AWS CEO says using AI to replace junior staff is 'Dumbest thing I've ever heard'
theregister.com
theregister.com
In the Swedish schoolsystem, the idea for the past 20 years has been exactly this, that is to try to teach critical thinking, reasoning, problem solving etc rather than hard facts. The results has been...not great. We discovered that reasoning and critical thinking is impossible without a foundational knowledge about what to be critical about. I think the same can be said about software development.
The most damning example I have about Swedish school system is anecdotal: by attending Saturday school, I never had to study math ever in the Swedish school. (same for my Asian classmates) when I finished 9th grade Japanese school curriculum taught ONLY one day per week (2h), I had learned all of advanced math in high school and never had to study math until college.
The focus on "no one left behind == no one allowed ahead" also meant that young me complaining math was boring and easy didn't persuade teachers to let me go ahead, but instead, they allowed me to sleep during the lecture.
It's like this in the US (or rather, it was 20 years ago. But I suspect it is now worse anyway)
Teachers in my county were heavily discouraged from failing anyone, because pass rate became a target instead of a metric. They couldn't even give a 0 for an assignment that was never turned in without multiple meetings with the student and approval from an administrator.
The net result was classes always proceeded at the rate of the slowest kid in class. Good for the slow kids (that cared), universally bad for everyone else who didn't want to be bored out of their minds. The divide was super apparent between the normal level and honors level classes.
I don't know what the right answer is, but there was an insane amount of effort spent on kids who didn't care, whose parents didn't care, who hadn't cared since elementary school, and always ended up dropping out as soon as they hit 18. No differentiation between them, and the ones who really did give a shit and were just a little slow (usually because of a bad home life).
It's hard to avoid leaving someone behind when they've already left themselves behind.
I’m curious, could you share your Saturday school‘s system? I’m very interested in knowing what a day of class was like, the general approach, etc.
https://gpseducation.oecd.org/CountryProfile?plotter=h5&prim...
https://gpseducation.oecd.org/CountryProfile?plotter=h5&prim...
The exams were typically essay-ish (even in science classes) where you either had to basically reiterate the reasoning for a fact you already knew, or use similar reasoning to establish/discover a new fact (presumably unknown to you because not taught in class).
Unfortunately, it didn't work for me and I still have about the same critical thinking skills as a bottle of Beaujolais Nouveau.
For example in electricity you need at least that amount of cross section if doing X amount of amps over Y length. I want to dig down and understand why? Ohh, the smaller the cross section, the more it heats! Armed with this info I get many more "Ohhs": Ohh, that's why you must ensure the connections are not loose. Oohhh, that's why an old extension cord where you don't feel your plug solidly clicks in place is a fire hazard. Ohh, that's why I must ensure the connection is solid when joining cables and doesn't lessen cross section. Ohh, that's why it's a very bad idea to join bigger cables with a smaller one. Ohh, that's why it is a bad idea to solve "my fuse is blowing out" by inserting a bigger fuse but instead I must check whether the cabling can support higher amperage (or check whether device has to draw that much).
And yeah, this "intuition" is kind of a discovery phase and I can check whether my intuition/discovery is correct.
Basically getting down to primitives lets me understand things more intuitively without trying to remember various rules or formulas. But I noticed my brain is heavily wired in not remembering lots of things, but thinking logically.
It is however not taught very well by some teachers, who skirt on explaining how to properly do it, which might be your case.
Why do you say so? Even just stating this probably means you are one or a few steps further...
In its/your/our defense, I think it’s a perfectly smart wine, and young at heart!
I'm loving this expression. May I please adopt it?
I'm not sure I'd agree that it's been outright "not great". I myself am the product of that precise school-system, being born in 1992 in Sweden (but now living outside the country). But I have vivid memories of some of the classes where we talked about how to learn, how to solve problems, critical thinking, reasoning, being critical of anything you read in newspapers, difference between opinions and facts, how propaganda works and so on. This was probably through year/class 7-9 if I remember correctly, and both me and others picked up on it relatively quick, and I'm not sure I'd have the same mindset today if it wasn't for those classes.
Maybe I was just lucky with good teachers, but surely there are others out there who also had a very different experience than what you outline? To be fair, I don't know how things are working today, but at least at that time it actually felt like I had use of what I was thought in those classes, compared to most other stuff.
In the world of software development I meet a breed of Swedish devs younger than 30 that can't write code very well, but who can wax Jira tickets and software methodologies and do all sort of things to get them into a management position without having to write code. The end result is toxic teams where the seniors and the devs brought from India are writing all the code while all the juniors are playing software architect, scrum master an product owners.
Not everybody is like that; seniors tend to be reliable and practical, and some juniors with programming-related hobbies are extremely competent and reasonable. But the chunk of "waxers" is big enough to be worrying.
Sweden is the 19th country in the PISA scores. And it is in the upper section on all education indexes. There has been a world decline on scores, but has nothing to do with the Swedish education system. (That does not mean that Sweden should not continue monitoring it and bringing improvements)
From Swedish news: https://www.sverigesradio.se/artikel/swedish-students-get-hi...
- Swedish students skills in maths and reading comprehension have taken a drastic downward turn, according to the latest PISA study.
- Several other countries also saw a decline in their PISA results, which are believed to be a consequence of the Covid-19 pandemic.
Having teenagers that's been through most of the primary and secondary schools I kind agree with GP, especially when it comes to math,etc.
Teaching concepts and ideas is _great_, and what we need to manage with advanced topics as adults. HOWEVER, if the foundations are shaky due to too little repetition of basics (that is seemingly frowned upon in the system) then being taught thinking about some abstract concepts doesn't help much because the tools to understand them aren't good enough.
The result is usually bottom of the barrel in the subjects that don’t fit that model well, mostly languages and math - the latter being the main issue as it becomes a bottleneck for teaching many other subjects.
It also creates a tendency for people to take what they learn as truth, which becomes an issue when they use less reputable sources later in life - think for example a person taking a homeopathy course.
Lots of parroting and cargo culting paired with limited cultural exposition due to monolingualism is a bad combination.
Well, I kind of disagree. The results are bad mainly because we have a mass immigration from low education countries with extremely bad cultures.
If you look at the numbers, it's easy to say swedes are stupid when in the real sense, ethnic swedes do very well in school.
Media can fill that gap. People should be critical about global warming, antivax, anti israel, anti communism, racism, hate, whitr man, anti democracy, russia, china, trump...
This thing is bad, imhate it, problem solved! Modern critical thinking is pretty simple!
In future goverment can provide daily RSS feed, of things to be critical about. You can reduce national schooling system to a single vps server!
You can’t teach critical thinking like that.
You need to teach hard facts and then people can learn critical thinking inductively from the hard facts with some help.
On a side note.. ya’ll must be prompt wizards if you can actually use the LLM code.
I use it for debugging sometimes to get an idea, or a quick sketch up of an UI.
As for actual code.. the code it writes is a huge mess of spaghetti code, overly verbose, with serious performance and security risks, and complete misunderstanding of pretty much every design pattern I give it..
Personally, I wrote 200K lines of my B2B SaaS before agentic coding came around. With Sonnet 4 in Agent mode, I'd say I now write maybe 20% of the ongoing code from day to day, perhaps less. Interactive Sonnet in VS Code and GitHub Copilot Agents (autonomous agents running on GitHub's servers) do the other 80%. The more I document in Markdown, the higher that percentage becomes. I then carefully review and test.
Perhaps that's part of it.
People here work on all kinds of industries. Some of us are implementing JIT compilers, mission-critical embedded systems or distributed databases. In code bases like this you can't just wing it without breaking a million things, so LLM agents tend to perform really poorly.
Which problem are you trying to solve?
At this point my assumption is they learned that talking about this question will very quickly reveal that "the great things I use LLMs for" are actually personal throwaway pieces, not to be extended above triviality or maintained over longer than a year. Which, I guess, doesn't make for a great sales pitch.
If you are writing code that is/can be "heavily borrowed" - things that have complete examples on Github, then an LLM is perfect.
In the long run, that is going to be what drives their quality. At some point the conversation is going to evolve from whether or not AI-assisted coding works to what the price point is to get the quality you need, and whether or not that price matches its value.
There is the huge variance in prompt specificity as well as the subtle differences inherent to the models. People often don't give examples when they talk about their experiences with AI so it's hard to get a read on what a good prompt looks like for a given model or even what a good workflow is for getting useful code out of it.
We have 1-2 small python services based on Flask and Pydantic, very structured and a well-written development and extension guide. The newer Copilot models perform very well with this, and improving the dev guidelines keep making it better. Very nice.
We also have a central configuration of applications in the infrastructure and what systems they need. A lot of similarly shaped JSON files, now with a well-documented JSON schema (which is nice to have anyway). Again, very high quality. Someone recently joked we should throw these service requests at a model and let it create PRs to review.
But currently I'm working in Vector and it's Vector remap language... it's enough of a mess that I'm faster working without any copilot "assistance". I think the main issue is that there is very little VRL code out in the open, and the remaps depend on a lot of unseen context, which one would have to work on giving to the LLM. Had similar experiences with OPA and a few more of these DSLs.
That would probably be 1000 line of Common Lisp.
I also think it’s that people don’t know how to use the tool very well. In my experience I don’t guide it to do any kind of software pattern or ideology. I think that just confuses the tool. I give it very little detail and have it do tasks that are evident from the code base.
Sometimes I ask it to do rather large tasks and occasionally the output is like 80% of the way there and I can fix it up until it’s useful.
No model ive tried can write, usefully debug or even explain cmake. (It invents new syntax if it gets stuck, i often have to prompt multiple AI to know if even the first response in the context was made-up)
My luck with embedded c has been atrocious for existing codebase (burning millions of tolkens), but passable for small scripts. (Arduino projects)
My experience with python is much better. Suggesting relevant libraries and functions, debugging odd errors, or even making small script on its own. Even the original github copilot which i got access to early was excellent on python.
Alot of people that seem to have fully embraced agentic vibe-coding seem to be in the web or node.js domain. Which I've not done myself since pre-AI.
I've tried most (free or trial) major models or schemes in hope that i find any of them useful, but not found much use yet.
We probably do, yes. the Web domain compared to a cybersecurity firm compared to embedded will have very different experiences. Because clearly there's a lot more code to train on for one domain than the other (for obvious reasons). You can have colleagues at the same company or even same team have drastically different experiences because they might be in the weeds on a different part of tech.
> I then carefully review and test.
If most people did this, I would have 90% less issues with AI. But as we expect, people see shortcuts and use them to cut corners, not give more times to polish the edges.
If you look, there are people out there approaching this stuff with more objectivity than most (mitsuhiko and simonw come to mind, have a look through their blogs, it's a goldmine of information about LLM-based systems).
Betting in advance that it's JavaScript or Python, probably with very mainstream libraries or frameworks.
The resulting code included an API call to run arbitrary SQL queries against the DB. Even after pointing this out, this API call was not removed or at least secured with authentication rules but instead /just/hidden/through/obscur/paths...
The big 3 - Opus 4.1 GPT5 High, Gemini 2.5 Pro
Are astonishing in their capabilities, it's just a matter of providing the right context and instructions.
Basically, "you're holding it wrong"
That said, observing attempts by skeptics to “unsuccessfully” prompt an LLM have been illuminating.
My reaction is usually either:
- I would never have asked that kind of question in the first place.
- The output you claim is useless looks very useful to me.
The less code I write to solve a problem the happier I am.
But I have not yet been able to consistently get value out of vibe coding. It's great for one-off tasks. I use it to create matplotlib charts just by telling it what I want and showing it the schema of the data I have. It nails that about 90% of the time. I have it spit out close-ended shell scripts, like recently I had it write me a small CLI tool to organize my Raw photos into a directory structure I want by reading the EXIF data and sorting the images accordingly. It's great for this stuff.
But anything bigger it seems to do useless crap. Creates data models that already exist in the project. Makes unrelated changes. Hallucinates API functions that don't exist. It's just not worth it to me to have to check its work. By the time I've done that, I could have written it myself, and writing the code is usually the most pleasurable part of the job to me.
I think the way I'm finding LLMs to be useful is that they are a brilliant interface to query with, but I have not yet seen any use cases I like where the output is saved, directly incorporated into work, or presented to another human that did not do the prompting.
I use aider and your description doesn't match my experience, even with a relatively bad-at-coding model (gpt-5). It does actually work and it does generate "good" code - it even matches the style of the existing code.
Prompting is very important, and in an existing code base the success rate is immensely higher if you can hint at a specific implementation - i.e. something a senior who is familiar with the codebase somewhat can do, but a junior may struggle with.
It's important to be clear eyed about where we are here. I think overall I am still faster doing things manually than iterating with aider on an existing code base, but the margin is not very much, and it's only going to get better.
Even though it can do some work a junior could do, it can't ever replace a junior human... because a junior human also goes to meetings, drives discussions, and eventually becomes a senior! But management may not care about that fact.
Using the following prompt:
Write a rust serde implementation for the ORB binary data format.
Here is the background information you need:
* The ORB reference material is here: https://github.com/kstenerud/orb/blob/main/orb.md
* The formal grammar dscribing ORB is here: https://github.com/kstenerud/orb/blob/main/orb.dogma
* The formal grammar used to describe ORB is called Dogma.
* Dogma reference material is here: https://github.com/kstenerud/dogma/blob/master/v1/dogma_v1.0.md
* The end of the Dogma description document has a section called "Dogma described as Dogma", which contains the formal grammar describing Dogma.
Other important things to remember:
* ORB is an extension of BONJSON, so it must also implement all of BONJSON.
* The BONJSON reference material is here: https://github.com/kstenerud/bonjson/blob/main/bonjson.md
* The formal grammar desribing BONJSON is here: https://github.com/kstenerud/bonjson/blob/main/bonjson.dogma
If you spend a lot of time thinking about what you want, describing the inner workings, edge cases, architecture and library choices, and put that into a thoughtful markdown, then maybe after a couple of iterations you will get half decent code. It certainly makes a difference between that and a short "implement X" prompt.
But it makes one think - at that point (writing a good prompt that is basically a spec), you've basically solved the problem already. So LLM in this case is little more than a glorified electric typewriter. It types faster than you, but you did most of the thinking.
Now it requires me to fully review all code and understand what the LLM is doing at the functional, class level and api level- in fact it works better at the method or component level for me and I had a lot of cleanup work (and lots of frustration with the models) on the codebase but overall there’s no way that I could equal the velocity I have now without it
I don't know what to tell you. I just talk to the thing in plain but very specific English and it generally does what I want. Sometimes it will do stupid things, but then I either steer it back in the direction I want or just do it myself if I have to.
The agents are able to iterate and work with the compiler until it gets it right and the combination of 1 and 2 means there’s fewer possible “right answers” to whatever problem I have. If i structure my prompte to basically fill in the blanks of my code in specific areas it saves a lot of time. Most of what I prompt is something already done, and usually 1 google search away. This saves me the time to search it up, figure out whatever syntax I need, etc.
AI is also pretty good if you get it to do small chunks of code for you. This means you come with the architecture, the implementation details, and how each piece is structured. When I walk AI through each unit of code I find the results are better, and it's easier for me to address issues as I progress.
This may seem some what redundant, though. Sometimes it's faster to just do it yourself. But, with a toddler who hates sleep I've found I've been able to maintain my velocity... Even on days I get 3 hrs of sleep.
Even I can see they have big blind spots. As the parent said I get overly verbose code that does run, but is no where near the best solution. Well, for really common problems and patterns I usually get a good answer. Need a more niche problem solved?You better brush up your Googling skills and do some research if you care about code quality.
// assign "bar" to foo
const foo = "bar";
They love to do that shit. I know you can prompt it not to. But the amount of PRs I'm reviewing these days that have those types of comments is insane.
Granted, I may be a inexpert prompter, but at the same time, I'm asking for basic things, as a test, and it just fails miserably most of the time.
Thank you, but I don’t feel that way.
I’d ask you a lot of details…what tool, what model, what kind of code. But it’d probably take a lot to get to the bottom of the issue.
I think a lot of us eventually arrive at a point where our jobs get a bit boring and all the work starts to look like some permutation of past work. If instead of going to work and spending two hours adding some database fields and writing some tests, you had the opportunity to either:
A) Do the thing as usual in the predictable two hours
B) Spend an hour writing a detailed prompt as if you were instructing a junior engineer on a PIP to do it, and doing all the typical cognitive work you'd have done normally and then some, but then instead of typing out the code in the next hour, you have a random chance to either press enter, and tada the code has been typed and even kinda sorta works, after this computer program was "flibbertigibbeting" for just 10 minutes. Wow!
Then you get that sweet dopamine hit that tells you you're a really smart prompt engineer who did a two hour task in... cough 10 minutes. You enjoy your high for a bit, maybe go chat with some subordinate about how great your CLAUDE.md was and if they're not sure about this AI thing it's just because they're bad at prompt engineering.
Then all you have to do is cross your t's and dot your i's and it's smooth sailing from there.Except, it's not. Because you (or another engineer) will probably find architectural/style issues when reviewing the code that you explicitly told it to follow, but it ignored, and you'll have to fix those. You'll also probably be sobering up from your dopamine rush by now, and realize that you have to either review all the other lines of AI generated code, which you could have just correctly typed once.
But now you have to review with an added degree of scrutny, because you know it's really good at writing text that looks beautiful, but is ever so slightly wrong in ways that might even slip through code review and cause the company to end up in the news.
Alternatively, you could yolo and put up an MR after a quick smell, making some other poor engineer do your job for you (you're a 10x now, you've got better things to do anyway). Or better yet, just have Claude write the MR, and don't even bother to read it. Surely nobody's going to notice your "acceptance critera" section says to make sure the changes have been tested on both Android and Apple, even though you're building a microservice for an AI-powered smart fridge (mostly just a fridge, except every now and then it starts shooting ice cubes across the room at mach 3). Then three months later when someone, who never realized there are three different identical "authenticate," spends an hour scratching their head about why the code they're writing is not doing anything (because it's actually running another redundant function that nobody ever seems to catch in MR review because they're not reflected in a diff.
But yeah, that 10 minute AI magic trick sure felt good. There are times when work is dull enough that option B sounds pretty good, and I'll dabble. But yeah, I'm not sure where this AI stuff leads but I'm pretty confident it won't taking over our jobs any time soon (an ever-increasing quota of H1Bs and STEM opt student visas working for 30% less pay, on the other hand, might).
Just yesterday I uploaded a few files of my code (each about 3000+ lines) into a gpt5 project and asked in assistance in changing a lot of database calls into a caching system, and it proceeded to create a full 500 line file with all the caching objects and functions I needed. Then we went section through section of the main 3000+ line file to change parts of the database queries into the cached version. [I didn't even really need to do this, it basically detected everything I would need changing at once and gave me most of it, but I wanted to do it in smaller chunks so I was sure what was going on]
Could I have done this without AI? Sure.. but this was basically like having a second pair of eyes and validating what I'm doing. And saving me a bunch of time so I'm not writing everything from scratch. I have the base template of what I need then I can improve it from there.
All the code it wrote was perfectly clean.. and this is not a one off, I've been using it daily for the last year for everything. It almost completely replaces my need to have a junior developer helping me.
I’m not even sure AI is good for any engineer, let alone junior engineers. Software engineering at any level is a journey of discovery and learning. Any time I use it I can hear my algebra teacher telling me not to use a calculator or I won’t learn anything.
But overall I’m starting to feel like AI is simply the natural culmination of US economic policy for the last 45 years: short term gains for the top 1% at the expense of a healthy business and the economy in the long term for the rest of us. Jack Welch would be so proud.
The CEO didn't express any concerns about "talent leaving". He is saying "keep the juniors" but he's implying "fire the seniors". This is in line with long standing industry trends and it's confirmed by the flowing quote from the OP:
>> [the junior replacement] notion led to the “dumbest thing I've ever heard” quote, followed by a justification that junior staff are “probably the least expensive employees you have” and also the most engaged with AI tools.
He is pushing for more of the same, viewing competence and skill as threats and liability to be "fixed". He's warning the industry to stay the course and keep the dumbing-down game moving as fast as possible.
The 2010's tech boom happened because big tech knew a good engineer is worth their weight in gold, and not paying them well meant they'd be headhunted after as little as a year of work. What's gonna happen when this repeats) if we're assuming AI makes things much more efficient)?
----
And that's my kindest interpretation. One that assumes that a junior and senior using a prompt will have a very close gap to begin with. Even seniors seem to struggle right now with current models working at scale on Legacy code.
There are lots of personal projects that I have wanted to build for years but have pushed off because the “getting started cost” is too high, I get frustrated and annoyed and don’t get far before giving up. Being able to get the tedious crap out of the way lowers the barrier to entry and I can actually do the real project, and get it past some finish line.
Am I learning as much as I would had I powered through it without AI assistance? Probably not, but I am definitely learning more than I would if I had simply not finished (or even started) the project at all.
> (…)
> I’m not even sure AI is good for any engineer
In that case I’m not sure you really agree with this CEO, who is all-in on the idea of LLMs for coding, going so far as to proudly say 80% of engineers at AWS use it and that that number will only rise. Listen to the interview, you don’t even need ten minutes.
Yes, but when there are certain mundane things in that discovery that are hindering my ability to get work done, AI can be extremely useful. It can be incredibly helpful in giving high level overviews of code bases or directing me to parts of codebases where certain architecture lives. Additionally, it exposes me to patterns and ideas I hadn't originally thought of.
Now, if I just take whatever is spit out by AI as gospel, then I'd be inclined to agree with you in saying AI is bad, but if you use it correctly, like any other tool, it's fantastic.
In the case of Amazon with a shit ton of money to throw at a team of employees to crush your little startup?
Okay, but what about work output? That's seems to be the only thing business cares about.
Also, maybe it's the HN bias but I don't see this notion where old engineers are rejecting this en masse. More younger people will embrace it. But most younger people haven't mucked in legacy code yet (the lifeblood of any businesses).
The thinking was that they could iterate faster, ship better code, and have an always on 10x engineer in the form of Claude code.
I've observed perfectly rational founders become addicted to the dopamine hit as they see Claude code output what looks like weeks or years of software engineering work.
It's overgenerous to allow anyone to believe AI can actually "think" or "reason" through complex problems. Perhaps we should be measuring time saved typing rather than cognition.
[1] vibebusters.com
https://www.ycombinator.com/companies/text-ai/jobs/OJBr0v2-f...
So far I don't see that notion disproved. Ai still doesn't truly "reason with" nor understand the data it outputs.
Pretty obvious conclusion that I think anyone who's thought seriously about this situation has already come to. However, I'm not optimistic that most companies will be able to keep themselves from doing this kind of thing, because I think it's become rather clear that it's incredibly difficult for most leadership in 2025 to prioritize long-term sustainability over short-term profitability.
That being said, internships/co-ops have been popular from companies that I'm familiar with for quite a while specifically to ensure that there are streams of potential future employees. I wonder if we'll see even more focus on internships in the future, to further skirt around the difficulties in hiring junior developers?
Independent thinking is indeed the most important skill to have as a human. However, I sympathize for the younger generations, as they have become the primary target of this new technology that looks to make money by completely replacing some of their thinking.
I have a small child and took her to see a disney film. Google produced a very high quality long form advert during the previews. The ad portrays a lonely young man looking for something to do in the evening that meets his explicit preferences. The AI suggests a concert, he gets there and locks eyes with an attractive young woman.
Sending a message to lonely young men that AI will help reduce loneliness. The idea that you don't have to put any effort into gaining adaptive social skills to cure your own loneliness is scary to me.
The advert is complete survivor bias. For each success in curing your boredom, how many failures are there with lonely young depressed men talking to their phone instead of friends?
Critical thinking starts at home with the parents. Children will develop beliefs from their experience and confirm those beliefs with an authority figure. You can start teaching mindfulness to children at age 7.
Teaching children mindfulness requires a tremendous amount of patience. Now the consequence for lacking patience is outsourcing your Childs critical thinking to AI.
There is also a movie called Her, with Joaquin Phoenix and ScarJo. Absolutely brilliant.
Ahmen! I attend this same church.
My favorite professor in engineering school always gave open book tests.
In the real world of work, everyone has full access to all the available data and information.
Very few jobs involve paying someone simply to look up data in a book or on the internet. What they will pay for is someone who can analyze, understand, reason and apply data and information in unique ways needed to solve problems.
Doing this is called "engineering". And this is what this professor taught.
Memorization is not a panacea. I never found memorizing l33t code problems to be edifying. I think it's because those kinds of tight, self-referential, clever programs are far removed from the activity of writing applications. Most working programmers do not run into a novel algorithm problem but once or twice a career. Application programming has more the flavor of a human-mediated graph-traversal, where the human has access to a node's local state and they improvise movement and mutation using only that local state plus some rapidly decaying stack. That is, there is no well-defined sequence for any given real-world problem, only heuristics.
A regular old competent developer can quickly pick up whatever stack is used. After all, they have to; Every company is their own bespoke mess of technologies. The idea that you can just slap "15 years of React experience" on a job ad and that the unicorn you get will be day-1 maximally productive is ludicrous. There is always an onboarding time.
But employers in this field don't "get" that. For regular companies they're infested by managers imported from non-engineering fields, who treat software like it's the assembly line for baking tins or toilet paper. Startups, who already have fewer resources to train people with, are obsessed with velocity and shitting out an MVP ASAP so they can go collect the next funding round. Big Tech is better about this, but has it's own problems going on and it seems that the days of Big Tech being the big training houses is also over.
It's not even a purely collective problem. Recruitment is so expensive, but all the money spent chasing unicorns & the opportunity costs of being understaffed just get handwaved. Rather spend $500,000 on the hunt than $50,000 on training someone into the role.
And speaking of collective problems. This is a good example of how this field suffers from having no professional associations that can stop employers from sinking the field with their tragedies of the commons. (Who knows, maybe unions will get more traction now that people are being laid off & replaced with outsourced workers for no legitimate business reason.)
- something we've been doing since forever
- the latest trend that can be picked up just-in-time if you'll ever need it
Now as a hiring manager I’ll say I regularly find that those who’ve had humanities experience are way more capable and the hard parts of analysis and understanding. Of course I’m biased as a dual cs/philosophy major but it’s very rare I’m looking for someone who can just write a lot of code. Especially juniors as analytical thinking is way harder to teach than how to program.
I miss her jokes against anxious nerds that just wanted to code :(
Don't forget the rise of boot camps where some educators are not always aligned with some sort of higher ethical standards.
However most of us are not in that situation. It is better for us to just look up those details as we need them because it gives us more room to handle a broader variety of situations.
Coding, doctors, plumber… different information, often similar skill sets.
I worked a job doing tech support for some enterprise level networking equipment. It was the late 1990s and we were desperate for warm bodies. Hired a former truck driver who just so happened to do a lot of woodworking and other things.
Great hire.
Unlike my teachers, none of my bosses ever put me in an empty room with only a pencil and a sheet of paper to solve given problems.
My experience as a professor and a student is that this doesn't make any difference. Unless you can copy verbatim the solution to your problem from the book (which never happens), you better have a good understanding of the subject in order to solve problems in the allocated time. You're not going to acquire that knowledge during your test.
Very anecdotally, but I hazard that most of these types of low-hanging fruit, low-value add roles are much less common since they tended to be blockers for operational improvement. Six-sigma, Lean, various flavors of Agile would often surface these low performers up and they either improved or got shown the door between 2005 - 2020.
Not that everyone is 100% all the time, every day, but what we are left with is often people that are highly competent at not just their task list but at their job.
In general, I also attend your church.
However, as I preached in that church, I had two students over the years.
* One was from an African country and told me that where he grew up, you could not "just look up data that might be relevant" because internet access was rare.
* The other was an ex US Navy officer who was stationed on a nuclear sub. She and the rest of the crew had to practice situations where they were in an emergency and cut off from the rest of the world.
Memorization of considerable amounts of data was important to both of them.
I think the problem with this is that it requires the professor to mentally fully engage when marking assignments and many educators do not have the capacity and/or desire to do so.
Finding the correct balance for a given class it hard. Generally, the lower level the education, the more it should be closed books because the more it is about being able to manually solve the smaller challenges that are already well solved so you build up the skills needed to even tackle the larger challenges. The higher the education level, the more it is about being able to apply those skills to then tackle a problem, and one of those skills is being able to pull relevant formulas and such from the larger body of known formulas.
I've had a frustrating experience the past few years trying to hire junior sysadmins because of a real lack of problem solving skills once something went wrong outside of various playbooks they memorized to follow.
I don't need someone who can follow a pre-written playbook, I have ansible for that. I need someone that understands theory, regardless of specific implementations, and can problem solve effectively so they can handle unpredictable or novel issues.
To put another way, I can teach a junior the specifics of bind9 named.conf, or the specifics of our own infrastructure, but I shouldn't be expected to teach them what DNS in general is and how it works.
But the candidates we get are the opposite - they know specific tools, but lack more generalized theory and problem solving skills.
I meam, yes, to an extent you can teach how to think: critical thinking and logic are topics you can teach and people who take their teaching to heart can become better thinkers. However, those topics cannot impart creativity. Critical thinking is called exactly that because it's about tools and skills for separating bad thinking from good thinking. The skill of generating good thinking probably cannot be taught; it can only be improved with problem-solving practice.
The bigger your mental toolbox the more effective you will be at solving the problems. Looking up a tool and learning just enough to use it JIT is much slower than using a handy tool that you already masterfully know how to use.
This is as true for physical tools as for programming concepts like algorithms and data structures. In the worst case you won’t even know to look for a tool and will use whatever is handy, like the proverbial hammer.
It’s also hard to teach people “how to think” while at the same time teaching them practical skills - there’s only so many hours in a day, and most education is setup as a way to get as many people as possible into shape for taking on jobs where “thinking” isn’t really a positive trait, as it’d lead to constant restructuring and questioning of the status quo
Forcing kids to sit and memorize facts isn’t suddenly going to make them a better thinker, but much of my process of being a better thinker is something akin to sitting around and memorizing facts. (With a healthy dose of interacting substantively and curiously with said facts)
Ahh, but this is part of the problem. Yes, they have access, but there is -so much- information, it punches through our context window. So we resort to executive summaries, or convince ourselves that something that's relevant is actually not.
At least an LLM can take full view of the context in aggregate and peel out signal. There is value there, but no jobs are being replaced
(I do mean memorisation fairly broadly, it doesn't have to mean reciting a meaningless list of items.)
For anyone looking for resources, may we recommend:
* The Art of Doing Science and Engineering by Richard Hamming (lectures are available on YouTube as well)
* Measurement by Paul Lockhart (for teaching mindset)
Most of us are also old enough to have had a chance to develop taste in code and writing. Many of the young generation lack the experience to distinguish good writing from LLM drivel.
in developing areas, they actually implement more modern models commonly, as its newer and free to implement newer things.
those newer models focus more on exactly this. teach a person how to go through the process of finding solutions. rather than 'knowing a lot to enable the process of thinking'.
not saying what is better or worse, but reading this comment and article it reminds me of this.
a lot of people i see, they know tons of interesting things, but anything outside of their knowledge is a complete mystery.
all the while ppl from developing areas learn to solve issues. alot of individuals from there also, get out of their poverty and do really well for themselves.
ofcourse, this is a generalization and doesnt hold up in all cases. but i cant help think about it.
a lot of my colleagues dont know how to solve problems simply because they dont RTFM. they rely on knowledge from their education which is already outdated before they even sign up.. i try to teach them to RTFM. it seems hopeless. they look at me , downwards, because i have no papers. but if shit hits the fan, they come to me. solve the prolbem.
a wise guy i met once said (likely not his words). there are 2 type of ppl. those who think in problems, and those who think in solutions.
id related that to education, not prebaked human properties.
My boss said we were gonna fire a bunch of people “because AI” as part of some fluff PR to pretend we were actually leaders in AI. We tried that a bit, it was a total mess and we have no clue what we’re doing, I’ve been sent out to walk back our comments.
VP->Public: "We'll replace all our engineers with AI in two years"
Boss->VP: "I mean we need to fire VPs because AI"
VP->Public: "Replacing people with AI is stupid"
You don't fire people if you anticipate a 100x growth. Who cares about saving 0.1% of your money in 10 years? You want to sell 100x / 1000x/ 10000x more .
So the story is hard to swallow. The real reason is as usual, they anticipate a downturn and want to keep earnings stable.
In other words, none of these stories make any sense, even if you take the AI superpower at face value.
All of sudden to ensure better support and separation of concerns people needed a team with a manager for each service. If this hadn't been the case, the industry as a whole can likely work with 40% - 50% less people eventually. Thats because at any given point in time even with a large monolithic codebase only 10 - 20% of the code base is in active evolution, what that means in microservices world is equivalent amount teams are sitting idle.
When I started out huge C++ and Java code bases were pretty much the norm, and it was also one of the reasons why things were hard and barrier to entry high. In this microservices world, things are small enough that any small group of even low productivity employees can make things work. That is quite literally true, because smaller things that work well don't even need all that many changes on a everyday basis.
To me its these kind of places that are in real trouble. There is not enough work to justify keeping dozens to even hundreds of teams, their managements and their hierarchies all working for quite literally doing nothing.
I think sometimes the definition of work gets narrowed to a point so infinitesimal that everyone but the speaker is just a lazy nobody.
There was an excellent article on here about working at enterprise scale. My experience has been similar. You get to do work that feels really real, almost like school assignments with instant feedback and obvious rewards when you're at a small company. When I worked at big companies it all felt like bullshit until I screwed it up and a senator was interested in "Learning more" (for example).
The last few 9s are awful hard to chase down and a lot of the steps of handling edge case failures or features are extremely manual.
You're committing the classic fallacy around microservices here. The services themselves are simpler. The whole software is not.
When you take a classic monolith and split it up into microservices that are individually simple, the complexity does not go away, it simply moves into the higher abstractions. The complexity now lives in how the microservices interact.
In reality, the barrier to entry on monoliths wasn't that high either. You could get "low productivity employees" (I'd recommend you just call them "novices" or "juniors") to do the work, it'd just be best served with tomato sauce rather than deployed to production.
The same applies to microservices. You can have inexperienced devs build out individual microservices, but to stitch them together well is hard, arguably harder than ye-olde-monolith now that Java and more recent languages have good module systems.
1.) Elon fired 80% of twitter and 3 years later it still hasn't collapsed or fallen into technical calamity. Every tech board/CEO took note of that.
2.) Every kid and their sister going to college who wants a middle class life with generous working conditions is targeting tech. Every teenage nerd saw those over employed guys making $600k from their couch during the pandemic.
Edit: huh, what’s with the downvote, is this wrong? Did I overstate it? Here’s the data: https://www.demandsage.com/twitter-employees/
https://nordicapis.com/the-bezos-api-mandate-amazons-manifes...
I think it depends on the industry. In safety critical systems, you need to be testing, making documentation, architectural artifacts, meeting with customers, etc
There's not that much idle time. Unless you mean idle time actually writing code and that's not always a full time job.
Big businesses don’t inherently require the complexity of architecture they have. There is always a path-dependent evolution and vestigial complexity proportional to how large and fast they grew.
The real purpose of large scale architecture is to scale teams much moreso than business logic. But why does headcount grow? Is it because domains require it? Sure that’s what ambitious middle managers will say, but the real reason is you have money to invest in growth (whether from revenue or from a VC). For any complex architecture there is usually a dramatically simpler one that could still move the essential bits around, it just might not support the same number of engineers delineated into different teams with narrower responsibilities.
The general headcount growth and architecture trajectory is therefore governed by business success. When we’re growing we hire and we create complex architecture to chase growth in as many directions as possible. Eventually when growth slows we have a system that is so complex it requires a lot of people just to understand and maintain—even if the headcount is longer justified those with power in the human structure will bend over backwards to justify themselves. This is where the playbook changes and a private equity (or Elon) mentality is applied to just ruthlessly cut and force the rest of the people how to keep the lights on.
I consider advances in AI and productivity orthogonal to all this. It will affect how people do their jobs, what is possible, and the economics of that activity, but the fundamental dynamics of scale and architectural complexity will remain. They’ll still hire more people to grow and look for ways to apply them.
The question is why. You mention microservices. I'm not convinced.
Many think it is "horizontals". Possible, these taxes add up it is true.
Perhaps it is cultural? Perhaps it has to do with the workforce in some manner. I don't know and AFAIK it has not been rigorously studied.
Finally, the c-suite is getting it.
"It just means that each of us has to get more in tune with what our customers need and what the actual end thing is that we're going to try to go build, because that's going to be more and more of what the work is as opposed to sitting down and actually writing code...."
https://www.businessinsider.com/aws-ceo-developers-stop-codi...
If you read the full remarks they're consistent with what he says here. He says "writing code" may be a skill that's less useful, which is why it's important to hire junior devs and teach them how to learn so they learn the skills that are useful.
Some subset of the population likes to pretend their workforce is a cost that provides less than zero value or utility, and all the value and utility comes from shareholders.
But if this isn't true, and collective skill is worth value, then saying anyone can have that with AI at least has some headwind on your share price - which is all they care about.
Does that offset a potential tailwind from slightly higher margins?
I don't think any established company should be cheerleading that anyone can easily upset their monopoly with a couple of carefully crafted prompts.
It was always kind of strange to me, and seemed as though they were telling everyone, our moat is gone, and that is good.
If you really believed anyone could do anything with AI, then the risk of PEs collapsing would be high, which would be bad for the capital class. Now you have to correctly guess what's the next best thing constantly to keep your ROI instead of just parking it in save havens - like FAANG.
Bedrock/Q is a great example of how Amazon works. If we throw $XXX at the problem and YYY SDEs at the problem we should be able to build Github Copilot, GPT-3, OpenRouter and Cursor ourselves instead of trying to competitively acquire and attract talent. The fact that Codewhisperer, Q and Titan barely get spoken about on HN or Twitter tells you how successful this is.
But if you have that perspective then the equation is simple. If S3 can make 5 XXL features per year with 20 SDEs then if we adopt “Agentic AI” we should be able to build 10 XXL features with 10 SDEs.
Little care is given to organizational knowledge, experience, vision etc. that is the value (in their mind) of leadership not ICs.
Not to say that‘s what the AWS CEO is doing—maybe it is, maybe it isn’t, I haven’t checked—I’m just commenting on the general idea.
Pasting the quote for reference:
> Amazon Web Services CEO Matt Garman claims that in 2 years coding by humans won't really be a thing, and it will all be done by networks of AI's who are far smarter, cheaper, and more reliable than human coders.
Unless this guy speaks exclusively in riddles, this seems incredibly inconsistent.
It can only mean one thing: the music is about to stop.
[1]: https://www.shrm.org/topics-tools/news/technology/ai-will-sh...
“My view is you absolutely want to keep hiring kids out of college and teaching them the right ways to go build software and decompose problems and think about it, just as much as you ever have.” - Matt Garman
"We will need fewer people doing some of the jobs that are being done today” - Amazon CEO Andy Jassy
Maybe they differ in degree but not in sentiment.
If you're quoting something, the only ethical thing to do is as verbatim as possible and with a sufficient amount of context. Speeches should not be cleaned up to what you think they should have said.
Now, the question of who you go to for quotes, on the other hand .. that's how issues are really pushed around the frame.
From the perspective of a former employee. I knew that going in though. I was 46 at the time, AWS was my 8th job and knowing AWS’s reputation from 2nd and 3rd hand information, I didn’t even entertain an opportunity that would have forced me to relocate.
I interviewed for a “field by design” role that was “permanently remote” [sic].
But even those positions had an RTO mandate after I already left.
There's an endless series of one pagers with this idea or that idea, but from what I witnessed first hand, the ones that stuck were the ones that made money.
Jassy was a decent guy when I was there, but that was a decade ago. A CEO is a PR machine more than anything else, and the AI hype train has been so strong that if you do anything other than saying AI is the truth, the light and the way, you lose market share to competitors.
AI, much like automation in general, does allow fewer people to do more, but in my experience, customer desires expand to fill a vacuum and if fewer people can do more, they'll want more to the point that they'll keep on hiring more and more people.
I'll never forget the sama AGI posts before o3 launched and the subsequent doomer posting from techies. Feels so stupid in hindsight.
Especially with the amount of money that was put into just astroturfing the technology as more than it is.
From a person who is responsible for delivering projects, I’ve never thought “it sure would be nice if I had a few junior devs”. Why when I can poach an underpaid mid level developer for 20% more?
You really want to believe, maybe even need to believe, that anyone who comes up with this idea in their head has never written a single line of code in their life.
It is on its face absurd. And yet I don't doubt for a second that Garman et al. have to fend off legions of hacks who froth at the mouth over this kind of thing.
> "Measuring software productivity by lines of code is like measuring progress on an airplane by how much it weighs." -- Bill Gates
Do we reward the employee who has added the most weight? Do we celebrate when the AI has added a lot of weight?
At first, it seems like, no, we shouldn't, but actually, it depends. If a person or AI is adding a lot of weight, but it is really important weight, like the engines or the main structure of the plane, then yeah, even though it adds a lot of weight, it's still doing genuinely impressive work. A heavy airplane is more impressive than a light weight one (usually).
I don't know how true this is in fact, but I do know how true this is in my work - you cannot apply some arbitrary "make the number bigger" goal to everything and expect it to improve anything. It feels a bit weird seeing "write more lines of code" becoming a key metric again. It never worked, and is damn-near provably never going to work. The value of source code is not in any way tied to its quantity, but value still proves hard to quantify, 40 years later.
0. https://www.folklore.org/Negative_2000_Lines_Of_Code.html
I just won't use that information in quite the excitable, optimistic way they offer it.
I bet Google has a lot of tools to say convert a library from one language to another or generate a library based on an API spec. The 30% of code these LLMs are supposedly writing is probably in this camp, not net novel new features.
I ask an AI 4 times to write a method for me. After it keeps failing, I just write it myself. AI wrote 80% of the code!
Undergraduate -> Graduate Student -> Post-doc -> Tenure/Senior
Some exceptions occur for people getting Tenure without post doc or people doing some other things like taking undergraduate in one or two years. But no one expect that we for whole skip the first two and then get any senior researchers.
The same idea applies anywhere, the rule is that if you don't have juniors then you don't get seniors so better prepare your bot to do everything.
In my experience if you do this and break the problem down into small pieces, the AI can implement the pieces for you.
It can save a lot of time typing and googling for docs.
That said, once the result exceeds a certain level of complexity, you can't really ask it to implement changes to existing code anymore, since it stops understanding it.
At which point you now have to do it yourself, but you know the codebase less well than if you'd hand written it.
So, my upshot is so far that it works great for small projects and for prototyping, but the gain after a certain level of complexity is probably quite small.
But then, I've also find quite some value in using it as a code search engine and to answer questions about the code, so maybe if nothing else that would be where the benefit comes from.
Appreciate you saying this because it is my biggest gripe in these conversations. Even if it makes me faster I now have to put time into reading the code multiple times because I have to internalize it.
Since the code I merge into production "is still my responsibility" as the HN comments go, then I need to really read and think more deeply about what AI wrote as opposed to reading a teammate's PR code. In my case that is slower than the 20% speedup I get by applying AI to problems.
I'm sure I can get even more speed if I improve prompts, when I use the AI, agentic vs non-agentic, etc. but I just don't think the ceiling is high enough yet. Plus I am someone who seems more prone to AI making me lazier than others so I just need to schedule when I use it and make that time as minimal as possible.
Haven't we learned that it almost always ends up in hollow PR and marketing theater?
Basically the solution to this is extending education so that people entering workforce are already at senior level. Of course this can't be financed by the students, because their careers get shortened by longer education. So we need higher taxes on the entities that reap the new spoils. Namely those corporations that now can pass on hiring junior employees.
You need people who can validate LLM-generated code. It takes people with testing and architecture expertise to do so. You only get those things by having humans get expertise through experience.
junior + llm, it even worse. they become prompt engineers
That's rich coming from AWS!
I think he meant "how do you think about adding unnecessary complexity to problems such that it can enable the maximum amount of meetings, design docs and promo packages for years to come"!
This is becoming unbreathable for hackers.
The hype train is going to keep on moving for a while yet though.
As someone who works in AI, any CEO who says that AI is going to replace junior workers has no f*cking clue what they are talking about.
And AI will hurt them in their own development and with it taking over the tasks they would normally cut their teeth on.
We'll have to find newer ways of helping the younger generation get in the door.
The challenge now is for companies, managers and mentors to adapt to more remote and AI assisted learning. If a junior can be taught that it's okay to reach out (and be given ample opportunities to do so), as well as how to productively use AI to explain concepts that they may feel too scared to ask because they're "basics", then I don't see why this would hurt in the long run.
Having AI function as a senior programmer for lots of junior programmers that helps them learn and limits the interruptions for human senior coders makes so much more sense.
Can SOME people's jobs be replaced by AI. Maybe on paper. But there are tons of tradeoffs to START with that approach and assume fidelity of outcome.
I don't mean that as a negative, he's doing great work explaining AI to (dev) masses!
LLMs are actually -the worst- at doing very specific repetitive things. It'd be much more appropriate for one to replace the CEO (the generalist) rather than junior staff.
This isn't exactly rocket science, and I'm pretty sure software development leaders have known this for 4+ decades, but for some reason the current crop of leaders don't "get" software. I think a lot of wisdom was destroyed in the dot-com bubble crash, and the survivors, or the ones that grew up in the rubble (Google, Amazon, Netflix, Facebook, Github) were the only ones who maintained these principles. Everyone else got a generic MBA running the show.
I’m really tired of this trope. I’ve spent my whole career on “boring CRUD” and the number of relational db backed apps I’ve seen written by devs who’ve never heard of isolation levels is concerning (including myself for a time).
Coincidentally, as soon as these apps see any scale issues pop up.
At least in my personal case, struggling with renewal at Virgin Broadband, multiple humans wasted probably an hour of everyone's time overall on the phone bouncing me around departments, unable to comprehend my request, trying to upsell and pitch irrelevant services, applying contextually inappropriate talking scripts while never approaching what I was asking them in the first place. Giving up on those brainless meat bags and engaging with their chat bot, I was able to resolve what I needed in 10 minutes.
Claude code is better than a junior programmer by a lot and these guys think it only gets better from there and they have people with decades in the industry to burn through before they have to worry about retraining a new crop.
Instead you should replace senior staff who make way more.
Better learn how to learn as we are not training(or is that paying) you to learn...
Stop getting played.
I remember someone that had a .sig that I loved (Can't remember where. If he's here, kudos!):
> I hate code, and want as little of it in my programs as possible.
I assumed it would happen at some point, but I am relieved that the change in sentiment has started before the bubble pops - maybe this will lesson the economic impact.