Andrej Karpathy
AI researcher behind the LLM Wiki pattern and early OpenAI computer-use agents; argues that model capability must precede reliable agent products and that demos can hide decade-scale engineering work.
Andrej Karpathy is an AI researcher and builder referenced here both as the source of the LLM Wiki Pattern and as an early OpenAI agent researcher. In the World of Bits project, he worked with Tim Shi and Jim Fan on reinforcement-learning agents that used simplified websites through keyboard and mouse actions.1
Karpathy’s retrospective is a model-first warning for agent builders: World of Bits did not fail because computer-use agents were inherently a bad idea, but because reinforcement learning and the available models were not ready. The productive detour was to build stronger language models; today’s agents returned with a different toolkit.1
He also warns that agent demos and agent products are separated by years of reliability work, comparing the gap to self-driving and VR. At the same time, he treats a language model as only one part of a digital cognitive architecture, which may also need planning, reflection, memory retrieval, sensory processing, and information-routing mechanisms inspired by neuroscience. This complements Harness Engineering: scaffolding can expose model capability, but cannot manufacture capability that the underlying model lacks.1
In the blog
SummaryAI agents are easy to demo and hard to make real