Robin Hansen

Beyond the San Francisco Bubble Discussions regarding Artificial General Intelligence (AGI) often shift based on proximity to tech hubs. In San Francisco, timelines feel aggressive, sometimes predicting a total shift in just two years. However, stepping outside that environment reveals a different reality. While raw processing power continues to climb, the distance between solving isolated coding problems and replacing a human worker remains vast. The optimism of the valley often ignores the messy, organic nature of professional growth and the nuanced layers of human contribution. The Failure of Continual Learning What makes a human worker indispensable isn't just their initial skill set; it is their capacity to build context over months and years. Current Large Language Models (LLMs) suffer from a "Groundhog Day" effect. They exist session-to-session, losing the specific knowledge of a user’s preferences and failures as soon as the window closes. A human employee becomes valuable because they interrogate their own mistakes and refine their approach. Models like ChatGPT provide high-quality output for self-contained tasks, but they cannot yet mirror the trajectory of a person who learns to anticipate needs through shared history. The Coding Mirage Coding has seen explosive AI progress because of the massive, structured repositories available on GitHub. This creates a mirage where it seems AGI is imminent. When a machine writes fifty files of working code in thirty minutes, it feels like magic. Yet, this success is difficult to replicate in other white-collar fields or robotics where data is less organized. Dwarkesh Patel notes that while these systems are objectively intelligent, they lack the "on-the-job training" instinct required for complex, collaborative labor. The Unpredictable Horizon Predicting the future of AI is notoriously difficult, even for experts. In his 2014 book Superintelligence, Nick Bostrom explored brain uploading and misalignment but failed to foresee the specific rise of deep learning as the primary catalyst. This history reminds us that the architecture for true AGI might not even be the one we are currently using. We must remain humble about our timelines, recognizing that the next leap often comes from a direction no one is looking toward.