AI innovators highlight human input speed as obstacle to full autonomy

Monday 15 December 2025 - 12:50

By: Dakir Madiha

AI innovators highlight human input speed as obstacle to full autonomy

OpenAI’s product lead for Codex, Alexander Embiricos, believes that the transition from human-supervised to autonomous artificial intelligence hinges on overcoming a surprising limitation: human typing speed. Speaking on Lenny’s Podcast, he noted that even highly capable AI agents remain constrained by the need for continuous user input and manual validation. He described this dependence as a “hidden bottleneck” slowing progress toward artificial general intelligence (AGI).

Embiricos explained that while AI systems are increasingly able to observe and learn from human workflows, their effectiveness remains limited if users must constantly approve or correct their outputs. “You can have an agent watch all the work you’re doing, but if you don’t have it validating its own process, you’re still stuck reviewing everything,” he said.

The OpenAI executive suggested that the next major breakthrough will come when these agents can perform tasks with minimal human supervision, becoming "default useful" rather than reactive tools. He predicted that smaller teams and individual innovators will begin seeing exponential productivity gains by 2026, with widespread adoption following among large enterprises soon after.

Embiricos linked these trends to the surge in usage of OpenAI’s Codex, which has grown twentyfold since August and now processes trillions of tokens each week. This expansion signals a shift toward agentic AI capable of handling multi-step operations autonomously. Still, he cautioned that building fully self-directed systems will depend on case-specific designs and iterative innovation.

Analysts observe that major AI labs are now engaged in a race to reach AGI, with some forecasters estimating a one-in-four chance of its arrival by 2027. For Embiricos, the rapid scaling of autonomous code generation and workflow management tools marks the beginning of that acceleration curve—one that could redefine how humans collaborate with intelligent systems.