I had to code something on a plane today. It used to be that you couldn't get you packages or check stackoverflow. But now, I'm useless. My mind has turned to pudding. I cannot remember basic boilerplate stuff. Crazy how fast that goes.
Use it or lose it, as it were.
0: https://www150.statcan.gc.ca/n1/daily-quotidien/241210/dq241...
This is similar to the photo-taking impairment effect where people will remember an event more poorly if they took photos at the event. Their brain basically subconsciously decides it doesn’t need to remember the event because the camera will remember the event instead.
If the tool is powerful enough to do a better job than our brains would, it's a big win. In fact, we built the entire technological civilization on one such fundamental tool: writing.
Or from another perspective: our brains excel at adapting to the environment we find ourselves in. The tools we build, the technology we create, are parts our environment.
make stock number go up and up and up and people get in the way
Edit - lol @ the bozo who downvoted my post. Is that you scam Altman?
I don't think there is that much value in memorizing rarely used, easily looked up information.
Facts alone are like pebbles on a beach, far better (IMO) to have a few stones mortared with understanding to make a building of knowledge. A fanciful metaphor but you know ...
(Can't say time is the limiting factor either -- we're both in HN comments, valuing our own time at zero.)
And of course, what if your phone dies?
There is limited time, of course - no one can learn everything, but you can pay attention to the important facts, and the connections between them.
In some ideal world you would learn every fact there is, and the connections would fall out on their own, but in the real world we have to construct theories and frameworks to organise facts.
Let's for example about html boilerplate, where you don't remember the syntax. What you remember is the components & why they are needed, then add them one by one as you recall your memory. Doctype, html tag, head, body, etc. It works because html is simple and common.
Then for express it is harder, because you need to recall javascript syntaxes and express syntaxes, and most of the time you don't get involved with express outside req and res. You recall that express need body parser, register routers, and finally listen, whether you use http server first or directly from express. Now you compose one by one, looking at docs or web for the forgotten pieces, but you don't lose the understanding / logic of express, you just forget the syntaxes.
As for stream where I keep forgetting it, I just need to remember that stream need source, event handler such as on data, error, finish / end. Pipe if needed. However I never remember whether to use writable, readable, streamable, etc because I seldom get involved with them, and can look up for references anytime.
To use another metaphor, you can't REALLY see the forest amongst the trees, if you don't consider the trees themselves.
One of the reasons I like history so much is because, with enough facts accumulated, you can see how one piece of information flows into another - e.g. dates matter, because knowing the precise order in which important events occur helps you determine how those events may or may not have affected each other in the course of their unfolding.
Sure memorizing dates is boring on its own, but putting them in contexts is exciting - you still need to comb the beaches to find the right stones!
I guess an interesting counterpoint to what I said is something like https://en.wikipedia.org/wiki/Phantom_time_conspiracy_theory (and similar) where a grandiose framework tries to fit inconvenient facts into a shape that is entirely invented.
easily looked up - we don't have that any more since Google decided to entshittify search. What you now have is not looked up information, as that would look the same each time you "looked it up" - instead of a quick 3-4 word search pattern you now write an elaborate verbose "query", and get a chewed up re-interpretation by the shitty LLMs. And then since sometimes its not quite what you asked for, you have to ask it again or redirect it and just like that, you've wasted 5 mins of your time arguing with a goddamn neural network!
> Yet I can code normally provided there's internet accessible.
I'm the opposite. Yes I need my computer to test things fully, but I'm able to code on paper. I want my computer to be a complete sufficient node, so I mostly install documentation and my computer is mostly not connected to the internet, unless I actively enable it to do a specific thing.
2 years ago you had downloaded onto your laptop an effective and useful summary of all of the information on the Internet, that could be used to generate computer programs in an arbitrarily selected programming language.
I wrote a post about it: Your toaster will know mesopotamian history because it’s more expensive not too.
https://wanderingstan.com/2026-03-01/your-toaster-will-know-...
I'm sure people would get a cheaper toaster in exchange of an ad being burned in your bread.
And as other commenter pointed out, a smart toaster with ads or data collection can be subsidized and thus be more profitable. (Oh what a world we're headed for!)
In any case, I think the LLM-everywhere thesis holds even strong for even moderate-complexity devices like power plugs, microwaves, and mobile phones.
But will it know the difference between too and to?
Ask your local model a verifiable question - for example a list of tallest buildings in Europe. I did it with Gemma on my laptop, and after the top 3 they were all fake. I just tried that again with Gemma-4 on my iphone, and it did even worse - the 3 tallest buildings in Europe are apparently the Burj Khalifa, the Torre Glories and the Shanghai Tower.
I wouldn't call that effective compression of information.
But what you can do with local models is give them actual data and tools to search it. Download a copy of Wikipedia locally, give the agent a way to search it and BOOM accurate information without an internet connection.
Also "small enough to live on disk" is a bit vague, especially when models get super stupid super fast when you get to the smaller size. At that point they're just basically 40k servitors that can use tools and nothing much.
> I cannot remember basic boilerplate stuff.
I don't know exactly what you mean by boilerplate stuff, but honestly, that's stuff we should have automated away prior to AI. We should not be writing boilerplate.
I'd highly encourage you to take the time to automate this stuff away. Not even with AI, but with scripts you can run to automate boilerplate generation. (Assuming you can't move it to a library/framework).
Last time I looked there were at least seven ways to do it.
As another commenter said, the affordability of LLM subscriptions (or, as others are predicting, the lack thereof) is the primary concern, not the technology itself stealing away your skills.
I am far from the definitive voice in the does-AI-use-corrupt-your-thinking conversation, and I don't want to be. I don't want LLMs to replace my thinking as much as the next person, but I also don't want to shun anything useful that can be gained from these tools.
All that said, I do feel that perhaps "dumber" LLMs that work on-device first will allow us to get further and be better, more reliable tools overall.
Were people actually physically typing every character of the software they were writing before a couple of years ago?
I on the other hand am a software engineer, so writing code is part of the job title.
Company? Helm? Whatever vi uses that's like company and helm? Haven't IDEs written function calls for you for like decades now?
I work on greenfield projects so I see my fair share of boilerplate, but honestly, it's just a minute part of work that's almost meditative to write a little bit of trivial code (i.e. a function signature) in between sessions of hard thinking. Writing boilerplate is very far down the list of things I seek to optimize.
Also I don't use LLM.
I'm not super worried, either I still do the last leg of the work, or I go back an abstraction level with my prompts and work there
I do use AI daily to help me enhance code but then... I also very regularly turn off, physically, the link between a sub-LAN at home and the Internet and I still can work. It's incredibly relaxing to work on code without being connected 24/7. Other machines (like kid's Nintendo switch) can still access the Internet: but my machines for developing are cut off. And as I've got a little infra at home (Proxmox / VMs), I have quite a few services without needing to be connected to the entire world: for example I've got a pastebin, a Git server, a backuping procedure, all 100% functional without needing to be connected to the net (well 99% for the backuping procedure as the encrypted backup files won't be synch'ed with remote servers until the connection is operational).
Sure it's not a "laptop on a plane", but it's also not "24/7 dependent on Sam Altman or Anthropic".
I'll probably enhance my setup at some point with a local Gemma model too.
And all this is not mutually exclusive with my Anthropic subscription: at the flick of a switch (which is right in front of me), my sub-LAN can access the Internet again.
I remember at that time, my "mentor" suggested to memorize all the "keywords" from C (which were few). But given my bad memory I had to constantly look at the book.
Aaah how times have changed.
So I don't think this is all AI tbh.
The early this year, in my country (Uganda) the entire internet was shutdown during elections for close to 1 week.
A lot of devs in the country couldn't do anything during that period.
So much, some actually claimed the moved to nearby Nairobi Kenya.
I learned the Q array language five years ago and then didn't touch it for six months. I was surprised how little I remembered when I tried to resume.
The fact that this is being called out is strange.
Yes, you lost some abilities. Install local model so you have someone to talk to while you are on the plane ;)
Repeat this for decades, and "very basic programming tasks" might be creating a cross-platform browser by using LLMs via voice dictation.
I'm wondering if this is something that hits new developers faster than more experienced ones?
Almost certainly, at least according to Ebbinghaus' forgetting curve.
People using AI for tasks (essay writing in the MIT study linked below) showed lower ownership, brain connectivity, and ability to quote their work accurately.
> https://arxiv.org/abs/2506.08872
There was a MSFT and Carnegie Mellon study that saw a link between AI use, confidence in ones skills, confidence in AI, and critical thinking. The takeaway for me is that people are getting into “AI take the wheel” scenarios when using GenAI and not thinking about the task. This affects people novices more than experts.
If you managed to do critical thinking, and had relegated sufficient code to muscle memory, perhaps you aren’t as impacted.
My theory is that if you're not full-time coding, it's hard harder to remember the boiler plate and obligatory code entailed by different SDKs for different modules. That's where the documentation reading time goes, and what slows down debugging. That's where agent assisted coding helps me the most.
As an example I have been fighting with agents re-writing or removing guard clauses and structs when dealing with Mach-o fat archives this week, I finally had to break the parsing out into an external module and completely remove the ability for them to see anything inside that code.
I get the convenience for prototyping and throwaway code, but the problem is when you don’t have enough experience with the quirks to know something is wrong.
It will be code debt if one doesn’t understand the core domain. That is the problem with the confidence and surface level competence of these models that we need to develop methods for controlling.
Writing code is rarely the problem with programming in general, correctness and domain needs are the hard parts.
I hope we find a balance between gaining value from these tools while not just producing a pile of fragile abandonware
I guess that's an advantage? People shouldn't have to burden their memory with boilerplate and CRUD code.
The people who used chatGPT had the most difficulty quoting their own work. So not boilerplate, CRUD - but yes the advantage is clear for those types of tasks.
There were definite time and cognitive effort savings. I think they measured time saved, and it was ~60% time saved, and a ~32% reduction in cognitive effort.
So its pretty clear, people are going to use this all over the place.
i dont think it is entirely a case of voluntary outsourcing of critical thinking. I think it's a problem of 1) total time devoted to the task decreasing, and 2) it's like trying to teach yourself puzzle solving skills when the puzzles are all solved for you quickly. You can stare at the answer and try to think about how you would have arrived at it, and maybe you convince yourself of it, but it should be relatively common sense that the learning value of a puzzle becomes obsolete if you are given the answer.
Maybe the difference between actually knowing stuff vs surface level? I know a lot of devs just know how to glue stuff together, not really how to make anything, so I'd imagine those devs lose their skills much faster.
Played with these coding agents for the last couple weeks and instantly noticed the brainrot when I was staring at an empty vim screen trying to type a skeleton helloworld in C.
Luckily the right idioms came back after couple of hours, but the experience gave me a big scare.
Very interesting, wonder what makes our experiences so different? For you "playing for a couple of weeks" have a stronger effect than for me after using them almost exclusively for more than a year, and I don't think I'm an especially great programmer or anything, typical for my experience I think.
Right, sounds very credible to me. What did you write, an addition function in each of those?
I wrote a little toy-calculator in each, ended up being ~250 LOC in each of them, not exactly the biggest test but large enough to see if my muscle memory still works which I was happy to discover it still did.
> But now, I'm useless. My mind has turned to pudding. I cannot remember basic boilerplate stuff
> Played with these coding agents for the last couple weeks and instantly noticed the brainrot when I was staring at an empty vim screen trying to type a skeleton helloworld in C.
This is very different from what I'm (not) experiencing. My test was for if I can remember the basic syntax of the language itself, I was never a big framework user, so of course using a framework is about the least interesting test I could do of myself.
Instead, I did the bare minimum to see if my "mind has turned to pudding" or "instantly noticed the brainrot", which would have been visible even for a toy calculator, obviously.
> the people pushing for "AI writing all the code" are not worried about you retaining the "muscle memory" to write addition and substraction functions
What are they worried about then? From your perspective, sounds like they're worried about "using these and those frameworks" but that's far from "real world work" in my experience, and really the least interesting thing you could remember as a developer.
So you only ever wrote code in academic setting? Not being sarcastic, but no realistic software development in commercial setting will happen without frameworks.
> That's not how I understood other's experience to be, they're describing something that won't let them even write toy calculators. Selected quotes:
Well that is literally not what they are describing. The man said "boilerplate code". Again, unless most of the code you wrote was code used in purely academic setting, whether research or teaching, we all understand boilerplate code to mean something like "code to open a file descriptor / iterate through database rows / poll web server for results" etc. So the typical stuff you would implement while relying on abstractions and concepts someone else already defined for you.
> What are they worried about then?
If I had to guess - they are worried about the growing LLM-backlash working against their hoped-for-industry-wide-adoption and their infinite money at some point running out, because at the end of the day, NVIDIA lending OpenAI 100B to invest in MS Azure and Azure using those 100B to purchase NVIDIA chips...is starting to look a lot like circular financing.
> sounds like they're worried about "using these and those frameworks" but that's far from "real world work" in my experience
Again, if your area of work is academic or teaching, then this probably matches your own real world experience. The problem for the LLM-crowd is that there is a lot more people here whose "real world experience" are not toy calculators and isolated algorithm implementation, but actually yes, using "this and that framework", as otherwise the implementation would take un-realistically long. I may know how TLS works, but would I implement it's handshake routine on my own? Only if I did it as a personal excercise, not in a commercial scenario. Same for UI, same for DB ORMs, etc. You name it.
I do still write stuff manually frequently - I often spend 5 minutes writing structs or function signatures to make sure that the LLM won't misunderstand or make something up. Maybe that's why I haven't lost it.
No we don't and we never should actually, compilers need to be deterministic.
With the right settings, a LLM is deterministic. But even then, small variations in input can cause very unforeseen changes in output, sometimes drastic, sometimes minor. Knowing that I'm likely misusing the vocabulary, I would go with saying that this counts as the output being chaotic so we need compilers to be non-chaotic (and deterministic, I think you might be able to have something that is non-deterministic and non-chaotic). I'm not sure that a non-chaotic LLM could ever exist.
(Thinking on it a bit more, there are some esoteric languages that might be chaotic, so this might be more difficult to pin down than I thought.)
Also, give the same programming task to 2 devs and you end up with 2 different solutions. Heck, have the same dev do the same thing twice and you will have 2 different ones.
Determinism seems like this big gotcha, but in it self, is it really?
"Do the same thing" I need to be pedantic here because if they do the same thing, the exact same solution will be produced.
The compiler needs to guarantee that across multiple systems. How would QA know they're testing the version that is staged to be pushed to prod if you can't guarantee it's the same ?