Comments by "Charles M." (@charlesm.2604) on "Claude 3.7 goes hard for programmers…" video.

  1.  @Pz916  You are out of your mind if you think a 24GB vram GPU can hold a state-of-the-art AI model. A 3090 can only run 12B parameters models and still have terrible throughout speeds because of how high the context size has to be for programming use cases like these (900GB/s of memory bandwidth for 200k tokens?). I have a double 3090 setup with a nvlink (48GB) to host 70B models at home (mostly qwen models for me) and I guarantee you that they don't replace writing code by hand. They are helpful to generate tests, code blocks and give an expressive description of new code when I pull the repository for changes, but that's about it. I also use AI to write utilities function (string manipulation, date utils, etc.) or RegExp/SQL queries with thorough human review (most of the time it spews out incomplete queries or doesn't make use of indexes and other performance related features). I mean, just look at this video. A state-of-the-art AI, the best on the market, doesn't even type properly, which is the #1 root step for any automation at later stages (testing, docs, CI, etc.). If Claude can't type then a consumer model won't do wonders and write you a functioning app. The best AI models, those that "could" replace human workers (they don't exist yet), will never be cheaper than paying a salary. Just the electricity to power datacenters is more expensive than hiring a full-time team. Also, salaries can be adjusted.
    1