> Work that introduces new methods, highly creative ideas, or solutions that have not been used or experienced before. More generally, an approach that introduces an innovative strategy to solve a complex problem.
Something that I've been thinking about for the past year or so is coming to grips with the fact that the vast majority (anecdote) of software engineering work is not novel (and maybe that's okay). Few opportunities lend themselves to doing truly novel work. Other than infrastructure work and highly specialized software, pause and ask yourself when you last encountered software were you said "how the hell did they do that?" or "damn, that's nice" (for me, the most recent was Ghostty). I think much of the angst that people have when they fear for their job is coming to the realization that LLMs can do most of the "standard" work that a lot of highly compensated individuals currently do. We've built livelihoods around this and the threat of that coming to an end is genuinely frightening.
Amd do it better in most cases imo. Which is also hard to come to terms with, because there is a good bit of elitism/entitlement going around. The idea that a SWE is working at a higher level, which is beyond the reach of mere mortals, so therefore the high compensation is justified. Meanwhile everyone is, for the most part, doing some slight variation of the same thing as you suggested.
After starting out working minimum wage jobs I've always thought that the work gets easier and easier from there. Compensation and hard work are negativity correlated.
If I had to highlight the one thing all those conversations had in common it would be precisely this:
I thought that having this knowledge would set me apart
People who _can_ see the wood for the trees, and are able to understand multiple (sometimes conflicting) requirements and work out a way through that solves the problems that arise, for all involved parties.
An understanding of domain, the ability to communicate effectively and a mind that can think laterally, will all be vital.
In the future, those who succeed will be the owners of capital.
The results will hopefully be a lot more tangible.
I think perhaps the problem is instead "I thought that having this knowledge would set me apart, forever, without me having to learn anything else"
If your argument is that the customer themselves could use an AI or whatever to learn plumbing, that was always an option (libraries, google, youtube). They pay you so they don't have to worry about flooding their house (or at least have someone else to blame).
They might be able to "one shot" simple fixes that you might previously have assigned to an apprentice, but believe me, AIs are not about to start doing complex things for the layman that actually required seniors previously in either programming or plumbing, because very few of those things were just "type better into a computer". (build trust, speak confidently, know what doesn't work, take responsibility, test without breaking systems, communicate and work together with other professionals, have opinions)
it's oft debated, but I do fall on the side of "you should still know maths even in the age of the calculator/matlab/llms". I have found productive employment, and indeed tickets to speak to the big boys in their gilded palaces many times because graphs and charts are their favorite toys and knowing maths got me there. They have always been able to make things with excel, with matlab etc. Often they actually can make charts themselves, but they don't care to become experts in what data is important and what isn't.
The LLM isn't yet good enough to tell you what data matters. People act like LLMs are magical gods that do everything, but it is but another tool. It has limitations, just as it has strengths. It is not ultimately convincing, it is not infallible, and experts will keep finding edge cases all the damn time. Anyone working with them every day knows this, and you need to know it too.
I could theoretically learn everything about plumbing but would still rather call a professional for the peace of mind that it was done "correctly" and it the process goes wrong, I would have an instant fix instead of trying to go back and educating myself on plumbing more.
Could you consider that as part of knowledge? Yeah and also no. Because the knowledge can be copied and put into a LLM but legally a LLM cannot sign off on things like NDAs or take accountability like a human has to in these roles.
We can argue about imagined future progress, but I don't see that getting much better, given that the literature doesn't often do that, and how often experts in one scenario end up being poorly suited given another set of facts.
Not producing holy code in the academic best language.
"Oh, we'll just ship production to China, and do the design and marketing in US, this is where the real value is anyway, China will never be able to do design and marketing as well as we do".
Literally same thing:
"Oh, we'll just let LLMs code, and we'll just do Taste. LLMs will never be able to do Taste"
It seems to me that knowledge doesn't always imply competence, but the lack of knowledge often very well explains incompetence. And, since the LLM is replacing the competence part without imprinting any knowledge on the one that wields it, it generates a lot of competent imbeciles that pass interviews and appear as though they not only do things, but know things as well. And once you reach that critical mass, sheeeeesh
* The curve of AI improvement will continue at the current pace
* AI companies will have the capital continue to expand infrastructure
* there will be some kind of functioning economy if all knowledge workers are replaced
There are strong headwinds to all three of these.
Hey it may come to pass but it’s very speculative at this point. I see a lot of tech people simply overlaying the progress curve of previous tech booms which is reductive.
Frontier AI is already good enough to be very useful for engineering. It's too costly for many places where it could be useful today.
The cost for the same quality of output is going to drop at least 10x over the next 18-24 months.
And likely again in the following 18-24 months.
At the same time, the cost per watt is going to down ~25%, and at the same time speed will increase (also valuable since time is money).
How do you know that?
In 2026 the prices have been spiking. It now costs orders of magnitude more than it did in November.
April of last year you'd get 1431 ELO[0] from o3-2025-04-16 for $8.00 per million output tokens. April of this year you can get 1436 ELO from deepseek-v4-flash for $0.2 per million output tokens.
[0]: https://huggingface.co/spaces/lmarena-ai/arena-leaderboard
Historic trends, every 18 months, performance for the same level of quality has gone down 90%.
See: https://www.reddit.com/r/LocalLLaMA/comments/1gpr2p4/llms_co...
And Chart 13 here: https://www.rdworldonline.com/ais-great-compression-20-chart...
And here: https://epoch.ai/data-insights/llm-inference-price-trends
The technology already exists now on the algorithmic front for the next 10x drop between everyone adopting DeepSeek's MLA, MoE (mostly already done), Medusa (a better version of Google's speculative decoding), Kimi's Attn Residuals, and Mimo's Sliding Window Attn, and (possibly) Microsoft's 1.58b (this may be a nothing burger).
Historically, algorithmic gains are only ~30% of the pie, but there's enough out there to get to 10x, with just what's available already. The other ~70% of the pie is better training data (often synthetic) and distilling frontier knowledge. There's no sign we are tapped out on that front.
> In 2026 the prices have been spiking.
That's not for the SAME level of output...
Regarding AI companies having capital to expand infrastructure; this is largely irrelevant. The cat is out of the bag, and you can already make serious gains by finetuning to local problems on a desktop machine. There is enough hardware out there to run these things en masse; it's more a question of power. Regardless, this stuff will always keep progressing, regardless of who is doing it.
Regarding the economy, it may be largely irrelevant if we, the people, don't do something very soon. The wheel keeps spinning as long as there are productive workers; it's just that those workers are being replaced by machines. The last year has increasingly demonstrated that you don't need normal people to buy your stuff to remain afloat. You can just keep selling amongst your rich friends while the masses starve, as long as _something_ is still producing what the wealthy want, and enough systems are in place to protect them.
I guess this is trivially true if you say "maximalism" (hell, the maximalists think it will speed up as the AI becomes a super-AI-researcher), but as long as the rate of change is positive and not miniscule, it's hard to predict what 2035 looks like in software development.
These things are very hard to quantify, but making the progress that happened from Jan 2025-December 2025 repeat twice in 10 years would be enough for me to say I couldn't predict the day-to-day of a software engineer in 2035.
No, it does not. There is no ceiling for complexity.
> No, it does not. There is no ceiling for complexity.
There's an upper limit on everything. Maybe there's no ceiling on incidental complexity for s/ware development, but there sure as shit a ceiling on the essential complexity.
Tools/improvements have rarely been negative in such a massive way except rare instances, and even then society moved on and past those tools to bigger & better things.
How many people today seriously consider agriculture as a career prospect but almost all humans who lived in the last 2000 years worked as peasant labor on a farm. We are thriving in comparison to that period of time.
You: you're saying "it's different this time."
I don't know. It looks like AI really rots people's brains. As if that they just shut down their minds when they see an anything AI-related. Imagine if this article were about anything else, like:
Article: the stock bubble is going to burst because...
Comment: your argument boils down to "the stock bubble is going to burst."
It'd be so stupid. But somehow when it comes to AI this kind of weird comment is tolerated even celebrated.
The argument boils down to: this is exactly the same as other times. And provides multiple examples.
AI is really good at making things that look like they work.
This is a steelman of your argument.
Parallels to the industrial revolution are apparent. And this is disturbing.
Until they go wrong because they are not good inside.
We’re in some weird stage of capitalism where everything is a grift and nobody really cares anymore.
I looked at some examples and couldn't tell the difference.
There is a visceral hate in the artistic community toward AI that doesn't really make sense to me tbh.
But pretty soon after that it's "Why am I paying a transcriptionist $3/minute when I can just have the machine auto-transcribe it and then my admin assistant can just scan it for mistakes."
Even if there still IS a quality difference between great writers and AI product, "good enough" is good enough for most customers, especially if you have to pay professional rates to get better.
Really? Have you seen how the CEOs marketed it and talked about people in that community? Artists hate it, because they listened to what AI community and leadership were openly saying.
The weirdest thing on this all is how people find the hate puzzling considering initial rhetoric coming from the industry itself. And current rhetoric for that matter.
"Why do you guys hate AI so much? All I did was tell you it's so great that it makes your skills worthless and how glad I am that I won't need people like you around in the future to make art and designs. What's wrong with that?"
I'm thinking of this awful slop "art" I saw on Wayfair yesterday. As a surfer, it's hilarious. That's not how you stand on a board. It's not even a board. And the wave is terrible-- nobody wants to surf shorebreak like that! https://www.wayfair.com/decor-pillows/pdp/design-art-4-hawai...
I guess it could be a useful signal-- if you meet someone and they have it up in their home, you know they don't surf.
More generally, I think anything AI produces that's dense with factual details is inherently trash.
Mass unemployment equals riots equals an end to the status quo.
I take it you haven't been listening to what the guys at the AI labs have been saying?
Plus that's what the whole article is about. I'm not sure how you could've missed that?
> Take copywriting. It was a profession that took years to master and paid well. This changed slowly as more professionals joined the market, even after the demand spike driven by ecommerce and adtech. Now, LLMs have destroyed the job for the vast majority of professionals.
It's quite hard to predict what will happen, but in a few years, I bet the unemployment rate of tech workers will be really high, we can just look at how many jobs are currently already replaceable but the owner of it is just lagging in the implementation of automation, it's probably already the large majority of tech jobs.
It's never been declared saturated, with one exception in the six months following the dot-com crash.
I've been in the industry since the mid-90s. I have not seen automation with the potential to automate away everything for the average office worker.
LLM will have an extremely large context window and extremely high communication bandwidth in the future. Therefore, even more complex large-scale software will emerge.
I feel that OP has reach that point because he went out of the basic tooling like Claude Code (at least in its default state) and embrace multi-model, automatic reviewing, fuse, loops and so-on, when it's done right, well, failure rate to solve issues is <1%, this is exactly why you arrive to that kind of depressing thoughts afterward and it's spot-on.
Many people will disagree because they are still at the vibe coding stage, not "as much as I can prompt will be automatically done stage". Claude Code imo is deliberately not implementing the best ways for users to work, they have recently implemented Workflows but that's almost a year late, many companies are doing this since always and that's just part of basic tooling nowadays.
People talk about models and benchmarks score while genuinely I'm baffled because they seem to ignore that that same benchmark can reach 99% by levering tooling intelligently, we don't really need better models (at least for coding), we just need adoption of proper methods. The day developers will discover that they are already able to solve 300 issues in a single day with ZERO supervision in complex Rust codebases, I'm sure they'll change their mind.
Our bottleneck in our team is currently just having the mental bandwidth to type as much as possible, it's kinda sad, it is becoming all absurd.
If you are still watching the output of the model for coding tasks, I bet you haven't challenged your own methodologies, yet.
We became for AI what our clients were for us. Some hate it, some love it.
To feel safe in life our clients needed to have an actual business. Now when we are the clients of our AI we are scared, because now we need to have an actual viable business. Economic machine that works. Because the old model of just selling our time and effort to a client no longer works, when we are the clients.
Extremely formal syntax, limited ambiguity, simple verifiable testing procedures, and colossal well-documented training sets.
I don't yet buy that the successes of coding agents will apply nearly as well to other professions. "Correct more often than not when asked a random accounting question" really isn't any indication to me that they'll get there.
Wouldn't that be true for humans as well? If you have documentation explaining a rule and you read it, you may not need to reach out to coworkers.
Otherwise I think the author's concerns are 100% valid.
TLDR: there will be less programmers and they will be better on average.
This entire section is backwards to me.
The current state of a lot of different domains I've been in is that they tend to center around 2-3 major, generic products that all get retrofitted to fit those smaller/medium-sized businesses. Now that the economics have shifted, it makes sense for those businesses to bring on software devs to build software tailored to their problem specifically.
And you can't compare copyrighting. It's a totally different field, with different goals and different time tables.
The problem looks something like (not a real example): Type Z hours maximum A per day, B per week, C per month, D per year. E more hours than A is allowed every F weeks but no more than G per month and H per year. More than B is allowed... etc Minimum rest hours I per day, J per week, K per two weeks, L per month. More is allowed every 7.5 days unless it is full moon and maximum hours per day were exceeded at least 3 times in the last 82 days except from solar eclipses or if the Kings is married 12.5 years or if the employee gave birth in the last 472 hours.
My employer has software to make the schedules. It cant tell where shifting around shifts is possible but you can try do it and it will tell you why it isn't possible.
I was hoping to calculate if multiple shifts can be shifted around to facilitate someones day off. Sometimes it just cant be made to work but if people are willing and there is a hole you end up doing it anyway. (I've done a triple shift once because the coworker wanted to bring his wife to the hospital.) Employees earn undocumented days off... and then you end up with multiple schedules, the real one and the official one. Possibly extra copies depending on who knows what is really going on. This cant be the way...
Better just have modern laws that make sense in code.
In the 1990s when crypto went to court. It was determined that really anything coming from AI is protected speech. Very few exceptions, AI cant export a few things.
So you're never seeing AI go away, which means you need to transition/adapt.
"Usually" is the keyword. Until it becomes "always" (counterintuitive for heuristic systems) or "almost always" some human experts will (/may?) be needed to babysit.
P.S. "_are_ usually right" since they are "LLMs". Methinks running the response through an LLM could've made it more "right".
"These AIs are usually right about things I don't know anything about" sounds like the textbook example of risky thinking though.
> > So blog with single post hyping LLMs. Oh and the domain name "human-in-the-loop". Call me suspicious.
> If after reading what I just said in the reply above you still think I'm an "AI shill" or "lab shill", there's nothing I can do for you.
Yes there isn’t. Because they look indistinguishable.
Replacement Inevitability with a human face, along with all the human concern; “I am part of it and it scares me.”
> Yeah, that's what I'm doing right now. I'm one of the engineers who's constantly committing to improve our agentic tooling, I use different models to do adversarial code reviews, I keep a toolbelt of skills and prompts, etc. I have effectively become the so-called "AI-native engineer" (gosh, I hate that term).
Some CEO gloating about replacing all-knowledge-work gets skepticism, eye-rolls and resentment. Someone in the trenches having human feelings about it generates both sympathetic and ecocentric fear.
---
And maybe autor intent does not matter? The original submission was massively “popular”. It served its purpose.
Literally today I got like 4 AI ads literally mocking "old people still using excel", trying shame and insecure people into some AI whatever product.
This is literally the first technology that is trying to scare and mock me into using it. All it actually does is that I am growing to hate it, honestly along with tech industry itself. Which I used to like.
This is a particularly ignorant thing to say.
(Also, both might be out of reach of the current AI architectures)
This is just silly. It's fairly clear that the current design (by which I mean the entire concept of the deep neural network) has its limits and that they just aren't that good. We're seeing lots of other AI and software engineering brought to bear, but there's nothing 'inevitable' that means this is close.
"at some point" is so vague as to be irrelevant. Fusion might be the dominant source of electricity "at some point". Equally, AI knowing good principles could be 30 years away.
Don't assume that hard intellectual challenges are solvable on faith. Look at what's currently possible.
AI has always been a field where https://imgs.xkcd.com/comics/tasks.png applies heavily.
Maybe, but people have been saying deep learning is about to hit a wall since 2012, and many reasonable-sounding "machines fundamentally can't do X" have since fallen.
Feels like we're standing on a roof with floodwater up to our ankles - maybe it stops rising now, but we didn't foresee it getting anywhere near this high in the first place.
I do agree that progress will probably be more slow/gradual than others seem to predict, no "hard takeoff", but even being decades away is still relevant to someone starting a career in software development.
That's why the VC and CEO crowd are so excited about it, while the average population is hesitant at best.
There is no addressing this issue without developing class consciousness.
The only two ways out of this are 1) communal ownership of the means of production, e.g. of compute or 2) technofeudalism with cleansing of the now unneeded, unproductive new underclass that only takes up resources our overlords want for themselves.
Which version do you want to see realized? It's time to make your choice.
Note that I am not shitting on the idea of option 1 at all, in fact personally I would very much like to see it succeed. I just think this is more of a global issue than a local one.
> Note that I am not shitting on the idea of option 1 at all, in fact personally I would very much like to see it succeed. I just think this is more of a global issue than a local one.
That’s why socialists argue for international revolution.
That's why they're obsessed (to the point of psychosis) with "mastering" the new technique.
That's why they're all building a "harness".
What they don't realize is that the ironworker still ends up in chains.
As every communist, you forget about economics of such a system. How would you prevent concentration of capital in this system? Planned economy? Planned by whom?
The AI-enhanced become more and more AI-integrated and internally AI-fused and they don't even realize they eventually are not humans at all.
The non-AI underclass just hasn't got enough access to resources to survive long term and dies out with a whimper.
- the Cul-de-Sac: AI progress flattens as scaling data and compute, RL and algorithmic improvements hit diminishing returns.
- democratization: LLMs decentralize, mirroring the shift from mainframes to personal computers.
- AI creates new jobs and thus new dependencies for the capitalist class
- Any combination of the above.
Clearly you feel you've made yours, so what are you doing differently now to what you did before?
The capitalist class doesn't control Capital, Capital controls the capitalist class.
It tastes bad, and poisons you slowly.
Some (less) food is produced on farms and kitchens.
It tastes good, and keeps you healthy.
I don't really care who/what wrote the code. I don't even really care about the code at all. What I care about is the end product.
The problem is not "code quality" the problem is that billionaire sociopaths have removed human judgement (and human morality) from the dev loop. This started long before AI.
Coders are hyperfocused on style and missing the substance. We are entering a world where rich bastards can produce evil software without any checks whatsoever.
At least when humans were required to write the code, they had to find and retain unscrupulous humans. Now they're completely unfettered, and we're soon going to learn the precise shape of the digital prisons they're constructing.
> They will come for finance, biology, law, marketing, all knowledge work. That's their stated goal and they're already teasing it with "ChatGPT for Health" and similar launches. They're working on "harnesses" for other fields, it's just a matter of time before we have "Claude Finance Analyst" or something.
…
> Beg to disagree. The models will learn good engineering principles at some point.
…
> Stop and think, don't try to predict the future using (bad) past examples.
Don't try to prediction the future based on the past.
Also, here is my doomsday prediction.
Thats kind of ironic.
Heres a more thoughtful take: everything is an s curve.
Things start out fast, then they slow down.
It happens in learning, in tech, in literally everything.
The question (unanswered) is where we are in that curve.
Will they get better? Yes.
A lot better? A bit better? /shrug
