Y Combinator AI Startup School

Successful innovation requires one to be contrarian, but ultimately right.

When OpenAI released GPT-1, Elon Musk said the company had a 0% chance of succeeding. That was the majority opinion of the time. Altman himself called the model crap. But he had the conviction to keep progressing, believing something great would come out of the project in due time. The advice is simple to understand, but incredibly difficult to follow in practice.

Scaling laws continue to guide AI progress. What’s left to build around it is simple.

The relationships discovered between compute, dataset size, parameters, and model accuracy has paved the way for massive AI progress over the last decade. The models will continue to improve.

For that reason, it is crucial that people build things that don’t (yet) work. Try to get ahead of the curve. That way, once science improves AI capabilities, the necessary workflows will already be in place waiting for new models to be plugged in.

This was my approach with my senior thesis, QuickCase. A model capable of complex problem-solving is necessary for verifying legal formatting minutia. OpenAI’s full version of o1 was released four months into the development of QC. I didn’t know when the new model would come, or exactly how powerful it’d be. But I knew it would come. And they will continue to come.

Competition is to be expected, because things worth building are competed over by well-funded companies.

The true moats today are moving fast, and both embracing and sleeping with the fear of competition every day. That’s real.

I am not a Perplexity user myself, so I was genuinely curious how they haven’t been stomped out by the incumbents yet. But people online really do believe Perplexity offers better search and citation capabilities compared to ChatGPT, Claude, etc. I’m excited to see how Aarvind continues to challenge the big guys over the next few years with sheer velocity and a chip on his shoulder.

First principles thinking unlocks solutions otherwise shrouded by abstraction.

People told Elon that rockets cost tens of millions of dollars. This top-down perspective is limiting. Instead, Elon asked, what is a rocket made of? How much does each component cost? What quantity of each component is necessary? He quickly discovered how much bloat could be avoided by creating rockets himself. Hence, SpaceX was born.

If you’re going to use energy, it better be socially acceptable. In other words, AI’s energy consumption will only be “worth it” if AI is showing up in real stats (like GDP surplus).

Earlier this month, the World Bank released growth forecasts which predict the worst decade for global growth since the 1960’s. At the same time, some analysts like those at J.P. Morgan project major boosts in GDP due to AI over the next decade. In April, the U.S. government released their own analysis of the macroeconomic effects of AI, assessing that most estimates for the short and long run effects on U.S. GDP vary, but are typically positive.

Nadella points to GDP as an easily identifiable metric for which to assess AI by— if GDP is increasing, the energy consumption associated with the progress is broadly “worth it” for society. This assessment kicks the can down the road a bit too far. The time lags which exist between AI adoption and GDP growth imply cause-effect relationships can only be assessed after time has passed since the new tech’s adoption. By 2028, AI is projected to consume up to 22% of all U.S. household electricity use, and the share of U.S. electricity going to data centers may triple. Policymakers won’t see the effects of this on GDP for another few years— by that time, energy consumption will likely rise even more.

I am not advocating for a halt on AI progress— I actually lean more towards AI optimist than doomer. But that’s beside the point. I would just ask Satya how engineers and researchers should think about this concept of progress being “worth it” in practice. Because the delay between AI adoption and GDP impact makes it impossible to assess whether the tech is socially acceptable before actually consuming more energy.

Build less Iron Man robots, and more Iron Man suits.

This talk was a fan-favorite. So many great nuggets of knowledge. Check out the full talk here.

My favorite part of the talk was Karpathy’s Iron Man analogy. The suit extends Iron Man using both augmentation (giving the user strength, information, and other tools) and autonomy (like having a mind of its own). Today, we can build both. We can build AI tools that augment our workflows, and we can build fully autonomous agents that act on our behalf.

This gives way to the concept of the “autonomy slider,” which offers users the ability to tune how much autonomy they are willing to give up for a given task. For example, in Cursor, one can use just tab-completion, leaving the developer almost fully in control. However, one can also give up full control to agentic mode to manipulate the code base as the AI assistant sees fit.

Karpathy thinks this autonomy slider will be crucial to successful apps in the coming years. Building “Cursor for X” will be a common pattern amongst winning startups in the AI age. But for now, while LLMs are still fallible (hallucinating, demonstrating jagged intelligence, and forgetting things often), it makes sense to build less Iron Man robots and more Iron Man suits.

Less building flashy demos of autonomous agents, and more building partial autonomy products.

The greatest predictor of startup success today is speed. And AI speeds up startups.

Early research has shown that AI can speed up writing production software by ~30-50%. But AI can speed up prototyping by ~10x.

Designing application architecture is now more of a two-way door than a one-way door. If something breaks when prototyping, it is now sometimes more cost-effective to rewrite from scratch than refactor entirely. AI has flipped typical software development conventions upside down.

Production software still requires another level of care and diligence. Wiring up integrations, patching up security concerns, and other considerations still require careful human oversight.

But for lean startups racing to product-market fit? Having concrete ideas, leveraging AI to bring those ideas to life, quickly capturing feedback, and adjusting based on that feedback is the new competitive advantage.

The same build-measure-learn feedback loop makes up the playbook. But now, AI unlocks a 10x speed advantage.

For complex tasks, training on broad datasets followed by finetuning on specific, consistent examples yields better performance.

I know very little about robotics. But watching Finn’s demonstration of her lab’s robot folding laundry made me feel like a kid in a candy shop. I want that!

Comparing performance across three training methods— training only on curated dataset, training on all the data, or a mix of both pre-training and post-training— demonstrated that the third method dramatically boosts performance.

Intuitively, this mirrors how humans learn new skills. First, the human generally learns about the world around them through diverse experiences. Then, the human intentionally practices a particular skill through high- quality repetitions. Only experiencing one or the other would hinder performance. Such is the case as well with robotics.

Find where people are outsourcing labor. Attack those markets.

Before: TAM = (# of seats) x ($ per seat).

Today: TAM = combined salaries of everyone completing a task.

Ask the question, how would the best X professionals in the world do their job with unlimited time and resources? Replace that work with reliable AI.

Most steps salary workers take in their jobs can be replaced by either deterministic algorithms or AI prompts. The more flexible in nature the process, the more agentic the solution.

To build truly reliable systems, half-baked, vague prompts won’t cut it. Well-defined evaluations help guide the system towards top-tier reliability. Heller offered a clear roadmap for achieving reliable AI:

• Ask: What does GREAT look like? For each outcome? For each micro-task? Start with dozens of evals per prompt (e.g. tell GPT to output T/F, or output a number between 1-7).
• Iterate until the system passes all test cases consistently. Then add 50 more. Realistically, this could take 2+ weeks. (Most people will stop iterating after a first round of prompts. Then, more people will drop out after a 1% improvement. More will drop out after a 2% improvement. In this game, however, EVERY % MATTERS!)
• Pre-release, ~100 evals per prompt and ~100 evals per task should exist (and system should pass all evals).
• Listen + Learn: every user complaint is an opportunity to add more evals
• KEEP ITERATING!

For your startup to succeed, you need to be irrationally optimistic and uncompromisingly realistic.

Mohan was irrationally optimistic at the beginning of his journey with Windsurf. Most assumed GitHub Copilot would be the clear winner in the AI dev-tools space (access to proprietary data, Microsoft distribution). But Mohan still believed his team’s grasp on training and running models themselves gave them a fighting chance at competing. His uncompromising realism guided him towards his pivot away from Exafunction and towards Codeium (now Windsurf).

“Strategic moats” are dead. Network effects, branding, etc. still exist, but they’re much more malleable now than ever before. Today, speed, focus, and catching the wave on the next innovation are what matter most.

Krishnan’s goal is to maximize American market share of tokens inferenced each month.

This is our next space race. The release of DeepSeek was “the Sputnik moment.” Right now, the competition with China is tight. If we stop progressing now, they could get ahead past the point of no return.

Learning The Bitter Lesson changed the way Krishnan thinks about AI. He reports sending someone the essay at least once a week. Internalizing the lesson and becoming “bitter-lesson pilled” sooner rather than later will speed up progress by maintaining focus on the research that truly moves the needle.