Ethan Mollick traces the spectacular evolution of AI over three years, from GPT-3 to Gemini 3, marking a fundamental paradigm shift: the move from conversational chatbots to genuinely autonomous agents. This transformation redefines not only technical capabilities but also the human-machine relationship in intellectual work.
The first major development concerns code as a universal interface. Mollick observes that AI's ability to write code transcends pure programming. Since everything on computers ultimately runs on code, agentic systems can now autonomously manipulate dashboards, websites, presentations, and file systems. Gemini 3 no longer merely describes what could be done; it "codes the engine and designs the interface" directly.
The second change concerns autonomous task management. Google's Antigravity platform illustrates the evolution of the human-in-the-loop paradigm. Humans no longer correct errors but direct AI workflows through an inbox system, approving sensitive actions. This transition marks the shift from a reactive role (debugging) to a strategic role (orchestration).
The third development, the most striking, concerns advanced research capabilities. Gemini 3 conducted original academic research in a near-autonomous manner: hypothesis generation, data collection, sophisticated statistical analyses, producing a 14-page paper of journal quality. This demonstration suggests the emergence of a "PhD-level intelligence," capable not only of executing but of designing and conducting complex investigations.
Mollick tempers this enthusiasm with critical observations on persistent limitations. The system required human corrections for fine judgment and methodological refinements. Hallucinations, though reduced, persist in edge cases. Genuine creativity and scientific intuition remain domains where human intervention provides irreplaceable value.
The evolution of the human role is the central insight: from "error corrector" to "research manager." This transition reflects a fundamental reorganization of intellectual work in which humans define objectives and validate approaches while AI executes with growing autonomy.
Mollick concludes that these three years represent not an incremental improvement but a qualitative transformation. The gap between GPT-3 generating plausible text and Gemini 3 conducting autonomous research marks the emergence of a new category of intellectual tools, redefining the boundaries of the possible in cognitive augmentation.