I see this as entirely surmountable. We’re still making geometric progress in small model accuracy, and breakthroughs like test-time training and synthetic data are poised to deliver immediate gains in self-training performance.

Your point about skepticism being warranted when viewing this linearly is well taken. But this isn’t a linear path. The Internet, at its core, was about connecting computers to unlock the value of those connections—a transformative but relatively straightforward concept.

What we’re dealing with now is the training of cognitive digital intelligence. This is an inherently dynamic and breakthrough-oriented process, one that evolves in ways far less predictable or constrained than simple network effects. While the metaphor of connectivity is useful, it doesn’t fully capture the parallel, multi-dimensional approaches at play here.

Pessimism, in my view, is deeply unwarranted, especially given the history of technological progress. Time and again, advancements have proven to be far more impactful and beneficial than even the most optimistic predictions. Consider the projections for AI in 2017—most futurists undershot its actual progress by an order of magnitude.

This research clearly illuminates a path forward:

https://ekinakyurek.github.io/papers/ttt.pdf

Deeply appreciate your thoughtful comment.