I don't think AGI is right around the corner (2025)

The case against near-term AGI rests on a gap between impressive one-off performance and the ability to function as a dependable employee. Current language models can handle many tasks tolerably, but they do not learn from experience in the persistent, subtle way people do: they struggle to retain feedback, build context, recognize recurring mistakes, and improve their workflows. Prompts and session summaries only partly preserve what a model has learned, while reinforcement learning requires task-specific training rather than everyday adaptation. If progress stopped now, the author estimates that fewer than a quarter of white-collar jobs would disappear: many tasks could be automated, but limited learning would keep models from taking responsibility for whole roles. Solving continual learning, by contrast, could eventually transform the economy as experience gained by one deployed model is shared across its copies.

The same caution applies to forecasts of reliable computer-use agents arriving soon. Long, multi-step tasks are difficult to evaluate and require costly visual processing; training data for interacting with software is far scarcer than text, and generating practice environments or devising effective training methods may take substantial engineering. The author therefore doubts that agents will soon competently manage complex work such as filing taxes. This skepticism is not a dismissal of current advances: reasoning models can break problems into steps, reconsider weak approaches, and produce genuinely thoughtful answers. The broader view is that progress is real but uneven, with continual learning and robust agency remaining major bottlenecks; once those are overcome, the resulting change could be abrupt and far-reaching.