The speaker argues that the most significant missing piece for AGI is the ability for models to learn continuously from experience and feedback in a real-world environment. This contrasts with the current paradigm of static, pre-trained models that cannot adapt without extensive, custom retraining loops.
A core tension is explored between the current AI lab strategy of 'baking in' skills (e.g., using Excel) via RL and the human ability to learn novel, context-specific tasks on the fly. The speaker contends that the latter is essential for most knowledge work, making the current approach fundamentally limited.
The speaker observes a divergence where AI models are becoming more impressive at the rate predicted by optimists, but more economically useful at the slower rate predicted by skeptics. He uses the vast gap between the tens of trillions in knowledge worker wages and AI lab revenues as evidence that current models are not yet true substitutes for human labor.
The speaker predicts that the development of AGI will not be a singular event creating a monopoly. Instead, progress will be incremental, and breakthroughs will be rapidly replicated by competitors due to talent poaching, reverse engineering, and intense industry competition, neutralizing any single lab's advantage.
Keep pulling the thread on Microsoft Excel.