This reinforcement learning with human feedback.

Yes, it's a challenging task to handle in large volumes. You need a relatively low cost with high-level experts and a very good throughput, which is not easy to achieve. They managed to do this, and it's what they've been earning their money from. Even within this market, you can find a niche that brings in a lot of revenue. If RLHF becomes obsolete, they will have to adapt. For instance, if OpenAI realizes there's a limit to how much they can improve the model with RLHF, I'm nearly 100% sure there is unless you change the whole architecture. Unless it's not just transformers, and you come up with something new, we're probably hitting the limit of what this model can do. There may be another architecture that might or might not use RLHF-type training. If it's not RLHF, Surge will have to pivot.