield moved from chat to work.
Agents will win because they finish jobs, not because they talk well.
The gap between model ability and product reality is the opportunity.
Founders win with speed, taste, and humility. Stubborn about the goal. Flexible about the path.
Pick a real workflow. Map the best human version. Code the repeatable steps. Let the agent handle judgment steps. Keep the human as the final signer.
Move fast. Stay concrete. Prototype, test, kill or grow. Build trust with clear checks, clear UIs, and a short leash.
Do not wait for the next model.
Start now.
Resources
Talks I watched and took notes from: