Prioritize Deep AI Expertise Over Early Startups

Paul Graham recommends that 17-year-olds focus on building LLMs from scratch to create a strong knowledge base for future startups. Yann LeCun adds that they should investigate why AI struggles with physical tasks and seek architectures beyond current LLMs. Together, they advocate for deep technical exploration over early entrepreneurship.

Key Points

• Paul Graham advises 17-year-olds to master building and training LLMs from scratch rather than starting a business immediately.
• Deep foundational knowledge of AI is expected to yield much better startup ideas in the long run than early entrepreneurship.
• Yann LeCun highlights the gap between AI's cognitive abilities and its inability to perform simple physical tasks.
• LeCun encourages the study of new architectures beyond LLMs to solve the problem of physical task execution.
• Both experts emphasize the importance of academic and technical depth over superficial application.

Sentiment

The overall sentiment is mixed and moderately contentious. A substantial share of commenters agrees with the spirit of the article: young technologists should develop deep, hands-on expertise in important new systems instead of rushing into shallow startup work. At the same time, many push back on the literal framing of building LLMs from scratch, arguing that it understates the required math, data, infrastructure, and compute, and that it reflects startup-world hype more than practical guidance. The most constructive middle ground is that small-scale models, fine-tuning, and toy experiments are worthwhile educational projects, while industrial-scale language model work is not a realistic benchmark for most learners.

In Agreement

• Building small language models or experimenting with fine-tuning is a modern version of the classic hacker path: learn the important technology of the era by taking it apart and rebuilding simplified versions.
• The point is not to match frontier labs, but to gain intuition about data, training, loss functions, model behavior, and inference through toy systems and accessible hardware.
• Young people often learn best by following tutorials, copying examples, and tinkering before they fully understand the theory; that exploratory route can later mature into real competence.
• Deep technical understanding is a better foundation for future startups than launching yet another thin product built around API calls.
• LeCun's emphasis on the limits of language models resonates with commenters who think the bigger opportunity is understanding architectures beyond current text-centric systems.
• Several commenters argue that ambitious technical hobbies should be encouraged rather than dismissed, especially because many HN users learned programming through similarly obsessive teenage projects.

Opposed

• Many commenters argue that building language models from scratch is not a realistic beginner project because it demands substantial math, training data, infrastructure knowledge, and compute.
• Some object that following tutorials or copying a small transformer is not truly building an LLM from scratch and may create cargo-cult understanding rather than deep expertise.
• Others think the advice reflects an older founder's hindsight more than a teenager's lived reality, priorities, or developmental needs.
• A recurring criticism is that the recommendation is too shaped by AI hype and YC incentives, pushing young people toward whatever the startup ecosystem currently wants.
• Some commenters would prefer teenagers to explore broadly, enjoy youth, or learn more accessible technical foundations before attempting modern transformer systems.
• Part of the thread turns into broader skepticism toward Paul Graham, venture capital, OpenAI, and startup culture, with some commenters treating the advice as inseparable from those institutions.