4/n James Webb Telescope took hundreds of scientists and engineers 5 years to design + 20 years to build and cost $10 billion. LLMs use 100s of thousands of CPUs & cost billions to train... Researchers in climatology, physics, applied math, etc. use tens of thousands of CPUs for simulations.
Comments
Do we think macro research is less consequential? The global financial cost of 2008 Global Crisis is estimated to be ~$2 trillion.
So, an algo that solves K-S in 10 secs vs. 10 minutes means nothing - unless it’s accurate.
This brings us to Accuracy for Large Shocks... -->
Background:
In HA models, linearization-based methods start with Reiter (JEDC, 2009). It is nonlinear in individual states but imposes linearity in aggr states. It solves Khan & Thomas (ECMA, 2008) model about 100x faster than Krusell-Smith method.
These methods promise to be much faster than even Reiter’s method.
So, I am posting my concerns here, hoping that it will lead to a discussion of the pros and cons of these approaches and that hopefully, the authors can respond and persuade us.