The discussion debates whether AI progress is slowing. While the era of massive gains from simply scaling up compute and pre-training data is hitting diminishing returns, progress is now shifting to other vectors. These include increased efficiency, where smaller models match the capabilities of last year's giants, and enhanced reasoning abilities driven by post-training techniques like Reinforcement Learning.
Frontier labs are aggressively investing in Reinforcement Learning to push model capabilities, particularly in domains with verifiable outcomes like math and coding. This has created a massive demand for RL environments and the talent to build them, with labs setting aside virtually unlimited budgets for this purpose.
The extreme salaries for top AI talent are justified by the concept of talent as a 'compute multiplier.' A skilled researcher can significantly increase the efficiency and performance of a model for a given amount of compute, making their high cost a rational investment. This competition has also led to novel 'effective acquisition' structures to secure entire teams.
The conversation identifies two primary paths to overcome current AI limitations. The first is recursive self-improvement, where an AI is used to design a better AI, as the most likely path to a major leap in intelligence. The second is the strategic use of synthetic data to bypass the 'data wall' and continue improving models as high-quality human data becomes scarce.
The episode explores the sky-high valuations of AI labs like xAI, questioning whether they are justified, particularly for companies seen as 'me too' players. It also highlights the significant challenge of monetizing AI 'answer engines' without compromising user trust, a hurdle that even giants like Google will face in their transition.
Keep pulling the thread on Ari Morcos & Rob Toews.