Yeah, but then responsibility for the consequences of action of the system falls onto person pressing the button. I hope that no one would be willing to take such responsibility without second thought.
What part of placing a human in the loop requires telling that human what they’re approving, or even that they are approving something? Pressing a button could be as simple as following orders—or giving orders.
Barely anyone here touched on your idea that tab completion is great use of AI. On paper I like the idea of tab completion - I'm closer to the code, I'm doing, not waiting, I have more control. However, I'm not programmer building some complicated application - I'm data scientist writing short fragments of code to analyze data, train models and serve them. So maybe I don't see a big picture of code development.
Having said that, I cannot internalize how tab completion fits into my workflow - what are the benefits compared to IntelliSense. So far I've used LLMs as glorified search engines or to better understand some concepts.
In order to get a feel for tab completion I've installed continue.dev VS Code extension and connected it to local model. I don't know if there are better free alternatives (I hope so, since continue.dev project was killed recently and will not be ... continued)
Funny, but I've recently started a blog to keep my math and physics muscles working and this exact problem is the first post I've written - it is useful in high order perturbation theory in quantum mechanics that I'm planning to describe down the line. Counting these paths is cool problem but generating all of them was more challenging.
Could you elaborate on Gemma 4 12B capabilities from your experience and benchmarks?