OpenAI: Typing Speed Bottlenecking AGI, 2026 Productivity Su

Alexander Embiricos, head of OpenAI's Codex, has suggested that human typing speed and multitasking abilities are hindering the development of artificial general intelligence (AGI). Embiricos forecasts that by 2026, AI's self-auditing capabilities will trigger a "hockey stick" surge in productivity, propelling humanity toward AGI.

Human Input as an AGI Constraint

Embiricos articulated this perspective on "Lenny's Podcast," identifying human typing speed as a limiting factor in the AGI timeline. This view aligns with observations from other prominent figures, including OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, and Tesla CEO Elon Musk, who have described the issue as a "communication" or "friction" problem in human-computer collaboration.

Altman has consistently stated that AI progress will not solely depend on larger models. He cited the development of AgentKit, a tool designed to reduce friction and enhance efficiency for developers building intelligent agent systems, as an example of addressing this bottleneck.

Amodei noted in April that AI could write 90% of code within three to six months and "basically all code" within 12 months. However, he emphasized that programmers would still need to define functions, application design, and decisions, underscoring the ongoing importance of human involvement in the short term.

Musk, in discussions about Neuralink, has also highlighted the disparity between human data transmission speeds and computer processing. He suggested that brain-computer interfaces could bypass the limitations of human input, likening human communication with computers to "very slow, tonal gasps, somewhat like whale sounds."

These perspectives collectively indicate that human capabilities, particularly text input, are becoming an impediment to AGI development. This parallels the early computer era, where manual code input was a significant constraint. As AI's processing power accelerates, human input speed is increasingly seen as a bottleneck.

The Challenge of Prompt Engineering

Embiricos's assessment stems from his observations of AI agent systems in practical applications. He noted that while current AI models excel at complex tasks, they still rely heavily on human guidance and verification through text input. The core difficulty lies in prompt engineering, which demands precise expression of human intentions.

He explained that human multitasking, involving the switching of working memory, introduces inefficiency in prompt engineering. Developers using Codex, for instance, must simultaneously monitor AI output and adjust prompts. Embiricos compared this to a driver manually controlling an autonomous vehicle, where humans engage in a complex interplay of problem-solving, prompt writing, and output review. This workload, he argued, extends beyond simple typing speed, forming a significant barrier to AI productivity breakthroughs.

An example cited was the development of the Android version of the Sora app, which OpenAI reported was completed by a four-person team in 28 days. Approximately 85% of the code was generated by GPT-5.1+Codex, achieving a 99.9% stability rate. Despite the accelerated development, the process highlighted AI programming's weaknesses. Codex, described as a "novice senior engineer with zero memory," required continuous clear instructions through prompts to generate quality code. Human intervention for supervision, review, logical confirmation, and debugging became the slowest and most critical part of the engineering process, making typing speed a tangible productivity barrier.

Ilya Sutskever, former chief scientist of OpenAI, previously discussed a "performance paradox" where models perform well in evaluations but struggle in real-world scenarios. He provided an example of a model getting stuck in an infinite loop while fixing bugs, oscillating between two issues. This illustrates that current large model training methods often prioritize evaluation metrics over human-like understanding and generalization, reinforcing the idea that human capabilities can limit AI productivity in practical applications.

Liberating Humanity for AGI

Embiricos posits that to overcome the "human typing speed bottleneck," AI agents must develop the ability to review their own work without human dependence. He stated that for a system to achieve "hockey stick" growth, agents must be sufficiently useful by default, reducing the need for human validation and prompt writing.

He predicts that starting next year, a cohort of early adopters will experience this "hockey stick" increase in productivity, with more large companies following suit in subsequent years. Embiricos believes this productivity surge will eventually feed back into AI research, bringing AGI within reach by 2026.

Demis Hassabis, founder of Google DeepMind, has also emphasized the transitional nature of human-computer collaboration before more autonomous systems emerge. He suggests that AI needs to further develop its reasoning, autonomy, and creativity, estimating that machines may take another five to ten years to master all human capabilities.