The Three Phases of AI Evolution
Alexander Embiricos, Head of Codex at OpenAI, defines the evolution of AI in three phases: simple search, multi-step orchestration, and finally, autonomous agency. He argues the primary bottleneck to AGI is not the models themselves, but the interface and orchestration required to make agents reliable at scale. This framework suggests curation algorithms will evolve from static ranking to dynamic, agent-powered systems.
- The evolution of OpenAI's Codex illustrates this three-phase framework: it began as a tool for simple, single-line code completion (Phase 1) and has now advanced to an autonomous agent that can independently handle complex tasks, like generating complete pull requests, for up to 30 minutes in its own cloud environment (Phase 3). - Embiricos argues the primary bottleneck to AGI is not model capability but human input/output speed; the time it takes for a person to type prompts and validate the AI's work is the limiting factor, pushing the need for agents that can also validate their own work. - The shift from "pairing" with an AI (like a co-pilot) to "delegating" tasks to an autonomous agent requires a change in user mindset; top users of the new Codex adopt an "abundance mindset," running 10 or more pull requests per day in parallel to see what works best. - This evolution directly impacts curation and personalization, moving beyond static content ranking to dynamic systems where AI agents can autonomously test variations, orchestrate cross-channel campaigns, and tailor content to individual user behaviors in real-time. - In practice, the new Codex agent has demonstrated significant autonomy, with one report indicating it had merged over 352,000 pull requests with an 85.5% success rate within a 35-day period. - While AI coding assistants are widely used, their impact on productivity shows a complex picture; one study of 121,000 developers found that while 26.9% of all production code was AI-authored, overall productivity gains have plateaued at around 10%. - The concept of "multi-step orchestration" (Phase 2) is being implemented in business operations through the coordination of multiple specialized AI agents; for example, a customer service journey can involve a chatbot for triage, a knowledge agent for answers, and a task automation agent for updates, all managed by a central orchestrator. - Studies on agent-driven productivity show significant gains in specific tasks, with one report indicating a 14% average increase in issue resolution for customer support agents, and another finding that developers using GitHub Copilot completed tasks 55% faster.
