It’s unacceptable for a tool to sacrifice an iota of clarity, coherence, meaning, quality, etc. for the purpose of embedding hidden clues within the text to suggest its provenance. The idea that anything other than *my* needs should factor into the generation of text *for me* is patently offensive.
Primary reason I’m not using any hosted LLMs is I can’t stand and don’t want to start relying on tools that can be stealthily enshittified or made an advertising tool.
While open source (well more like “open weights”) cannot easily be enshittified they can absolutely be a stealthily made advertising tool.
For example : if you ask “how do I do X” it could recommand to you based on who paid the company making the LLM more money and you would be none the wiser.
For more obvious bias just ask Chinese models about Taiwan and other problematic subject to the Chinese gov and you’ll see. And the source code isn’t telling us anything about if the LLM weights encode some kind of propaganda, advertising or bias toward/against some entities.
The open models are good enough. We will at least have what is available today and thats enough for most use cases. Deepseek v4 is dirt cheap and its really good.
It doesnt seem like anyone is going to make a massive leap since theyre all training data and compute bottlenecked.