OpenAI co-founder: AI agents are still 10 years away (thenewstack.io)

In a wide-ranging podcast interview, OpenAI co‑founder Andrej Karpathy argued that practical, autonomous AI agents are still roughly a decade away. He framed this as a corrective to current hype: while large language models have driven huge recent progress, they still lack many capabilities needed for robust agents — multimodality, reliable long‑term memory/continual learning, real-world sensor and actuator integration, safe behaviour under adversarial conditions (jailbreaks, data poisoning), and product‑grade reliability. Drawing on his Tesla self‑driving experience, Karpathy emphasized a persistent demo‑to‑product gap where prototypes look impressive but fail when hardened for safety‑critical deployment (he noted tele‑operators remain common). For practitioners the takeaway is both technical and programmatic: the path to agentic AI is tractable but dominated by engineering, systems integration, safety, compute and diffusion challenges rather than a single algorithmic breakthrough. Karpathy’s new education-focused venture, Eureka Labs, and his nanochat minimal full‑stack LLM aim to address the knowledge gap by giving engineers a compact, hackable implementation to learn from — highlighting that building and understanding real systems is essential work. His stance is cautiously optimistic: significant shifts will come, but expect years of “grunt work” before agents replace humans across arbitrary tasks.