I was taking a quick spin through Mastodon today and saw a post from someone I don’t know talking about “LLM Brain” and how bad it is. I don’t know the person so I don’t want to quote them, but the gist of it was they were complaining about trying to teach someone who would take every error message they encountered and feed it to ChatGPT to get a solution. There were some extenuating circumstances; in this case the LLM was often wrong and I guess the student was very resistant to stopping to think about the problem, and I can definitely see where that could be an issue.
But holy smokes in about 2 months I’ve completely accepted having “LLM Brain” and no regrets! Maybe I don’t have imposter syndrome and I really am an imposter, but let me describe my old and new ways of working. The scenario for these work flows is that I’m trying to install some new open source project on my PC. Despite following the documentation, I encounter an error that makes no sense to me.
Old Way: Paste the error into a search box. Skip past ads and sponsored results and YouTube video suggestions to get to the real results, which invariably point me towards something like Stack Exchange or maybe Reddit. But rarely to a traditional article. Next, start following links and skimming the pages. Check for how out of date the proposed solutions are, discarding really old results. Discarding the posts that are people yelling at the questioner for using the wrong format or whatever. Eventually finding a solution, trying to suss out what it is going to do, and then trying to use the solution. Sometimes it works, sometimes not If not, start all over.
New Way: Open ChatGPT, explain what I’m trying to do and share the error message. Almost instantly get a response that both gives me a solution and generally explains WHY I hit this error in the first place. Then I suss out what the solution is going to do, and finally try it. If it’s something really spooky I’ll take a minute and get a second source. But generally the first solution works.
And, if I then get another error, ChatGPT still has context of what I’m doing, so I don’t have to start from scratch again.
I work through problems orders of magnitude more quickly when using an LLM than I used to do using search and user-generated pages. And I don’t think it is making me dumber. I mean I still try to understand why things are happening and how we’re going to fix the issue, with the bonus that the LLM is happy to dive into these details. Does it get things wrong? Yes, sometimes. I still have to sanity check and all that. But plenty of search results give wrong answers, too.
So I kind of reject this idea of “LLM Brain” being a bad thing. In a way it reminds me of how they used to say no one would be able to do basic math once cheap calculators became available. They WERE kind of right, but does it matter? We all walk around with calculators in our pockets. I guess after the apocalypse we’ll be screwed but… I also bet our collective handwriting has gotten REAL bad since the invention of personal computers, but that doesn’t mean you can’t learn to do calligraphy if that’s what you want to do. It just means you don’t HAVE to spend the time learning good penmanship if you don’t want to. We have choices that we never used to have.
Real world example. This morning I wanted to fiddle with an open source project that required me to have node.js on my PC. I downloaded a Node.js installer for Windows and it failed spectacularly. Now I had a mess. I turned to ChatGPT which first guided me through cleaning out all the cruft the botched installation left on my machine, including temp files in AppData and such that I never would’ve thought of on my own. Then it guided me through a manual installation of node.js that was actually easier and faster than using the “installer.”
Again, I just don’t see this as an issue; I see it as being more productive.
I KNOW at this point I’ve drunk deep of the AI Koolaid. But I drank deep of the calculator Koolaid, and the Personal Computer Koolaid, and the cell phone Koolaid, too. And I’ve turned out OK so far.
My only real hesitation is just the power requirements of all of this and how it is impacting the environment, but I also see how models are getting smaller and more efficient so I’m hoping that will level off over time. If the current administration was so pro fossil fuel and anti renewable energy I’d be less concerned about all this, but maybe in 3.5 years that will change.
[Header image created using Imagen at https://aistudio.google.com/gen-media, I believe. I didn’t keep notes on that one.]
I recommend reading the article I linked in my post today. It’s the fairest, best-reasoned account of what AI is and isn’t good for, currently, why, and how to make sure you’re getting the best from it. He doesn’t really touch on the creative side much, where I think the main issues now are ethical and legal rather than practical, but he does confirm that the kind of thing you’re using the software for is one of the things it’s best at and best suited to do.
I have to say that as far as using LLMs/AIs for general search, my experience continues to be pretty poor. That’s with Gemini, mostly, or at least I assume that’s what Google is using. Now that it includes links to its sources wothj the AI summaries, it’s common for me to find that, if I click through to those, they’re either not at all authoritative or just plain wrong. It’s not so much the infamous hallucinations as it is that the Ai doesn’t seem to have the means or the motivation to fact-check its ownm sources. It just finds them and rapackages them without much regard to their intrinsic value.
I find it really interesting that your search results aren’t good, because mine are excellent. As I’ve said before, that is probably due to the kinds of things we tend to search for. I’m usually searching for syntax for some code language or maybe how to solve a math problem or something. These kind of feel like very binary searches, right? There’s 1 correct answer and nothing really subjective in there.
My guess is that you’re searching for more general interest topics that have a lot of subjectivity or just stuff that a lot of folks have posted bad info about, or something along those line?
I did go and read that post and pretty much agree with most of what he said when talking from a business point of view.
I also use it as a search engine, mainly, for code information that I know I knew at one point, but forgot somewhere along the way. I have no notes, really, because I know — as any developer does — that the non LLM route is going to lead to Stack Overflow or Reddit, and then it’s really a matter of parsing the context for info on whether or not the proffered answer is relevant to my specific use case. In dealing with LLMs (my company is actually shelling out for “Pro-pilot”) I have had pretty good luck in what I’ve been served. We’ve also touched on this before, but often I’ll just examine what’s returned to me for information or refreshment, and then go about putting it “into my own words”.
Bottom line, whether it’s SO/Reddit or LLM, I’m asking _someone_ else for information, and if it’s all coming from the same or similar sources — sources designed for people asking people for this exact purpose and this exact info — I have no issues cutting out a middle-man and getting a distilled version of the answer.
“but often I’ll just examine what’s returned to me for information or refreshment, and then go about putting it “into my own words”.”
Yup I do this too.
Lately I’ve been learning ComfyUI which has been described to me as “Blender for AI Models” in that it can do a ton but the UI can be really obtuse, and I’ve been leaning on ChatGPT pretty heavily for “explainers” [eg
What’s the difference between a checkpoint model and a diffusion model”] and another benefit here is that ChatGPT retains context so knows I’m still talking about ComfyUI so I don’t have to declare that with every question.
I was deep into Comfy about a year and a half ago, and really enjoyed it. It sent me down the rabbit hole of how LLMs were created in an attempt to understand how to not just craft a better prompt, but to tweak the various node settings to get the best responses.
That’s where I am now, and I’m asking a lot of questions. Like what exactly is the difference between a checkpoint and a model. And I have to confess I AM taking the answers more or less on faith for now. But so far everything Monday has told me seems right based on all the YouTube vids and stuff I’ve also been watching.
I found the thread you were referring to, and I think the key difference in experience here is… well, *experience*. As in, you have it and this student in question did not.
‘LLM Brain’ in your context is more akin to discussing with a peer with unlimited time and patience; you can talk through the issues and work the problem. Even if the peer in this case is ‘giving the answers’, you stand a much better shot of understanding how it got there and the explanations being provided.
‘LLM Brain’ in the student context was… not that. It was more akin to copying someone else’s homework – but without really checking or even understanding if the homework being copied was for the same question, assignment, or even subject. There was utter disconnect between learner and ‘teacher’ (AI). I think the example to most clearly highlight that is where the student was calling the wrong function (by a matter of a single letter) and yet somehow could not identify this issue, either for themselves nor from the AI side.
Fascinating subject.
I’ve been using AI for code a bit more recently as well, in making a child theme for the blog and messing around with what CSS can do. But I feel like there might be a post here, and I’ve left enough in comment form here as to almost be one already. xD
I wonder if it’s not also about trust, as in some subsets of people (possibly younger or more inexperienced or with less expertise in the subject matter) putting too much trust in what the LLM tells them.
Here, there’s a human checking through the AI work and seeking a second source for verification.
In other cases, some are just taking the answers the LLM produces like the be-all and end-all and/or passing it off as their own work, without the willingness or the wherewithal to fact check it.
I don’t know if there’s a correlation between the above and a phenomenon I’ve observed lately in the Reddit circles I’m frequenting (and I wonder if this also happens on other social media). Certain people constantly crowd sourcing opinion and asking for permission. “Is it okay if I do this, or that?” Subtext being essentially, will I get attacked or canceled or socially ostracized in some manner if I don’t check in with other people online on what’s acceptable or not?
Most of them sound pretty young. Maybe it’s a generational/cultural shift, like that Verge article a few years back suggesting that a number of Gen Z students didn’t understand file folder systems, because the phones they’re using do it differently. Or this Reddit thread:
https://www.reddit.com/r/Professors/comments/1ftvyrb/do_many_students_nowadays_really_not_understand/
Maybe with the advent of the internet and Google and social media, a number of people have basically outsourced thinking. And trusting AI to give them answers is just another step down that path.
It did for sure sound like the “student” in this case was blindly accepting whatever the LLM told them. But I wonder if this student didn’t have an LLM if he’d be doing the exact same thing, putting his error into google and then just accepting the first answer from the first result and plugging that into his project.
And y’know I kind of feel like that was a talking point ‘back in the day’ about how having Google (or back then, Yahoo or Alta Vista or something) was making it so people didn’t know how to do research on their own. Y’know, heading to the library and digging through the card catalog and the stacks!
Stuff like your WordPress Child Theme project is so perfect. I’m sure the LLM quickly told you how you should format your style.css to ensure it was a proper child theme. And it’s also great for revealing WordPress commands that do something you want to do in a way much easier than you writing all the code for it. It was not TOO long ago that I learned wp_kses in this way. Prior to that I was writing PHP functions to strip out illegal HTML tags, and thanks to a LLM I learned “Oh, WordPress provides a function to do just that!”