Exponential Curves
2026-01-03
I'm awake, and so far my morning has been peaceful and routine. It always feels nice when things go the way we expect them to. I'm a bit upset though because I've still been dealing with more sleepiness and fatigue lately. Yesterday I took another long nap in the evening. I don't like being so tired when I'm in a decent mood because I still want to try and do things. I want to switch over to a local LLM setup so that I can cancel my ChatGPT subscription and have more control over my data, the model I use, and my privacy.
Also, I am over the honeymoon phase with LLMs. I've explored the whole garden, as it were. It's always nice to have an interesting conversation with an LLM or work through different ideas with it. There is a different flow to brainstorming with an LLM compared to collaborating with other humans. It has its advantages and disadvantages. I would say the principal advantage of brainstorming with an LLM is that I can fully dictate the pace. That level of control is nice.
However, a main disadvantage I've seen as the technology has progressed is that I've started to see a similar uncanny valley with AI prose as I have with images. Every writer has a distinct style, including LLMs. Many have talked about the "fifteen most glaring signs you're reading AI-generated text," and I think that we've all gotten sick of its prevalence and monotonous use of syntax and rhetorical devices.
The more that I use ChatGPT myself, the more my disdain for the writing style grows. This is also because with their latest models, the style has become even more pronounced and jarring. That makes sense, though, considering that most of the training data used to improve those models came from, well, previous ones. There's not enough human-made data for the companies to gobble up anymore, so they settle with inferior synthetic data and convince themselves it's actually better. Funny how that happens.
I've outlined my general opinions and uses for AI here and those words still carry their weight here. Enshittification is most certainly upon us, and I want to find my way out of it. ChatGPT has been a great tool for my writing. Most people use it as their personal freelancer and make it generate all of their writing. That's fine if your focuses are beyond just writing and things like performance metrics matter more than actual performance.
I find that ChatGPT works best for me as a copy editor. It finds errors that standard spell checkers miss, especially more complex ones like idiomatic errors. It also helps with clarity and trimming certain sentences down, which I am always a big fan of. Then, of course, it also helps with big picture ideas for certain pieces and makes sure that there's follow-through. The best part about this is that I can choose to agree with its editing choices and make the changes or disagree with them and keep them as is.
This is the main thing I use it for. I'm not a fan of how it still constantly glazes everyone and tries to make everyone think they're a genius, but I've learned to harden myself towards its sycophancy. I use it a lot less than I did a year or two ago, and this lowered usage is another major reason I want to switch to a local option. I don't need as much compute as I used to, and I think that I might even be able to lower my costs in using it from OpenAI's $20 per month price tag.
From what I understand, I can use programs like Ollama to host models and then rent a high-end virtual GPU that the model uses to compute answers and work either in the terminal or make some kind of basic web front-end. I've also seen KoboldAI's stuff too and that seems like a good choice too, even though it's geared more towards users who use LLMs as roleplay partners.
Ultimately, I just want to keep myself ahead of the curve.
It just sucks that it's an exponential one.