Paul Graham Says He Would Learn to Build LLMs If He Were 17

Y Combinator Co-Founder Paul Graham and AI researcher Yann LeCun have offered contrasting answers to a question about what they would do if they were 17 today.

Graham said he would focus on learning to build and train large language models. LeCun, meanwhile, said he would look beyond LLMs towards AI systems that can understand and perform physical tasks.

Graham added that he would use his teenage years to develop a deep understanding of LLMs rather than immediately trying to build a startup.

“Someone asked what I’d do if I were 17. I’d learn how to build LLMs from scratch, and then train ones as powerful as I could with whatever hardware I could get access to,” Graham wrote on X.

He followed that by explaining why he would not prioritise entrepreneurship at that age.

“Notice that what I would not do is try to start a startup,” Graham wrote. “Instead I'd build the foundation of knowledge to base a startup on later.”

Graham argued that understanding the underlying technology would be more valuable than trying to identify a business opportunity without that technical foundation. 

“Way better startup ideas would grow out of understanding LLMs deeply than trying to start companies based on what I knew at 17,” he added.

Meanwhile, LeCun wants AI to move beyond language

He took a different approach, focusing on one of the limitations of current AI systems. Responding to Graham, he said he would investigate why LLMs can perform language-based tasks but struggle with basic interactions with the physical world.

“I would try to figure out why LLMs can write my essays but not clean my bedroom,” LeCun wrote. “I’ll figure out a set of methods and architectures beyond LLMs that can quickly learn to perform physical tasks as efficiently as humans and animals.”

LeCun added that this would remain his focus regardless of age, saying it is what he would do “if I were 30, 40, 50, or 66 years old”.

The comment comes as LeCun is building a company around that same idea.

He co-founded Advanced Machine Intelligence, or AMI Labs, after leaving Meta in late 2025. The Paris-based startup is developing world models that are intended to understand the real world, retain memory, reason and plan.

AMI Labs raised $1.03 billion in March at a $3.5 billion pre-money valuation. The round was led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, with other investors including NVIDIA and Samsung.

The company’s approach is based in part on LeCun’s Joint-Embedding Predictive Architecture, or JEPA, which seeks to make AI learn representations of the world and predict how it will change rather than simply predict the next piece of text.

AMI CEO Alexandre LeBrun has said the company wants its systems to work with robotics, manufacturing and other industries that operate in the physical world. 

He has also described LLMs and world models as complementary, with language models handling language while world models provide an understanding of physical environments.