LLMs are plateauing. At this point, anyone who has cared enough knows their fundamental limitations. Don’t get me wrong, they do provide immense value and have grown quickly in just a few years. But, they will probably never get past the “we must treat it like a junior engineer” phase, especially as LLM outputs inevitably leak back into the training data since everyone is using it (or being forced to by employers). Notice that everything major from these companies lately hasn’t been a vastly improved model from the previous iteration, but products and tooling: see Anthropic allowing you to create in-Claude apps, every companies own version of agentic CLI, and so on.