Why I don’t think AGI is right around the corner

In this monologue podcast, the speaker narrates a blog post from June 2025, outlining their perspective on the timeline for achieving Artificial General Intelligence (AGI). The speaker expresses skepticism about AGI being imminent, particularly due to the current limitations of Large Language Models (LLMs) in continuous learning and adaptation. While acknowledging the impressive capabilities of current AI models, the speaker argues that their inability to improve over time like humans significantly hinders their practical application in transforming workflows. The speaker predicts that solving continuous learning will lead to a significant breakthrough, potentially resulting in a broadly deployed intelligence explosion, but anticipates a broken version of continual learning being released before a fully functional one. The speaker also expresses doubt about the near-term arrival of reliable computer use agents capable of handling complex tasks like taxes, citing challenges related to long horizon rollouts, limited pre-training data, and the complexity of algorithmic innovations. Despite these reservations, the speaker acknowledges the rapid progress in AI reasoning and sets tentative timelines for AI achieving human-level competence in specific tasks, such as tax preparation by 2028 and seamless on-the-job learning by 2032, while also noting that AGI is likely to occur this decade or experience significantly reduced probability.