AI pioneer and former OpenAI and Tesla researcher Andrej Karpathy believes the industry is still years away from achieving true Artificial General Intelligence (AGI), despite the growing hype surrounding “agentic AI.” Speaking in a recent interview, Karpathy described the excitement around autonomous agents as “far ahead of reality”, suggesting that meaningful progress will take steady, grounded work rather than leaps of faith.
Karpathy explained that today’s large language models — no matter how powerful they seem — still fall short of core requirements for general intelligence. Modern systems, he said, are essentially autocomplete engines lacking true memory, reasoning, multimodal understanding, and continual learning. While these models can produce text that feels intelligent, they don’t truly “understand” or retain information across sessions, limiting their reliability as real assistants or autonomous agents.
He also highlighted that training data quality remains one of the biggest barriers to progress. Most AI models are trained on massive but messy internet datasets filled with low-quality or repetitive content. To move forward, Karpathy argued, the AI community needs to prioritize curated, high-quality datasets, along with smarter architectures and improved learning methods — not just scaling up model sizes.
Karpathy was particularly critical of current reinforcement learning techniques, describing them as “terrible”, though he acknowledged they remain the best available option for now.
Despite his critiques, Karpathy’s outlook wasn’t pessimistic — just realistic. He envisions a future where breakthroughs come from better engineering, cleaner data pipelines, and modular AI systems rather than hype-driven experimentation.
Why It Matters
Karpathy’s projection — that AGI may still be a decade away — stands as a much-needed reality check amid Silicon Valley’s race to build “agentic” systems. His message emphasizes a more sustainable path forward: focus on incremental improvements, human-in-the-loop designs, and practical AI applications that provide real value today, instead of betting everything on AGI arriving tomorrow.
🔗 Full interview: Dwarkesh Podcast – Andrej Karpathy
Key Takeaways
True AGI could still be 10 years away. Current models lack memory, reasoning, and continual learning. Data quality is the biggest obstacle to progress. Progress will come from smarter architectures and cleaner training pipelines. The focus should shift from hype to real, modular, useful AI systems.
Leave a Reply