Overview
- On the Dwarkesh Podcast, OpenAI cofounder Andrej Karpathy said current agents “just don’t work,” citing insufficient intelligence, lack of multimodality, weak computer-use skills, and no reliable continual learning.
- In follow-up posts on X, he criticized the industry for overshooting present capabilities and for envisioning fully autonomous code-writing entities that leave humans “useless.”
- Coverage contrasts his stance with investor narratives branding 2025 the “year of the agent,” as he projects roughly ten years of foundational work before agents are broadly reliable.
- He favors collaborative workflows where systems show their sources, ask for guidance, and learn with users, noting he still uses tools like Claude and Codex but finds them far behind human performance.
- Enterprise reporting notes real-world friction, with many pilots faltering and firms scaling back automation plans, while Karpathy’s timeline also stands in contrast to Sam Altman’s prediction that agents would join the workforce in 2025.