I've been feeling more and more that generative AI represents the average of all human knowledge. Which has its place. But a future in which all thought and creativity is averaged away is a bleak one. It's the heat death of thought.
Dostoevsky said that if all human knowledge could ever be reduced to 2 + 2 = 4, man would stick out his tongue and insist that 2 + 2 = 5. That was a 19th century formulation—he was a contemporary of Boole. I wonder what the equivalent would be for the LLM era.
That may or may not be true, but the expression of thought and creativity matters to transfer meaning. If you average that out, it loses momentum. Example: https://news.ycombinator.com/item?id=47346935. Compare the posters first and second, LLM assisted, paragraph. The second one is just bleak. If I had to read several pages like that, my eyes would glaze over. It cannot hold attention.
The why not is: human beings are valuable in and of themselves, not just because of what they can do. If you raise the bar too high, you kick people out. And our society just isn't setup for that, and is unlikely to ever be in our lifetimes.
And I'm talking about a radical shift in the concept of ownership, where shareholding is radically democratized. Basically every random Joe needs the option to live comfortably on passive income generated by things he owns.
I mean that it's a kind of lowest common denominator average where it's more important to seem reasonable and to not upset anyone rather than be really good in some ways and bad in others.
https://en.wikipedia.org/wiki/Mode_(statistics)
If human knowledge were a pyramid, LLMs just make the pyramid flatter, i.e. shorter, wider at the bottom, and narrower at the tip. It makes Humans dumber.
The capital M had meaning that I didnt grasp since I hadn't heard of Mode in that way before.
Today's learning!
The comment by Joseph Greenpie[0] is just marvellous, what a gem!
-----
No, it's far worse. It's the mode of all human knowledge. The amount of effort you have to put into an LLM to get it to choose an option that isn't the most salient example of anything that could fit as a response is monumental. They skip exact matches for most common matches; it's basically a continuity from when search engines stopped listening to your queries and just decided what query they wanted to respond to - and it suddenly became nearly impossible to search for people who had the same first name as anyone who was famous or in the news.
I've tried a dozen times to get LLMs to find authors for me, or papers, where I describe what I remember about them fairly exactly. They deliver me a bunch of bestsellers and popular things, over and over again, who don't even match at all large numbers of the criteria I've laid out.
It's why they're dumb and can't accomplish anything original. It's structural. They're inherently biased to deliver lowest common denominator work. If you're trying to deliver something original or unusual, what bubbles up is samplings of the slop that surrounds us every day. They're fed everything, meaning everything in proportion to its presence in the world. The vast majority of things are shit, or better said, repetitions of the same shit that isn't productive. The things that are most readily available are already tapped out. The things that are productive are obscure.
You can't even get LLMs to say some words by asking them to "say word X." They just will always find a word that will fill that slot "better." As I said, this is just google saying "did you mean Y?" But it's not asking anymore, it's telling.
edit: It's also why asking it to solve obscure math problems is a dumb test. If the math problem is obscure enough, and there's only one way to possibly solve it, and somebody did it once, somewhere, or referred to the possibility of solving it that way, once, somewhere, you're going to have a single salient example. It's not a greenfield, it's not a white sheet of paper: it's a green field with one yellow flower on it, or a piece of white paper with one black sentence on it, and you're asking it to find the flower or explain the sentence.
edit: https://news.ycombinator.com/item?id=47346901 - I'm late and long-winded.
At this point I'd rather review LLM generated code than a poor developer's.
It's literally what it is. Fairly sure that mathematically it's a fancier regression/prediction so it's a form of average.
Have you tried the paid versions of frontier models? They certainly do not feel like they spew the average of all human knowledge. It's not uncommon for them to find and interpret the cutting edge of papers in any of the domains that I've asked them questions about.
It is not uncommon for me to read a recently published review and find 2-3 interesting papers in the lot. Plus the daily Google scholar alerts. It can definitely be beneficial to have a LLM summarize a paper. Of course, at this point, one should definitely decide "is this worth reading more carefully?" and actually read at least some parts if needed.
Windows is the most widely used because of history and inertia. It is not the best option in just about any metric, and a lot of this is objective. It's slow, the software is poorly designed and obtuse to use, and it lacks functionality. However, people already kind of know it, so it remains.
If Microsoft didn't get the IBM gig early on, we would not be using Windows.
Start going to the profiles of every comment from a green account you see for a week and you’ll see how bad it is.
There will be friendly fire but unfortunately that’s to be expected when you click the top comment in a thread and realize an account has been posting 100% slop for months.
I won't comment further.
Why not?
Personally, I don't have the specialized knowledge, nor the time needed, to read and understand papers outside my own 2-3 domains. LLMs do. And I appreciate what they can do for me. They do it better, faster, and more accurately than most 'popular science', provide better coverage and also provide the ability to interact with the material to any degree or depth that I care to, better than any article.
It would be silly to pass up this capability to make my life better simply because random folks on the Internet disparage the quality of the output (contrary to my own experience) and make hand-wavy points about 'someone else's computer) while offering no credible or useful alternative :)
Tough question. I think the straightforward answer is that you can't.
That said, there is some confidence gained in an LLM's abilities based on its performance on papers in domains that I do understand. Yes, it's not going to be the same across all domains, but the frontier labs do publish capability scores across different domains, and that helps scrutinize the answers it provides, and how much salt to take with those.
It could be that the LLMs are good at stringing words together in a way that seems reasonable when you are not an expert yourself, much like people from other fields seem very knowledgeable until you compare many of them or hear/see them talk with each other.
I have, and it does, hence my confidence in its ability to do the same in other domains. Depending on what you're using it for, it is advisable to maintain some level of quality control (spot checks, sampling, deep dives, more rigorous continuous review) as in any process control.