Author here. The search algorithm was the easy part. The LLM already encodes domain knowledge from ML papers; it knows learning rate warmup helps with transformers, that batch size and learning rate are coupled. It converged on the winning GRPO config by iteration 1. Grid search needed 8 iterations.
The hard part was per-iteration GPU isolation. A botched run that leaves stale optimizer state or corrupted weights in memory will poison the next iteration. Each iteration needs a fresh CUDA runtime, fresh filesystem, fresh memory. No state leaks. That's where most of the engineering went; ephemeral containers with TTL-based cleanup, one A100 per iteration, torn down after metrics are emitted.
I will only partially agree with you; mainly because indeed, that's what they're pitching crypto as - To be your own bank. But practically, most blockchains are replacing the inter-banking networks (like SWIFT) rather than the banks themselves, as you still need centralised entrypoints.
To save some comments; Yes, they do have the potentials to create their own closed-circle economic ecosystems, but for the retail banking services to work (eg overdrafts, mortgages etc) you will end-up with centralised players as these kind of services need to be backed by wealth (it's not a tech-issue).