>>11876197There is no addressing performance issues. Its a problem inherent to the current manifestation of AI/ML. You can make small optimizations and move algorithms onto silicon, but its only a temporary fix.
It's not virtue signalling. >90% of modern AI/ML research builds upon a shitty foundation and ends up reaching the point where no one outside of "learned" individuals can even have a change of understanding.
Just look at this from mathematics:
In mathematics, Hodge theory, named after W. V. D. Hodge, is a method for studying the cohomology groups of a smooth manifold M using partial differential equations. The key observation is that, given a Riemannian metric on M, every cohomology class has a canonical representative, a differential form which vanishes under the Laplacian operator of the metric. Such forms are called harmonic.
No one can understand this outside of specialists who have built their way up to understanding it. Instead of trying something new at a lower-level and possibly advancing the field in a huge way, they just keep building on top of the same tired knowledge, and from what I can tell, it isn't sustainable.
