LLMs are making me lose my savviness(pgaleone.eu) |
LLMs are making me lose my savviness(pgaleone.eu) |
Of course there are errors and gaffes and weird bits all over, but that helps keep me in the loop and understanding it better.
I can now (more) boldly go where I feared to tread before.
To be clear, I am closing in on 60 years old. I’ve written a lot of code in a lot of languages. I have not bothered to check every fad that has come down in front end web and Javascript for the past umpteen years. Front end hasn’t been my thing. With current AI, I can leap frog a lot of bull crap I don’t care about and see what at least the AI thinks is good. I still retain a brain, I can still investigate what it did and form an opinion about it.
I treat the AI stuff as idiot savants that can make really nice stuff or go totally off in the weeds. But it is more than good enough to help me be a better me at what I do.
Writing things down and then arguing with an agent that constantly gets things wrong is just a new kind of Rubber Duck.
Are Rubber Ducks offloading thinking?
These are more like search plus memory tools.
Memory loss is real but that has been happening for me way before agents.
Uh, have you asked it?
These sorts of articles are funny. Like the author assumed that you have to use an LLM in a certain way.
“I can just ask the LLM and there it is! I didn’t learn anything! :(“
You can interrogate the LLM. You can ask it all sorts of things. Tell it to justify its decisions. Tell it to give you supporting evidence. Tell it to give you options. Tell it to explain things in detail.
I’ve been using it as a learning tool and it’s quite effective.
Granted the stuff I don’t care about really does get worse. If I have a config typo I just go “There’s a typo in blah. Find it and fix it.” Or setting up the boiler plate for a project. I’m getting quite useless at that.
But it doesn’t have to be any particular way. It’s a tool. Exploit it.
But if I end up feeding it JIRA tickets, playing agent management, and being QA on the other side, that's really not the job I signed up for.
It helps me pickup and understand new things pretty fast.
And when I need to learn something new I hate doing courses because they're always too slowly paced for me. All courses are paced for the slowest possible participant. Especially those from the US where trainers have this "training voice" where they speak slowly and articulated and usually non-skippable which makes me scream in agony. My brain can't slow down to that level, I just lose focus then. It's like driving a ferrari at 5km/hr. Especially the non-IT courses are so slow and constantly repeat everything, as if I didn't fully understand them the first time, argh. So I tend to spend my time on these awfully designed mandatory HR courses trying to fuck around in developer mode and re-enable the skip segment features :) The more boring the training the more I learn about moodle and its frameworks.
I just love to dive right in and just do something but I often skip that first analysis of what method would be best. And I tend to stick with what I know. Which is often not the best way in a greenfield project. And often the boring first steps are so awful for me that I just put it off. If I do start I often do it half-assed. Then I run into a wall, have to regroup etc. I learn from that (mistakes are always a learing experience), but it's slow going because I take the wrong direction first.
AI really helps me get past that first step meaning I actually start it instead of procrastinating the research forever and move onto the next thing. And I can do great things because I know what I want and how I want it. It's a real enabler for me especially something like cowork. It helps me widen my horizon efficiently instead of wasting my time and energy like online courses do. It picks up things that my weak points are (patience) and feeds me information in a highly dense format I can easily digest. Helped by custom prompts of course ("don't tell me I make great decisions, evaluate everything on its merits, just tell me the meat of the content, don't elaborate"). I still work along with it and take ownership of its output, I'm way too much of a control freak to let the AI do everything without understanding it. But it gets me there quicker and it avoids procrastination by skipping the boring steps. I ask it to suggest methods and techniques, compare their pros and cons so I fully understand what I'm selecting for.
I have to say this was not really the case with early simple question/answer LLMs. The hallucinations were too bad, the searches too limited. The knowledge too hit and miss. It would come up with wrong stuff or overlook the best options. But the new high-context agentic technologies like cowork that really investigate the source material for the purposes I provide, they really are a game changer for me.
For example, I recently asked it about an obscure error I'd been having. Previous simple LLMs would just collapse into a directionless troubleshooting mode, "Did you try...?". Cowork actually went and retrieved all the relevant code from github without being asked to do so, and evaluated what would have happened in the background for that message to appear. Really good to see because it's exactly what I would have done. But I would have taken much more time to do it.
So no, for me it's not making me lose savviness. It's helping me grow it in exactly the way that's efficient for me.
With AI, I can describe my exact use case and see what it gives me. Sometimes it will just cough up a furball like a boilerplate example, but usually it gets pretty close, and I can see how my (maybe weird) edge case fits into the framework or API or whatever.
It may be that I approach problems in a weird way, and that is why examples are less helpful than with some, but AI generally does better here for me.
Just curious. No right or wrong answer possible.
However, when we look at the claim with a more critical eye, it becomes obvious that this view is myopic. In fact, current progress is in fact indebted to the invention various tools, which has led us to grand theories and hypotheses: things the Ancient Greeks could only dream about.
Is coding about writing assembly? Is it about writing boilerplate? About writing endless if-then statements? I would argue that it is not: it's about coming up with ideas, building products, finding customers, iterating, producing code that's easy to read, easy to understand, easy to extend.
Coding (at least the kind LLMs do) was never really the hard part, and I'd like to know specific examples of savvyness the author is losing. They've made me more critical of code and able to create better patterns, simply because I can iterate so quickly.
I think you're blaming tools for what are other problems.
No carpenter ever said "I used to plane wood by hand and carve magnificent pieces, now that I have an orbital sander and power tools, the love just isn't there".
I don't buy this argument at all.
I also had a lot of ideas that I wanted to execute and I'm finally able to do so. Some of them are wild and big, unthinkable last year. I also really enjoy building product/software that people are delighted to use. So from my perspective, the net positive outweighs the losses.
My work position is less certain, I grief the years spent mastering things just to see AI doing it, but I went to software engineering because I enjoyed building products, from ideas to something people can play with, and this has never been easier.
Now I delegate all the grinding to AI. I still feel accomplishment when I get the final result. But more importantly, I can now build something in couple of weeks that would otherwise take me half a year. I can focus now less on the for-loop vs while-loop and more on the big picture, like keeping things tidy and separation of concerns. The front-end and how end-users will interact with what I build is now more important to me, as well as the design. Previously, I had brain capacity only for writing my clever backend algorithm. Now I can make an equally clever front-end for it.
Maybe we never enjoyed the grind? We just came to accept it and convinced ourselves to enjoy it, because there was no other way. If you wanted something cool, it meant you had to put in hours. Now you just need to word it correctly and get the prerequisites.
What's nevertheless true is that they've hardly touched hardware hacking. There are so many interesting opportunities right now in robots, drones, materials, batteries, and custom embedded systems that still demand dirty hands and physical intuition. These still reward savvy, and they're fun!
This is it for me. It’s a pretty incredible tool but it killed the vibe and challenge of building things.
preferred: LLMs are making me lose my savvy
Today I had claude figure out that using SIMD would not improve performance, and it gave me an analysis of why not. So my savvy improved without having to spend days figuring that out.
Having said that, I get that using AI can also make things less fun and reduce your learning too. I guess it depends on what you use it for.
For example, I wanted to build my own note-taking app for a long time because I got sick of Obsidian. Now I built my own, and it is magnitudes better than Obsidian, at least to me and for my use. I built smart home systems with Claude because, without an LLM, it would be too time-consuming to troubleshoot APIs and read all the api documentation. I wanted to test a TTS model, and it took me 10 minutes waiting to test it without reading the documentation. I use it for everyday random creations all day long.
Instead of one tiny project consuming all my week, I can now work on all of them in hours. My /srv folder has over 100 projects of my own, and some of them are parent folders to smaller dependency projects.
This is something I realized when I became a manager of programmers rather than doing the work myself. When I write code I do it from a position of having a deep understanding of what all the moving pieces are, when someone working for me writes code I just have to trust they had that level of understanding[1].
But being a manger, the thrill of solving the puzzle was replaced by the thrill of helping some new engineer get better at the craft of writing code. They took pride in their growing skill set and I took pride in helping them grow.
Managing consultants however was very transactional. Did they deliver what was asked of them? Did they fix the flaws I pointed out? Do I still want to pay them?[2] And I could of course fire them easily and try new ones that had a reputation for being "better" at the tasks I needed done.
If history is any judge, there have been many companies over the years where some founder/leader has a "great idea" and just needed someone technical to make it real. And the typical arc for those companies was they get founded, they spend a ton of money on consultants to build their technology, something changes (fashion, systems, something) and their product is no longer a good fit for the market so they try to hire consultants to fix it, which now costs more than 2x because first the consultants had to understand what the previous consultants did, then they have to change it in a way that leadership wants. The result is a product that got more and more brittle until there wasn't enough money to or enough consultants to anneal the product back into something people would pay enough for to support the salaries of the visionary leaders who never had any idea at all of how the product they sold actually worked.
It feels like we're going to speed run that. In the LLM era.
[1] And of course that level of trust is not always warranted but that's a different comment :-)
[2] With consultants there is a velocity vs organizational expertise equation that always needs to be balanced. Dropping people into a legacy code base they've never seen is just as hard for them as dropping them into a consultant's work product that they hadn't previously seen.
Moreover, I want to understand them as I did for any other tools in the past, that's why I find myself using and exploring them also outside work. It's just a tool, but with a huge impact of the crafting process (that is the fun part of coding for me).
This is not a fair analogy. Those things are like compilers, languages, and type-checkers. They are deterministic, carefully proven, hand-crafted tools, that allow people to go safely to a higher level of abstraction and still make things in a deterministic way. The jump into the fuzzy world of LLMs is a completely different thing. It's different than the jump from assembly to a higher-level language.
So this idea that doing math is some kind of fully deterministic & mechanistic process is a bit misleading.
Mathematics isn’t about calculating numbers but formulas and proves.
I have to put it politely: until the advent of AI no machine ever “calculated” a formula for me, because I didn’t have to calculate anything.
It is like driving a Tesla and laughing about the manufacturer when he mentions how hard it is to get the product right, because you have an easy ride.
I can relate to the article’s AI depression sometimes.
Whatever comes to your mind, AI makes it happen in a couple of minutes in most cases no matter what you imagined.
You invest days and the AI just build it is a couple of minutes way better than you and your 10 teams together did before.
So much hard work and sacrifice - and now? Poof - there it is.
It is nice but also depressing. And not being challenged is the actual problem.
The abacus and calculator doesn't do high order thinking or prove theorems for you. Math as a field aren't about manually adding numbers together or doing simple algebra formulas, and wheren't that even 2000 years ago.
Those calculators are to math what something like an IDE or a helper library is to programming.
The analogy would be with theorem provers, and even better, with the exact same tool as in the programming case: LLM doing the proofs.
If mathematicians had LLMs and theorem provers do their job, they'd certainly have gotten "less savvy". And if they let them write their coursework, starting from university, they'd might never even become savvy enough in the first place.
Builders/makers can be roughly subdivided into two camps: some like building for the final products; others like building for the process itself.
I'm in the latter camp. OP probably too. You probably are in the former.
And that's fine. But let's just stop pretending the other side doesn't exist.
EDIT: And this goes for OP as well. They forgot a lot of "for me"s in the article.
I would guess probably lots did say that?
At the very least I have no idea why you would be able to claim otherwise, we are far removed in time from when there were carpenters straddling the introduction of power tools.
Despite now having all these tools, a carpenter still needs a lot of skill to create something. If you hand me all the modern tools in existence, I'll probably still cut off both my thumbs and end up with a chair which breaks the moment I try to sit on it.
LLMs allow any person with no skills whatsoever to create almost anything they want. Quality is often good enough or doesn't matter at all. There's no skill needed because Claude will happily tell the person what they are missing and do it for them in the same step.
I totally feel what's expressed in the article: I've been in web development for 25 years now and I have never been this depressed in my life, because AI took most of what I cared for in my job and left me staring into a bleak future of becoming unemployed or pivoting into something I'm gonna hate doing for the rest of my life.
So more of a home-gamer chainsaw? Except that I'm not positive the once/twice per year chainsaw is nearly so effective.
and skills from building with hand tools help you judge the results, but not necessarily with actually using the tools.
For a woodworker, I think your analogy should be a “furniture factory,” not power tools. (Power tools in software would be like a great IDE and powerful language and framework tooling; AI is… way more than that.) I bet a “woodworker” in a furniture factory (if that even exists) doesn’t much care about their output, and doesn’t love the process anymore.
And getting paid for it doesn’t make you care about your work either, it just makes you care about money you can use on other things you do care about.
It would be foolish for anyone looking at the great change happening in software, and ignore the fact that we’re crossing a line where it’s less about craft, and where developers are just managing factory-like processes. Which is very different from what most signed up for.
And of course you can still do your craft, just like there are custom woodworkers as well as factories. But the space for that in AI-driven companies is vanishing as much as we might hate it or push back against it.
And of course there are also people like you who like the opportunity that more output provides, or who use it more like a tool and less like a factory. And that’s fine.
But no, the transition happening in the industry is not deliberately trying to keep the craft of it or even to keep developers interested in the job.
The transition is turning people who input stuff into a software factory and do QA on the output, or turning people into “factory automation engineers.” Those are both very different jobs from what people who like programming signed up for.
And yeah maybe no one “deserves” to have craft in their jobs. But let’s be honest, it is very natural, and very human, to be upset when the main motivation and reason for doing your work disappears. The same thing has happened in countless industries, and people have been talking about it for decades if not hundreds of years.
E.g. small-time farming replaced by exploitative big agriculture. A farmer might have a reason and identify in the work before, owning their own bit of land they tender. And now if they’re just hired to drive a tractor, they don’t. It’s a massive difference, and no wonder our society is increasingly depressed.
I guess what you’re saying is you like reading fully fledged code, almost like an apprentice. I would never think to delegate that responsibility to a next word guesser.
Depends on your usage. A rubber duck won't upheave the entire history of written human knowledge to reach for your answer. It doesn't have inherent bias, and its purpose could also be handled by a wall, or the air infront of you if you can stretch your mind a little. What I'm going for here is that as you said, its just a hack to make your brain look for answers in a more efficient manner, by clearly underlining your thought process out loud, tricking yourself into conversation mode, yada yada. Its still you doing all the work, you're just being smarter about it.
A LLM is more akin to a senior pair programmer. It may seem like a rubber duck on the surface, in the sense that you're having a conversation, but the senior programmer may come up with ideas you wouldn't think of. They may have more innate experience in the domain which might pull you away to very different implementations from what you had in mind initially. This is all good, but a common pitfall with LLMs is that they're much more conformative than a senior engineer, much faster, much more responsive. This is where it gets dangerous: they don't tell you you're wrong 80% of the time, like a rubebr duck. They also have innate experience of a domain, like a senior engineer. The result is that your brain never gets a 'hang' signal, and it doesn't feel the need to reevaluate.
I don't know, it seems more like every question goes to a contract agency, where a random programmer answers it. Generally they are very good at bringing up relevant things you've never heard of, but depending on who gets assigned to that answer, the implementation might be copy-pastable into production, or it might be a questionable hack, or it might be workable but clearly not the right approach if you're looking for those things. The agency must work by the hour, though, because if you challenge the assertions of the previous guy, the next guy will obsequiously say that you are, of course, quite right and possibly insightful and spit out a new solution more along the lines of the new guidance.
Just that an actual rubber duck doesn’t do anything. You solve the problem you have by talking, and in doing so, thinking, to come up with a solution, an idea, or gain better understanding.
After that you either implement something yourself or have learned something.
With an LLM you offload all of that, the only thing you still do is tell it what the problem is. The agentic duck does the rest and you look at the output.
Even if you have to argue, you argue without having gone through the steps to gain anything yourself.
Hopefully you realize that it's not actually introspecting into its previous thought process, and instead it is just making up a plausible after-the-fact explanation that matches the format of similar explanations found in its training set?
I too often find these explanations enlightening, but please don't forget that you need to read all such responses as those of a pathologically lying psychopath desparate to trick you, not as a trusted expert who actually understands its internals.
If you start blindly accepting everything it says, good luck to you.
But you can do things like get one agent to do something. Spin up another agent to review and critique. Use different LLMs to critique. I’ve had the same agent write piece of code and then come back and say “there are bugs x, y and z”.
For most code, it’s going to do a fairly good job and tell you best practices, give you things to research by yourself.
I've been working software development for close to 40 years now. I've worked in plenty of languages, software stacks, ecosystems, from assembler to VAXen to Unix (pre-dating Linux), to cloud stacks to...well, a lot of systems.
The software ecosystem has ballooned so tremendously in that time frame that no human can really be an expert in all of them, or even in a tiny subset of all of them. As I mentioned in another comment, I couldn't be bothered to keep up with the Javascript/Typescript/Web front end drama. I don't do rollercoasters, and that ride is enough to make even the staunchest stomach heave with distress from the endless loop-de-loops.
With LLMs, I can describe a little problem unique to my problem space, inspect the output, research it and even reverse engineer it, and see how well it fits (or doesn't). This is not unlike being a software engineering manager leading teams.
When I said being able to see the results was "cool", I did not mean or imply it was about me trying to "look cool". In fact, I really don't know where that is coming from. I do this for my day job and also on the side for a lot of different causes and personal projects, and no one could give a rats ass about how "cool" I am. Seriously.
People need to stop taking this all so personally. It's a tool. Maybe you all need to step back to my world in 1981 when I first started playing with code as a kid, and compare it to 2016. The advances from '81 to '16 are probably far more dramatic than what AI is doing now.
As for 'looking cool', I meant the code with its fancy formats and highlighting, no snark intended :)
LLMs are a great tool for building, especially for those who could conceivably be engineer team leads/project managers. For greenfield learning, its sub-par unless its guided well.
I have found the LLMs work best where I have some knowledge already, and I can spot the gaffes and mistakes more quickly. It has to be rough on people flying them blind, as it were.
Unrelated but I have tried out LLMs on some legal research, and it has been appallingly bad. The LLMs get the gist of a lot of it, but falls down completely because the legal realm requires a certain amount of exactitude. Time and time again I will ask about laws in a certain U.S. State, and it will show me a law from Arizona or something which has no or limited applicability (you can try to cross reference other State laws in certain State courts, but usually that doesn't go very far unless it is a very novel argument).
You hit the nail on the head with the fancy outputs, I am guilty as charged there. The LLMs are frankly better at it then I am.
Terry Tao talks about this challenge in some of his blog posts and interviews, it's pretty interesting.
Yes, for inspiration or search they are great tools.
Review enough of those without doing the proving yourself, and you'll soon lose the ability to review. And have LLMs do the course work for you at math school, and you might never acquire such ability to begin with anyway.