ai thought leaders
Below are the ideas of influential Thought Leaders in AI.
Andrej Karpathy (@karpathy)
Andrej Karpathy is a leading voice in applied AI, with experience at Tesla, OpenAI, and Stanford. His focus is on building practical, large-scale AI systems and communicating deep learning concepts clearly.
Software 2.0
Karpathy describes the shift from traditional programming to neural networks as “Software 2.0,” where learning replaces manual code.
End-to-End Deep Learning
He advocates for training entire systems directly from data inputs to outputs, removing hand-engineered steps where possible.
The Hacker Way for AI
Karpathy encourages learning by doing. The idea is to experiment with code, building models, and iterating often, as a way to develop intuition.
Neural Nets as Differentiable Programs
He conceptualizes neural networks as soft, modular, and differentiable computational graphs that can be composed like functions.
Understanding GPT Through Tokens
He emphasizes that understanding tokenization and next-token prediction is key to demystifying how GPT models work.
Yann LeCun (@ylecun)
Yann LeCun is a foundational figure in AI research, known for his work on convolutional neural networks and as a vocal proponent of self-supervised learning and modular AI systems. As Chief AI Scientist at Meta, his work shapes the long-term vision of general intelligence.
Self-Supervised Learning
LeCun believes SSL is the most promising path toward human-like AI, enabling models to learn from unlabeled data.
World Models and Common Sense
He argues that AI must build internal models of the world to reason, plan, and act—just as humans do.
Energy-Based Models
LeCun supports the idea of energy functions that evaluate the plausibility of outcomes, providing an alternative to probabilistic modeling.
Modular AI Architectures
He promotes AI systems with separate but interacting modules—such as perception, memory, planning, and reasoning—to mirror human cognition more effectively.
Critique of Reinforcement Learning
LeCun has expressed skepticism about reinforcement learning as a dominant approach to intelligence, pointing to its limitations in scalability and sample efficiency.