Yeah this does make me wonder: even now, models need to be constantly trained with new programming languages, libraries,etc. They can ingest new documentation during a session, but they often refuse to believe it. Ex: “the project uses .Net 10, but thats clearly a typo- the current version of .Net is 8”. So then if the bubble collapses and training budget drops, the models capability would suffer drastically

I think thats fairly reasonable. I have a personal project im working on- an sdl blend of DOS mode13h and the SNES retro engine. Ive used local models for specific issues- converting loaded images to closest palette matches, but for the most part its hand written. Im fairly happy with that middle ground. And im professionally frustrated by vibe coders that dont even check their output(not even code, they dont even check that the changed program works at all before submitting it).

They said 50- i looked up strata, and it uses smaller quantizations/optimizations of that model depending on your system.

I have 48g vram and 192g system ram and get ~15t/s on qwen 3.8 flash next at Q4. Off the top of my head. I think Q6 is 12.

Im fine with local only, but theres so many people that just vomit out exactly what claude says.