OpenAI cofounder Andrej Karpathy warned that expectations for artificial general intelligence are outpacing reality, saying in a podcast interview that current large language models remain far from delivering reliable, agentic performance. Karpathy put AGI at least a decade away and criticized the industry’s reliance on demos and benchmarks that he said overstate progress, particularly around autonomous “agents” marketed as digital workers. His remarks contrast with bullish forecasts from leaders including OpenAI CEO Sam Altman and Nvidia’s Jensen Huang, and follow growing skepticism such as AI pioneer Richard Sutton’s recent critique of LLMs. Karpathy cautioned that immature agent systems could create “slop,” exacerbate software vulnerabilities and security risks, and disappoint investors chasing quick wins. Still, he called the technical hurdles surmountable with time, research, and stronger safety practices—tempering the boom narrative without dismissing long-term potential.
Related articles:
LLM-powered Autonomous Agents
AgentBench: Evaluating LLMs as Agents
Auto-GPT (open-source autonomous agent)
Reflexion: Language Agents with Verbal Reinforcement Learning
Voyager: An Open-Ended Embodied Agent in Minecraft
