A new artificial intelligence research organization named Flapping Airplanes was unveiled this Wednesday, launching with a substantial $180 million in seed capital provided by Google Ventures, Sequoia, and Index Ventures. The founding team is notably accomplished, and their objective—to discover methods for training large models that require less data—stands out as a particularly compelling ambition. Based on the information available at this stage, their approach to generating revenue appears to be at an early, formative level.

However, there is an even more intriguing aspect to the Flapping Airplanes initiative that became clearer upon reading an analysis by Sequoia partner David Cahn. As Cahn outlines, this lab is among the pioneering groups to shift focus away from the prevailing industry trend of scaling, which has been characterized by an intensive accumulation of data and computing power.

The conventional scaling perspective advocates for directing a massive portion of societal and economic resources toward enlarging current large language models, with the expectation that this will eventually lead to artificial general intelligence (AGI). In contrast, the research paradigm suggests that achieving AGI-level intelligence is likely contingent on just a few key scientific breakthroughs. This viewpoint supports allocating resources to extended, foundational research projects, including those that may require five to ten years to fully develop.

A strategy centered primarily on computing power would emphasize expanding data center infrastructure above all else, naturally favoring shorter-term achievements measurable within one to two years over longer-term, more uncertain ventures. A research-driven strategy, however, would distribute investments across various time horizons and be open to supporting numerous projects, each with a relatively low individual chance of success. The collective aim would be to broadly explore new technological possibilities.

It remains possible that the proponents of massive compute scaling are correct, and that focusing on anything other than rapid infrastructure expansion is inefficient. Yet, with so many firms already committed to that path, it is refreshing to observe an organization charting a different course.