🎇 Yikes! What if AGI isn't possible with LLMs?

It would be a bummer but not necessarily the end of humanity's journey toward superintelligence

My fellow pro-growth/progress/abundance Up Wingers,

I write frequently about the possible creation and potential impact of artificial general intelligence, AI that can match or exceed humans across many mental tasks — not just one niche job — and can learn new skills on its own.

And with good reason. The tantalizing prospect of such advanced AI systems — experts who used to talk about AGI being 50 years away now see a timeline of five years or even sooner — has galvanized unprecedented investment. More than $1 trillion in American business capital has flowed into generative AI alone. From routine automation to scientific breakthroughs, the promise is incredibly vast. AGI might unlock unprecedented economic acceleration. The full Up Wing, in other words.

Metaculus

Then there’s the national security dimension. In "Superintelligence Strategy," Dan Hendrycks, Eric Schmidt, and Alexandr Wang argue that advanced AI represents a transformative force-multiplier for military capabilities, potentially compressing decades of defense R&D into a single training run. Superweapons ahead, let’s America owns them first. As the analysts state, "A nation with sole possession of superintelligence might be as overwhelming as the Conquistadors were to the Aztecs."

Given those ginormous stakes, there’s also good reason to consider the unpleasant possibility that hyperscaling these amazing large language models — the idea that simply making LLMs ever bigger with more data and more computer power — won’t actually push them all the way to AGI. It’s a scenario explored in a new Rand Corporation report, “Charting Multiple Courses to Artificial General Intelligence” by William Marcellino, Lav Varshney, Anton Shenk, Nicolas M. Robles, and Benjamin Boudreaux.