Slop.

Thank you very much for the explanations! I appreciate you!

I said less, not none. That particular AI will give the relevant page citation(s) for any stat or claim it makes so you can correct for errors, but obviously that doesn’t account for any data it just skipped so there’s still a possibility for errors of omission.

I’ll also note, as I think this is an under appreciated aspect of the AI narrative, is my wife is not some AI sycophant. She hates when people use AI to “write” an email. However, she is horrendously understaffed and gets way too much work dumped on her because she’s the only one that seems to have any institutional knowledge and gets things done without drama. So she’d love to go through these documents manually. But when you’ve got eight proposals dumped on your desk, each hundreds of pages of dense technical details, and are told “get them all evaluated by the end of the week,” there’s not much of an option there. When there’s more fallout from not getting the review done in time than there is from some bad data making it through the analysis of the AI summary, it’s not surprising that people are using it out of necessity.

LLMs are sort of like a bastard child outgrowth of machine learning systems. They aren't necessarily hot garbage -- e.g., DeepSeek is surprisingly good at scraping and summarizing research papers that have been fed into it -- but the general-purpose commercial shit like ChatGPT, Bard, Copilot, etc. are all more or less snake oil. Doing this stuff with manuals, technical whitepapers, and so forth is probably viable, but once they started feeding in social media posts -- even StackOverflow -- they lost the plot entirely. I say this as a senior-level software developer who constantly has to look shit up because I bounce between too many languages/platforms and nearly always need a refresher for whichever one I am currently trying to beat into submission, and I have been skimming the Google AI overview shit and checking its work more often than not lately. It doesn't always understand what I am looking for and tries to shit something out of left field anyway, so I just go straight to whatever Reddit/Stack/Baeldung/NerdRanch links pop up in the first few pages of results, but when it does "get" it, it's pretty close. If it's doing anything more than summarizing a Linux man page, I still click through to its source links because the highest-updooted answer isn't always the most correct for what I'm doing.

Regarding other actual uses, I think it was @microfiche@hexbear.net a few weeks back that had an anecdote about using one of the public LLMs to lay out some plumbing plans for a residential space given a set of codified rules, and the slop machine came out surprisingly close to the mark because it had so many guard rails around it (due to the building/plumbing code). I could see it working as a sanity check for tradespeople, civil engineers, and architects if they're running specialized models and the LLM is ultimately just a user interface. We're not really there in the US though. Shit's too unregulated and consumes entirely too much energy for what is still ultimately a novelty.

In computer science, starting in the 50s, there was a discipline called "artificial intelligence". Despite some interesting things coming out of that, by the 90s/2000s it was widely decried as being vaguely defined, over-promising and under-delivering. It was a good idea then to ditch the "artificial intelligence" label if you wanted funding. Better to do "machine learning", a more defined field (slash rebrand), focusing (mostly) on approximating functions using large amounts of sample data.

For example, one might feed it 100 000 scans of hand-written digits (from postal codes), and that would result in a function that could tell you whether some scan of a hand-written digit was a 1 or a 2 or any other digit. This was an early application, used to automatically sort letters for the postal service.

The "GenAI" version would be that in reverse, i.e. you tell it draw a 1, and it will generate something that looks like a hand-written 1.

"Deep leaning" just refers to the use of so-called "artificial neural networks", one of these "machine learning" algorithms, which has now taken center stage. The "deep" here refers to the many layers of the "neural net". The word "neural" here is misleading. There is a very superficial similarity between this and biological neurons, as they do differ in pretty much every imaginable way.

Neural nets are used in LLMs, but they are also widely used in e.g. image recognition. So one might call both of them "deep learning". If you wanted to sell a drone that autonomously dive-bombs a tank, that may very well use an "artificial neural net" to identify the tank. But it's not in any way "generative", so you can't call it "GenAI". You could call it "deep learning" or "AI" though, if you were a marketing person.