AI Agents are a Decade-Long Trend: He consistently states that achieving truly capable, human-level agents is a 10-year endeavor, not an overnight success, characterizing the current era as the 'decade of agents.' [10, 26, 33, 38]
Strong Critique of Reinforcement Learning: He views standard RL as a highly inefficient and 'terrible' learning paradigm for complex tasks, arguing it extracts too little signal from experience. [7, 43, 60]
Intelligence Should Be Separated from Knowledge: He advocates for a research direction focused on isolating a 'cognitive core' from the vast memorized knowledge in LLMs, believing this knowledge may be holding back true reasoning capabilities. [11, 44]
Pragmatic Optimism on Open Source: He views the current 6-8 month lag of open-source models behind frontier models as a 'healthy' dynamic for the industry and speculates that distributed, open collaboration could eventually outperform centralized labs. [13, 16, 18]
The Primacy of Digital over Physical AI: He predicts progress in digital AI agents will vastly outpace robotics because manipulating bits is millions of times easier and cheaper than manipulating atoms. [15]
▶The Decade of AgentsFeb–Mar 2026
Karpathy consistently frames the current era not as the 'year of agents' but the 'decade of agents.' This theme encompasses his predictions of a 10-year timeline for maturation [33, 38], his personal workflow shift to delegating nearly all coding to agents since December 2023 [50], and his hands-on work creating agents for home automation and auto-research. [47, 49]
This long-term framing suggests that the primary investment and innovation opportunities will be in building the infrastructure, tooling, and applications for agentic workflows, rather than expecting a single 'AGI' breakthrough to immediately change everything.
▶Critique of Current AI ParadigmsFeb 2026
Karpathy expresses significant skepticism about the foundations of current AI development. He criticizes the inefficiency of reinforcement learning ('sucking supervision through a straw') [7], the low quality of internet pre-training data ('total garbage') [5], the lack of diversity in synthetic data ('silently collapsed') [4], and the over-reliance on memorization. [11]
Karpathy's focus on these fundamental flaws indicates that he believes true progress is bottlenecked by methodology, not just scale, creating opportunities for startups or researchers with novel approaches to data generation, training, and model architecture.
▶The Cognitive Core vs. Memorized KnowledgeFeb–Mar 2026
A key research direction proposed by Karpathy involves separating an AI's learned knowledge from its core reasoning ability. He argues that the vast memorized data from pre-training may hinder true intelligence [11] and proposes isolating a 'cognitive core' that could be as small as one billion parameters. [9, 44]
This pursuit of smaller, more efficient reasoning engines over massive knowledge databases could disrupt the current compute-heavy landscape, potentially lowering barriers to entry and shifting the competitive advantage from data hoarders to those who can build the most effective 'cognitive core'.
▶Recursive Self-Improvement and the Future of AI DevelopmentMar 2026
Karpathy identifies recursive self-improvement—using LLMs to improve the next generation of LLMs—as the central goal of all frontier labs. [17] He extends this concept to his own work with an 'auto-researcher' agent that improves model hyperparameters overnight [49] and envisions a future where a distributed 'swarm of agents' could collectively improve models, potentially outpacing centralized labs. [13]
This theme highlights that the process of building AI is itself becoming the product. Companies that can create the most effective feedback loops for AI-driven development will likely gain a compounding advantage that is difficult for competitors to overcome.