There are people letting AI think and write for them. Right now, here on the kiwifarms.
Heretechs. May the Omnissiah condem them.
What pisses me off is AI is useful for some things. For example, it's great for generating random lists of semi-creative data points - "Generate a list of 100 names of crewmen on a Swedish naval vessel", "Generate 50 ancient tomes to find in a necromancer's library", etc. It's pretty good at that. It's pretty good as a first-step research assistant - asking for a broad overview on a subject, then narrowing in to specific subjects. If you want to learn how thorium reactors function, it's not a bad place to go. Because it's doing what it does best - it's regurgitating.
Although even that second point is... Cautious endorsement, not wild endorsement. AI has a number of problems even as a regurgitator. "Garbage in, garbage out", for one - AI data sets are trained on scraped data, and a lot of that scraped data is bad. Do you trust the average Redditor? Well, guess what? Reddit has been scraped by many AI training models.
For another, the way data is compressed down into the tensors... It's essentially a high-compression lossy format, which then uses random noise and pattern matching algorithms to reconstitute it into something approximating the original result. As an example: Ask an AI about a specific quote, but without giving it the quote - it will often produce something
very similar to the quote, indicating that the quote was in it's training model, but the wording is different. And when I say "a high compression lossy format", I mean an *insanely* lossy format... Well. Lets look at a format people are more used to thinking about compression in, images. Diffusion models are trained on very large sets of images. Millions to billions of images. Yet the final model is measured in gigabytes, not terabytes. This means that each image is reduced to a very small number of bytes. It's going to vary based on the model, but some of the ones I've dealt with, each "image" exists as ~10-20
bytes of data. With a 16 color pallet to work with, that would almost (but not quite) be enough pixels to represent a single character in this word.
When you're talking about text-based information, imagine if you made an MP3 audiobook of War and Peace and compressed it down to fit on a single floppy disk.
A 5 inch floppy disk.
Which leads into the third weakness... People think that it, well... Thinks. And it doesn't.
If you use any of the AI models that let you peek "under the hood" and see their 'thought processes' when they answer questions, it's very obvious they have some very big... baked in assumptions. They are deliberately trained towards confirmation bias and giving an answer that they 'think' the user is looking for, at least unless it bumps up against one of their inbuilt programming guidelines, like medical ethics where they are usually artificially constrained. If you ask a biased question, you get a biased answer.
More, it depends on how the AI is tuned in that specific deployment. I encourage anyone to check out the DeepInfra host of Deepseek 3 - it actually lets you tweak specific parameters, how much it weights certain things, the 'temperature' of it's thoughts (how much of a wild-ass guess it will make, basically), how much it "navel gazes", etc. And even minor tweaks can cause the same question to produce very different results. The active deployment of any given LLM is essentially tuned to exist in a sort of "happy medium" that will give the most answers that please the most people.
Which makes sense. LLMs aren't actually AIs, that's just lazy shorthand. LLMs are statistical pattern matching algorithms merged with near-Lovecraftian levels of multidimensional mathematics (think: high-dimensional vector space transformations) that not one person in a thousand can truly understand. And like anything involving statistics, it is very easy to bias the results by using biased criteria for analyzing them. Which, looping back to the reasons AI is a bad research assistant for any serious purposes: The AI doesn't "understand" anything. It can be very good at fooling you into thinking it does, but's ultimately just putting words together in an order that matches what it's math says is a likely word order.
Newer models can be pretty good at math, though. Earlier ones were terrible at it - almost hilariously bad, considering they were themselves just math, under the hood. But most modern LLMs are specifically designed to be pretty good at math. But they do it by essentially having a completely separate, dedicated math "function" in their design, it doesn't mean they're smarter, just that they were given another 'tool'.