In 2025, Andrej Karpathy delivered a one-hour lecture at Stanford University on the topic of AI engineering. The talk, which drew on his extensive experience in deep learning and artificial intelligence, provided a technical overview of large language models (LLMs), prompting strategies, and the emerging paradigm of AI agents. A central theme was the importance of graph-based approaches for structuring and reasoning about complex AI systems, a perspective that has influenced subsequent discussions in the field.
Karpathy, a Slovak-Canadian researcher, is known for his contributions to computer vision and natural language processing. He co-founded OpenAI, served as director of AI at Tesla, and later founded Eureka Labs, an AI education platform. His Stanford lecture reflected his pedagogical focus, aiming to make advanced concepts accessible to a broad audience of students and practitioners.
Background and Context
Karpathy's lecture came at a time of rapid advancement in generative AI, following the widespread adoption of transformer-based models. His previous work included teaching Stanford's CS 231n course, which became one of the university's largest classes, and creating popular educational content on neural networks. The 2025 lecture continued this tradition, offering a structured introduction to the practical engineering of AI systems.
The talk also occurred against the backdrop of Karpathy's public advocacy for accessible AI education. In 2024, he launched Eureka Labs with the goal of creating an AI-native school, and his lecture series, including the "Zero to Hero" tutorials, have been widely used by learners. The Stanford lecture served as a condensed version of these efforts, targeting an academic audience.
Core Topics: LLMs, Prompts, and Agents
The lecture covered three interconnected pillars of modern AI engineering. First, Karpathy explained the architecture and training of large language models, emphasizing the role of the Transformer (architecture) architecture and techniques like Multi-Head Attention. He discussed how these models are trained on vast corpora of text, using methods such as Supervised learning and RLHF to align outputs with human expectations.
Second, he addressed prompting, the practice of designing inputs to elicit desired responses from LLMs. Karpathy highlighted the importance of clear instructions, context, and iterative refinement, noting that effective prompting often requires understanding the model's Tokenization and temperature parameters. He also touched on advanced techniques like Chain-of-thought prompting, which encourages models to reason step by step.
Third, the lecture introduced AI agents - systems that use LLMs to perform tasks autonomously. Karpathy described how agents can be built by combining LLMs with tools, memory, and planning mechanisms. He emphasized that agents often require careful orchestration of multiple components, and he proposed graph-based models as a way to represent and manage these interactions.
Graph-Based Approaches
A distinctive contribution of the lecture was Karpathy's advocacy for graph-based approaches in AI engineering. He argued that many AI systems, from neural networks to agent pipelines, can be naturally represented as graphs, where nodes correspond to components and edges to data flows or dependencies. This perspective, he suggested, facilitates debugging, optimization, and scalability.
For example, Karpathy discussed how computation-graphs are used in deep learning frameworks to enable automatic differentiation and efficient training. He extended this idea to higher-level systems, proposing that agent workflows could be modeled as directed acyclic graphs, with each node representing a step such as Information retrieval or Tool use (function calling). This approach, he noted, aligns with practices in software engineering, where modularity and clear interfaces are key.
Karpathy also referenced his own experience at OpenAI and Tesla, where graph-based thinking helped manage complex, real-world AI deployments. He encouraged the audience to adopt such mental models when designing their own systems.
Reception and Impact
The lecture was well received by the Stanford community and the broader AI research community. Attendees praised its clarity and depth, noting that Karpathy's ability to distill complex topics into intuitive explanations is a hallmark of his teaching style. The talk was recorded and later shared online, where it accumulated significant viewership, further extending its reach.
Following the lecture, several discussions emerged in academic and industry circles about the utility of graph-based approaches for AI agent design. Some researchers began exploring frameworks that explicitly model agent interactions as graphs, building on the ideas Karpathy presented. The lecture also reinforced Karpathy's reputation as a leading educator in AI, complementing his ongoing work with Eureka Labs.
Legacy and Continued Influence
Karpathy's 2025 Stanford lecture is part of a broader trajectory of contributions to AI education and research. His earlier work on convolutional neural networks and recurrent neural networks helped shape modern deep learning. His subsequent ventures, including his return to OpenAI and later move to Anthropic in 2026, have kept him at the forefront of AI development.
The lecture's emphasis on graph-based approaches has been cited in subsequent papers and tutorials, particularly in the context of agent orchestration. As AI systems become more complex, the need for structured representations like graphs is likely to grow, making Karpathy's insights increasingly relevant. The talk remains a valuable resource for students and practitioners seeking to understand the practical aspects of AI engineering.