AI maximalism is making a lot of assumptions that I think are not a given
* 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.
> The curve of AI improvement will continue at the current pace
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.
I've had quite a few conversations and read many thoughts on the subject of job security in the software industry through the years. New technologies, various crisis and crashes, just age, incoming "hordes" of less prepared developers, or whatever.
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
I think in the future, those who succeed will be equivalent to wayfinders.
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.
does it never? seems to me that people pay me precisely for my knowledge, learned over many years. The knowledge translates into action, sure. But thats like the old parable about a plumber being paid €150 for a 5 minute consult that involves turning a single screw. "i could have turned that screw!" the customer cries, ignoring that yes, they could have. But they didn't know to.
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"
I think a more sane minded customer would not mind paying for the assurance and having someone to blame in case things go wrong, not necessarily because of their domain knowledge.
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.
Knowledge often does not produce competence, especially in the applicable market. I work on the system administration side of things, and I have encountered many output-competent developers that were immeasurably stupid, but very little incompetent ones with tons of cryptic knowledge and intuitive understanding of the systems they worked on.
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
"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"
>> The demand for software most certainly has an upper limit.
> 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.
Exactly and this is true of many things. Much of the world is not zero sum, otherwise we'd have fallen into the "malthusian trap" several productivity booms ago.
I strongly agree with the author replies. I cannot grasp the reasonment of those who underestimate the power of these tools and their growing potential.
We should remember that the outside world care about things that work, not about how good they are inside sadly.
The outside world doesn't even care that things work, they care that it looks like it works long enough. Investors don't care that it's snake oil, as long as they're not left holding the bag.
AI is really good at making things that look like they work.
Well yes. This has been the history of the web. Frontpage generated really crappy code but people still used it to create websites. They didn't care about code quality just how it looked.
This is a sentiment a highly skilled framework knitter could have shared. Investors don't care if those newfangled steam-powered knitting machines produce inferior textiles as long as people buy it.
Parallels to the industrial revolution are apparent. And this is disturbing.
I was just reading comments the other day where people who dragging a company because they apparently used AI for some low level copywriting stuff. No art assets, no code (so far as anyone knows), not actually writing copy but more like "is everything spelled right, does the copy structure flow, have all these points been addressed, etc." Not only that but the only reason anyone even knew is because the company was completely up front and transparent about what they used AI for and what they didn't.
There is a visceral hate in the artistic community toward AI that doesn't really make sense to me tbh.
I think you can't tell the difference until the "art" shows details of something you know well -- a place you've been, out a hobby or sport you do.
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.
I agree, his takes should not be dismissed lightly. I'm not sure about "demand is fixed" though. I feel like software demand has been declared saturated at least a few times.
Exactly, if we look at what projects are on-going now, look at Startups, they are practically solving all the same thing and most of them will be dead soon, we need to finally reach the era where tools to "zeroshot" anything becomes widespread to create new problems, but even by then, we will have an oversupply of tech workers, many will have to convert to a different field, many will not want to be paid based on callcenter type of work which is prompt-as-much-as-you-can, understandably.
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.
Agreed. The limitations of human context window and communication bandwidth restrict the complexity of large-scale software.
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 think people are far too dismissive of just how well-suited programming is to the exact form of LLMs.
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.
Every freelancer that switched to AI feels exactly what happened even if they can't name it.
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.
I agree with all of it, and I think author did a really good job at actually saying what's true, it's almost like developers don't want to hear it.
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.
> If the models (and harnesses) keep getting better at the same pace for the foreseeable years, we are heading to a world where the profession is commoditized to the ground. There's this talk about Jevons Paradox but I disagree. The demand for software most certainly has an upper limit.
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.
Tax I could do to some extend but I once (for laughs) had a go at scripting up Dutch work hour laws because no one could do it in their head. This was so terrifyingly complex that I'm convinced many laws should be rewritten to make it easier to code.
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.
>Agents used to be bad at this kind of stuff in my workplace as well, but newer models + agent-friendly documentation + AGENT.md begging agents to read the fucking docs before coding changed this landscape for us here.
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.
You are correct that your career is changing, but it's not like AI is going to go away.
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.
"I'm finding LLMs also competent at explaining and giving advice on other domain stuff I'm totally new to, which I have cross-checked with Legal/Product Managers and is usually right."
"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".
I don't entirely disagree, but as with many other posts on this topic…
> 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.
> > This anonymous article is likely more FUD from the AI industry. "Just give up,you can't beat the machine. Please go quietly, we want to take your place and it's easier for everybody if you don't resist because you believe it's pointless"
> > 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.
> This anonymous article is likely more FUD from the AI industry.
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.
> The models will learn good engineering principles at some point.
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.
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.
LLMs are an ideological tool for the capitalist class to finally replace their dependency on labor and its pesky demands like sick leave and a living wage. A way for capital to finally become completely self-reproducing, for power structures to cement themselves and never be challenged again.
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.
> 1) communal ownership of the means of production, e.g. of compute
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?
I attribute the excitement of the VC and CEO cast to the same underlying motive, but I think there are at least several other ways all this could play out:
- 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
AI maximalism is making a lot of assumptions that I think are not a given
* 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.
> * The curve of AI improvement will continue at the current pace
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).
> The cost for the same quality of output is going to drop at least 10x over the next 18-24 months.
How do you know that?
In 2026 the prices have been spiking. It now costs orders of magnitude more than it did in November.
2 replies →
> The curve of AI improvement will continue at the current pace
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.
I've had quite a few conversations and read many thoughts on the subject of job security in the software industry through the years. New technologies, various crisis and crashes, just age, incoming "hordes" of less prepared developers, or whatever.
If I had to highlight the one thing all those conversations had in common it would be precisely this:
And it never does.
I think in the future, those who succeed will be equivalent to wayfinders.
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.
> I think in the future, those who succeed will be equivalent to wayfinders.
In the future, those who succeed will be the owners of capital.
2 replies →
In a perfect world, yes. However, the current tech world is akin to a flea market. Those who shout out more stand out more.
4 replies →
does it never? seems to me that people pay me precisely for my knowledge, learned over many years. The knowledge translates into action, sure. But thats like the old parable about a plumber being paid €150 for a 5 minute consult that involves turning a single screw. "i could have turned that screw!" the customer cries, ignoring that yes, they could have. But they didn't know to.
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"
There's a good chance the apprentice plumber could've fixed it just as quickly. That's where we are now.
3 replies →
I think a more sane minded customer would not mind paying for the assurance and having someone to blame in case things go wrong, not necessarily because of their domain knowledge.
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.
Some knowledge does set you apart - the ability to ship things, people pay for.
Not producing holy code in the academic best language.
Ability can't really be compared to knowledge... e.g. you might lose the ability to play the piano, yet retain the knowledge about how to
1 reply →
Knowledge often does not produce competence, especially in the applicable market. I work on the system administration side of things, and I have encountered many output-competent developers that were immeasurably stupid, but very little incompetent ones with tons of cryptic knowledge and intuitive understanding of the systems they worked on.
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
This is the old China fallacy.
"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"
> The demand for software most certainly has an upper limit.
No, it does not. There is no ceiling for complexity.
>> The demand for software most certainly has an upper limit.
> 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.
Exactly and this is true of many things. Much of the world is not zero sum, otherwise we'd have fallen into the "malthusian trap" several productivity booms ago.
And when the required complexity of software to do the task gets high enough, you assign an agent to do the task instead.
Entropy makes sure that you can't scale systems into infinite completely.
I strongly agree with the author replies. I cannot grasp the reasonment of those who underestimate the power of these tools and their growing potential. We should remember that the outside world care about things that work, not about how good they are inside sadly.
The outside world doesn't even care that things work, they care that it looks like it works long enough. Investors don't care that it's snake oil, as long as they're not left holding the bag.
AI is really good at making things that look like they work.
This is a steelman of your argument.
Well yes. This has been the history of the web. Frontpage generated really crappy code but people still used it to create websites. They didn't care about code quality just how it looked.
2 replies →
This is a sentiment a highly skilled framework knitter could have shared. Investors don't care if those newfangled steam-powered knitting machines produce inferior textiles as long as people buy it.
Parallels to the industrial revolution are apparent. And this is disturbing.
> We should remember that the outside world care about things that work, not about how good they are inside sadly.
Until they go wrong because they are not good inside.
This is true. I have artist friends that are boycotting any company using AI art for their flyers/ads.
I looked at some examples and couldn't tell the difference.
I was just reading comments the other day where people who dragging a company because they apparently used AI for some low level copywriting stuff. No art assets, no code (so far as anyone knows), not actually writing copy but more like "is everything spelled right, does the copy structure flow, have all these points been addressed, etc." Not only that but the only reason anyone even knew is because the company was completely up front and transparent about what they used AI for and what they didn't.
There is a visceral hate in the artistic community toward AI that doesn't really make sense to me tbh.
2 replies →
I think you can't tell the difference until the "art" shows details of something you know well -- a place you've been, out a hobby or sport you do.
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.
The outside world itself will stop working if we replace labor with LLMs.
Mass unemployment equals riots equals an end to the status quo.
The next big revolution probably involves burning down datacenters.
This doesn't seem at all related to the above comment - or anything, for that matter. Nobody is suggesting we "replace labor" with LLMs.
4 replies →
Your argument boils down to: it’s different this time.
Isn't that a perfectly fair argument if you can articulate why?
ok, so?
[dead]
Did you read it?
The argument boils down to: this is exactly the same as other times. And provides multiple examples.
He literally did not provided multiple examples of such a thing.
Yes; that is literally the opposite of what this article does.
I agree, his takes should not be dismissed lightly. I'm not sure about "demand is fixed" though. I feel like software demand has been declared saturated at least a few times.
I have been making software professionally for 25 years and in all that time i have never run into the problem that we have run out of things to do.
Exactly, if we look at what projects are on-going now, look at Startups, they are practically solving all the same thing and most of them will be dead soon, we need to finally reach the era where tools to "zeroshot" anything becomes widespread to create new problems, but even by then, we will have an oversupply of tech workers, many will have to convert to a different field, many will not want to be paid based on callcenter type of work which is prompt-as-much-as-you-can, understandably.
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.
Do not use past events to predict the future, or you risk end up becoming a turkey: https://peteweishaupt.medium.com/talebs-tu-e406eb8859a8
> I feel like software demand has been declared saturated at least a few times.
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.
Agreed. The limitations of human context window and communication bandwidth restrict the complexity of large-scale software.
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.
Whenever someone complaints about LLMs eroding their career, I advise them to read The Profession by Isaac Asimov.
TLDR: there will be less programmers and they will be better on average.
I think people are far too dismissive of just how well-suited programming is to the exact form of LLMs.
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.
Every freelancer that switched to AI feels exactly what happened even if they can't name it.
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.
I agree with all of it, and I think author did a really good job at actually saying what's true, it's almost like developers don't want to hear it.
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.
Just 300 a day? That's only one ticket every 1.5 minutes. I hope in a year we can fix an issue under 30 seconds with ZERO supervision.
We will, most work can be parallelized, the same way as developers are able to work together on large codebases, tools can as well.
May I ask what are some of methods you’re using for this level of productivity?
> If the models (and harnesses) keep getting better at the same pace for the foreseeable years, we are heading to a world where the profession is commoditized to the ground. There's this talk about Jevons Paradox but I disagree. The demand for software most certainly has an upper limit.
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.
Tax I could do to some extend but I once (for laughs) had a go at scripting up Dutch work hour laws because no one could do it in their head. This was so terrifyingly complex that I'm convinced many laws should be rewritten to make it easier to code.
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.
>Agents used to be bad at this kind of stuff in my workplace as well, but newer models + agent-friendly documentation + AGENT.md begging agents to read the fucking docs before coding changed this landscape for us here.
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.
You are correct that your career is changing, but it's not like AI is going to go away.
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.
"I'm finding LLMs also competent at explaining and giving advice on other domain stuff I'm totally new to, which I have cross-checked with Legal/Product Managers and is usually right."
"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".
I think technically it's referring to the advice, which is in the singular.
"These AIs are usually right about things I don't know anything about" sounds like the textbook example of risky thinking though.
Maybe it's the advice that's usually right.
I don't entirely disagree, but as with many other posts on this topic…
> 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
> > This anonymous article is likely more FUD from the AI industry. "Just give up,you can't beat the machine. Please go quietly, we want to take your place and it's easier for everybody if you don't resist because you believe it's pointless"
> > 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.
> This anonymous article is likely more FUD from the AI industry.
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.
Seriously. Something is wrong with these people. It’s completely unacceptable.
> Domain knowledge can be learnt much quicker than how to apply good engineering principles.
This is a particularly ignorant thing to say.
It's classic https://imgs.xkcd.com/comics/physicists.png
(Also, both might be out of reach of the current AI architectures)
> The models will learn good engineering principles at some point.
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.
Some food is mass-produced in factories.
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.
LLMs are an ideological tool for the capitalist class to finally replace their dependency on labor and its pesky demands like sick leave and a living wage. A way for capital to finally become completely self-reproducing, for power structures to cement themselves and never be challenged again.
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.
> 1) communal ownership of the means of production, e.g. of compute
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 boosters imagine they'll be annointed and rewarded by their new overlords.
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.
IMHO that is the most likely of the many dystopian robots-replace-humans scenario:
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.
I attribute the excitement of the VC and CEO cast to the same underlying motive, but I think there are at least several other ways all this could play out:
- 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.
> It's time to make your choice
Clearly you feel you've made yours, so what are you doing differently now to what you did before?
You should read Nick Land, you've only went half-way with the argument.
The capitalist class doesn't control Capital, Capital controls the capitalist class.
[flagged]
[flagged]