It Ain't AGI Until H100s Go for $30 an Hour

You can't just multiply out the numbers because it assumes knowledge of the supply and demand curves that aren't available. Namely, we don't have access to the emperical supply and demand curves of any good or service.

Every supply and demand curve we've ever seen is a pedagogical tool for teaching the concept, not for what actually exists in the world. The author mistakes the teaching model as fact invalidating all downstream conclusions.

It's likely that the shift from Opus -> AGI will (at least for an initial design), be inefficient. Engine designs, rocket designs, manufacturing equipment designs, computer programs, etc., are built for results rather than efficiency.

Because of this, would expect whatever we consider to be "v1 of AGI" to require at least 10x as much compute as today's Opus.

Plugging this into the calculator, I see that it gives me a GPU price of $3 an hour, telling me we've already achieved AGI.

---

Realistically, I think the author has the correct direction, but there are many supply and demand curves, each with their own research frontiers focused on growth and efficiency. Simplifying the variable space may result in an equation that fits today's and prior data, but I would not rely on it for future predictions.

jagged frontier means AI may become vastly superhuman in capabilities in many areas that matter (even up to existential or mass disempowerment risk) without reaching this definition of AGI.

This is idiot economics. Ai is already taking over jobs by undercutting them, similar to offshoring.

At minimum wage of $7 and Anthropic's profit margin being closer to 10% we are already there

Why exactly aren’t the AI companies using this compute to get rich themselves?