two years ago i wouldn't have been able to tell you these are the parts for a cycloidal gearbox. i've come so far
it's kind of crazy and sad how X has successfully removed all politics from my fyp but the negativity that usually comes around elections is still here in a non-political format
like building robots. every time a new model drop i get 10x more excited
the future looks really bleak when you can't build
i'm feeling this presently as i cough out my left lung
the context IS the thought process
I cannot definitively prove this, but if you demand professionalism in its communication, it appears that its thinking process itself altered. The code it writes has a different quality of modularity/composability. We have a usgc-handbook plugin for internal use, I cannot describe the utter pleasure of working with Claude that speaks proper english and responds in a console style that's professional in true sense of the term.
I see a bunch of memes where the midwit gives a massive md file of instructions and the grug/elite just say "Fix it". I am not sure if this is a good idea. "Write a to-do list app" -> it's output isn't objective and falsifiable unless you are also accepting a wide range of outputs.
From talking to Claude why this may be true:
> Attention heads activate on register features (terseness, formality, vocabulary) and those features are correlated in pretraining with authorship traits — senior engineers, documentation, well-maintained repos. Conditioning on the register moves the model toward that region of the data manifold, and code quality is a property of that region.
I am not an expert at LLM in the slightest, but what I can say is that my observation aligns with it. YMMV.