

LLM driven code generation already works by gluing deterministic tools (linter, test runner, etc.) together. The LLM part is non-deterministic but the tools are deterministic. In other, limited and controlled environments this model can be extended to real physical systems. I would be surprised if automated warehouses couldn’t already be controlled via English instructions.
Claude and ChatGPT are stupid because they are general purpose. They fill the empty space between their training points with bullshit. A constrained domain specific AI, even GenAI has way smaller dimensional space to fill, therefore less gaps to bridge with bullshit. It is a viable and useful technology if it is not used the way it is used by oligarchs. In oligarch hands it is a tool of deception and oppression. In productive hands it could make it possible for people to not have to toil for 40+ hours a week to survive.



The only renaming that needs to happen is from White House to Shite House.