Llms Are Steroids for Your Dunning-Kruger

I'm not sure I follow. LLMs are probabilistic next-token prediction based on current context, that is a factual, foundational statement about the technology that runs all LLMs today.

We can ascribe other things to that, such as reasoning or knowledge or agency, but that doesn't change how they work. Their fundamental architecture is well understood, even if we allow for the idea that maybe there are some emergent behaviors that we haven't described completely.

> It's a possible outcome of the process, but it's not the only possible outcome.

Again, you can ascribe these other things to it, but to say that these external descriptions of outputs call into question the architecture that runs these LLMs is a strange thing to say.

> Looking at a LLMs I/O structure and training process is not enough to conclude much of anything. And that's the misconception.

I don't see how that's a misconception. We evaluate all pretty much everything by inputs and outputs. And we use those to infer internal state. Because that's all we're capable of in the real world.

I am very confused by your stance.

The aim of the function approximation is to maximize the likelihood of the observed data (this is standard statistical modelling), using machine learning (e.g., stochastic gradient decent) on a class of universal function approximators is a standard approach to fitting such a model.

What do you think statistical modelling involves?

Calculation isn't what makes us special; that's down to things like consciousness, self-awareness and volition.

> The main reason to say "LLMs are just next token predictions" is to stop thinking about all the inconvenient things. Things like...

They do it by iteratively predicting the next token.

Suppose the calculations to do a more detailed analysis were tractable. Why should we expect the result to be any more insightful? It would not make the computer conscious, self-aware or motivated. For the same reason that conventional programs do not.

I'm an SWE working in AI-related development so I have a probably higher baseline of understanding than most, but even I end up awed sometimes. For example, I was playing a video game the other night that had an annoying box sliding puzzle in it (you know, where you've got to move a piece to specific area but it's blocked by other pieces that you need to move in some order first). I struggled with it for way too long (because I missed a crucial detail), so for shits and giggles I decided to let ChatGPT have a go at it.

I took a photo of the initial game board on my tv and fed it into the high thinking version with a bit of text describing the desired outcome. ChatGPT was able to process the image and my text and after a few turns generated python code to solve it. It didn't come up with the solution, but that's because of the detail I missed that fundamentally changed the rules.

Anyway, I've been in the tech industry long enough that I have a pretty good idea of what should and shouldn't be possible with programs. It's absolutely wild to me that I was able to use a photo of a game board and like three sentences of text and end up with an accurate conclusion (that it was unsolvable based on the provided rules). There's so much more potential with these things than many people realize.

None of this is surprising? Like, I think you just lack a good statistical intuition. The amazing thing is that we have these extremely capable models, and methods to learn them. That process is an active area of research (as is much of statistics), but it is just all statistics...

Using Perplexity or Claude in "please source your answer" mode is much more like a conventional search engine than looking up data embedded in 5 trillion (or whatever) parameters.

Today they might have better trust for Wikipedia-- and I know I use it as a source of truth for a lot of things-- but back in my day teachers were of the opinion that it couldn't be trusted. This was for like middle and high school, not college or university, so we would cite encyclopedias and that sort of thing, since we weren't reading cutting edge papers back then (maybe today kids read them, who knows).

Edit: Also, I think the GP comment was proven correct by all of the replies claiming that Wikipedia was never controversial because it was very clear to everyone my age when Wikipedia was created/founded that teachers didn't trust the internet nor Wikipedia at the time.

An encyclopedia does not cite its sources, and does not claim to be a primary source, its potentialymistakes cannot be checked.

(Wikipedia has the additional problem that, by default, the version cited is the ever-changing "latest" version, not a fixed and identified version.)

Either way, you are correct, we weren't allowed to cite any encyclopedia, but they were meant to be jumping off points for papers. After Wikipedia launched when I was in 9th grade, we weren't allowed to even look at it (blocked from school computers).

I definitely used it though.

1) encyclopedias are a tertiary source. They cite information collected by others. (Primary source: the actual account/document etc, Secondary source: books or articles about the subject, Tertiary source: Summaries of secondary sources.)

2) The purpose of writing a research paper was.. doing research and looking up an entry in an encylopedia is a very superficial form of research.

Also the overall quality of Wikipedia articles has improved over the years. I remember when it was much more like HHG with random goofy stuff in articles, poor citations, etc. Comparing it to, for instance, Encarta was often fun.

Of course that's demonstrative of the genesis fallacy. Anyone can write or publish a book, too. So it always comes down to "how can you trust information?" That's where individual responsibility to think critically comes in. There's not really anything you can do about the fact that a lot of people will choose to not think.

I’m reminded of when people at work assured me the internet was never going to impact media consumption because 28.8kbps is not nearly enough for video.

And, for instance, I have barely any knowledge of how my computer works, but it's a tool I use to do my job. (and to have fun at home.)

Why are these different than using LLMs? I think at least for me the distinction is whether or not something enables me to perform a task, or whether it's just doing the task for me. If I had to write my own OS and word processor just to write a letter, it'd never happen. The fact that the computer does this for me facilitates my task. I could write the letter by hand, but doing it in a word processor is way better. Especially if I want to print multiple copies of the letter.

But for LLMs, my task might be something like "setting up apache is easy, but I've never done it so just tell me how do it so I don't fumble through learning and make it take way longer." The task was setting up Apache. The task was assigned to me, but I didn't really do it. There wasn't necessarily some higher level task that I merely needed Apache for. Apache was the whole task! And I didn't do it!

Now, this will not be the case for all LLM-enabled tasks, but I think this distinction speaks to my experience. In the previous word processor example, the LLM would just write my document for me. It doesn't allow me to write my document more efficiently. It's efficient, only in the sense that I no longer need to actually do it myself, except for maybe to act as an editor. (and most people don't even do much of that work) My skill in writing either atrophies or never fully develops since I don't actually need to spend any time doing it or thinking about it.

In a perfect world, I use self-discipline to have the LLM show me how to set up Apache, then take notes, and then research, and then set it up manually in subsequent runs; I'd have benefited from learning the task much more quickly than if I'd done it alone, but also used my self-discipline to make sure I actually really learned something and developed expertise as well. My argument is that most people will not succeed in doing this, and will just let the LLM think for them.

With the caveat of for our field in particular, it's one of the few that require continuous learning and adaptation, so tech workers in a way are better predisposed to this line of thinking and tool adoption without some of the potential harmful side effects.

To pick on spell check, it has been showing that we can develop a dependency on it and thereby losing our own ability to spell and reason about language. But, is that a bad thing? I don't know.

What I do know is humans have been outsourcing our thinking for a long time. LLMs are another evolution in that process, just another way to push off cognitive load onto a tool like we've done with stone tablets, books, paper notes, digital notes, google, etc.

As a software developer, your job is to understand code and business constraints so you can solve problems the way most appropriate for the situation. If you aren't actually keeping up with those constraints as they change through time, you're not doing your job. And yeah, that's a kind of fraud. Maybe it's more on yourself than your employer most of the time, but... It's your job. If you don't want to do it, maybe it's more respectful of your own time, energy, and humanity to move on.

especially large, old, convoluted domains where you want to be able to quickly map where things are, it's indeed a massive time saver

And worse of all is that it's not even consistent, because I tried the same searches again and I couldn't get the same answer, so it just randomly decides to assert complete nonsense sometimes while other times it gives the right answer or says something completely unrelated.

It's really been a major negative in my search experience. Every time I search for something I can't be sure that it's actually quoting anything verbatim, so I need to check the sources anyway. Except it's much harder to find the link to the source with these AI's than it is to just browse the verbatim snippets in a simple list of search results. So it's just occupying space with something that is simply less convenient.

Kidding aside, it's a good rule of thumb to distrust someone who appears very confident, even with people. Especially if they can't explain their reasoning. There are so many experts who are going to confidently tell you how it is based on their three decade old outdated knowledge that's of very questionable accuracy.

The fact that LLMs seem like people but aren't, specifically have a lot of the signals of a reliable source in some ways, I'm not sure how these processes will map. I'm skeptical of anyone who is confident about it in either way, in fact.