The post misses a few things:
- The ai bubble is currently being subsidized to an unimaginable degree. If you were to actually pay true cost for your token usage, you wouldn’t be saving that much over an engineer’s salary. Probably even worse once AI companies start to extract a real profit. 95% of companies diving into agentic labor will be in for a rude awakening when they balance next year’s budget.
- The cost to keep ai useful in its current form has a high floor. Unless you keep up with expensive training, your models will drift. You can only scale your model intelligence with more hardware (roughly). In two years, Claude opus 4.8 will still be bloating context to learn about the latest cloud platforms and libraries. A human engineer will get those passively at no cost to the company.
- As the complexity of the task grows the complexity of the ai babysitter must match it. Even if Ai stays cost effective, companies can now save money by spinning up bespoke in-house software to cut out vendors (think observability platforms, task tracking, product design, marketing systems, etc…). No matter how many adversarial reviews and sub agents you spin up, an Ai can’t grasp the full context of your company and it’s shifting priorities. The software engineer role transitions to a pseudo-sysadmin + product architect.
C-suites don’t want know about software and don’t care about non functional requirements (security, availability, audit ability, etc…). They just want to wave a magic wand and have a product appear, which is what Ai provides the illusion of. That’s why all current Ai software is garbage, but the smarter companies will catch on
The only thing to do is ride the wave of the new technology. You either do that or drown. I’ve written a literal book’s worth of text on how AI should think and behave when writing software and it straight up cannot write code as well as me when left to its own devices, even with all my guidance. LLMs, no matter the size, are structurally incapable of holding all the context needed to write good code in complex systems because they cannot compress complex topics into ideas they’ve never encountered and reason about the compressed state with their limited reasoning capacity (otherwise known as “learning”).
Yeah, I find it difficult, too, especially since management hasn’t caught onto this yet and still wants me to specialize.
And of course, the answer is that I should specialize in AI, because there’s currently a lot of new development happening there. But that knowledge is also getting obsolete by the minute, with ever more tools coming out and then again other tools that operate those tools for you.The one thing I hold onto, is that no matter how the situation evolves, the basic job requirements for software engineering, i.e. being smart and being able to learn quickly, will always be an advantage.
I don’t think it’s possible to hold onto the confort zone from before, even if the industry implodes from the AI costs becoming transparent. But yeah, I do think we’ll land on our feet in one way or another.I’m not sure about specializing, but there’s a lot of stuff you can learn in the AI field that was useful before the current hype bubble and will remain useful going forward. Traditional ML is huge and doesn’t move quite as fast as e.g. LLM applications.
Oh yeah, for sure, but they do mean generative AI in this case, because it’s the hype du jour.
Nobody needs A or B-grade codebases anymore because they’re being made for LLMs, not for humans to read.
That will come back to bite them in the arse, mark my words.
It’s also patently false.
A good chunk of good patterns are to make sure humans understand it sure. But a good chunk of patterns exist to make individual components reusable and make sure you’re encapsulating requirements and testing them correctly.
A lot of LLMs take the “easy” way out and duplicate code, suppress listing, etc to make a prompt work. It works at that point in time but when you suddenly have a bunch of spaghetti and repeated code littered across multiple services suddenly making changes without causing massive regressions becomes a headache.
Companies are going to pay for this mess in several months as token prices go up and the codebase is a massive pile of slop.
That’s going to be the bubble. When AI has to be able to actually pay for itself, no one is going to be able to afford it, and if you happen to be one of the companies that went all in any used AI to build your codebase and fire not devs and front line workers, you’re going to be the hardest hit. Possibly the only hope is that they saved enough from partial and didn’t pass any savings on to the customer (because of course they wouldn’t) that they can almost survive the actual unsubsidized token costs. But then you will be in direct competition with everyone else who can write a prompt with likely literally no differentiator outside of maybe name recognition in an industry.
If the only problem is that your code is slop and nobody can work with it without AI, then it’s probably not that bad. Text models I can run locally on my five year old Macbook are maybe a year behind in terms of coding assistance. So AI for coding is probably never going away. The worst case for someone in this scenario is just that it gets a bit slower and dumber and that they have to hire more engineers again. It’ll suck but I think it’s survivable. Someone would have to make a new Stackoverflow though if we’re going to google stuff again.
Now if you integrated multiple AI services into all your business workflows and into the products you sell, on the other hand, that might be a different story. In a way the risk is the same as with cloud providers. You get locked into a stack and then your product literally dies if the provider decides you’re not paying enough, because you have no feasible way out. Tbh I would much prefer working at a post-bubble era software company fixing the codebase to working at a random company now extracting their IT from a hyperscale cloud. But in reality, most companies that bet on AI are in this scenario. Nobody only installed Claude and called it a day.
I think there’s a fundamental misunderstanding here.
All of the qualities that make a codebase easier to read, maintain, and consume by humans do the same things for LLMs.
A codebase designed for humans is a codebase that is designed for LLMs. It’s just that most teams don’t even know how to design a codebase for humans. And those same teams just kind of accept LLM and Agent Slop as “Designed for LLMs”. When it most definitely is not.
- Patternization
- Structural consistency
- Naming conventions
- Style Opinionation
- Organizational conventions
- Safety and Quality Standards
- …etc
All these things matter just as much for humans as they do for LLMs. And like I said previously, most human developers don’t understand these things and do not optimize for them anyways. Which means that most human developers are ill-equipped to create codebases that are not degrading rapidly under the use of agents.
This is a bit of a rant of mine… because I’ve spent the last decade learning how to optimize software engineering to best fit the needs of humans. Now that LLMs are crashing onto the scene, teams that already were writing slop by hand can now write slop at twenty times the rate. And then seem to think that all the things that make for good software no longer apply to them
Some of these are arguably much more important for LLMs because of limited context sizes. The more of the code in the context window follows good practice, the more likely the LLM is to align with it. Any nonsense in the context window will multiply and beat that one document with the style guide that the LLM might not even see.
ya, but in the mean time i still need to fund my landlords retirement/pay rent
That will come back to bite them in the arse
Nothing ever bites them in the arse. It always runs downhill.
_
Y’all Software engineers went out to become programmers and made upwards of 300k a year as entry level job. Now you have a tough time like the rest of the working population? What 70k a year not good enough anymore? Get fucked.
it’s coming for whatever job you got too
Lmao the vast majority of us never come close to 300k much less start off there. Some of you are totally delusional about tech salaries and believe everything you read.
I feel you, but most of us don’t work for silicon valley, in most countries the salaries are average to slightly above average, and being fired or forced to work with tools you don’t want sucks, which is also a workers issue
+1 for a nuanced take. My America-as-default bias is strong and you rightfully call me out on that.
I also notice this isn’t the shitposting sub… ¯\_(ツ)_/¯
Even in the US I started at 65k which was normal back in the early 2010s for a junior dev. I didn’t reach 6 figures until I’d been in the industry for 5 or 6 years, after changing jobs and then getting promoted. It’s not that different from other stem jobs.
It’s a good job and I can’t complain much, but the Ivy League brats who bragged about their ridiculous starting salaries created a false impression of the industry. They were always a tiny minority and they’re basically non-existent now.
I think the author is mostly right about the current state of AI, but his future predictions (or worries) are based on a false premise: that the massive LLMs will keep improving in the future.
As far as I have seen the improvements have clearly slowed down, while the energy consumption is rising linearly (or worse). It’s like the energy (money) vs. performance graph is logarithmic, and the companies are expending double the energy to get a 10% improvement. Something like that is not sustainable, and the money seems to indicate so.
I really think that LLMs are a dead-end for AI. A really useful dead-end, once the bubble pops and with time, we get a useful working model for them, probably based mostly on local LLMs, maybe using specialized training data.
Energy efficiency has improved by orders of magnitude - leading to much higher energy use. It’s the Jevons paradox and it’s as old as coal-gas lighting. Last year some guy recreated GPT2 for twenty bucks. Corpus to model in one hour. OpenAI never said how much the original cost, but there was at least one comma.
But yeah, LLMs are fundamentally limited, because ‘what’s the next word’ shouldn’t work. The fact it’s accidentally this flexible and powerful, even with its many infamous fuckups, is a reminder that neural networks in general will permanently alter computing. Models trained on supercomputers can run on any potato. Any problem with good examples can be addressed, without first being solved.
LLMs are fundamentally limited, because ‘what’s the next word’ shouldn’t work.
Yes, you’re right. However, for fear of coming off as an AI sycophant (I’ve yet to sacrifice my brain at the altar of our future AI overlords), LLMs aren’t the whole picture. Plenty of research is dedicated to essentially combining the best of each class of AI algorithms into a composite model of intelligence. For instance, “Neuro-Symbolic AI” is really just the result of giving an LLM (good at translation, search, synthesis, bad at symbolic reasoning) a symbolic inference engine like Prolog (good at symbolic reasoning, no native ability for translation/search/synthesis). I’ve been coding for over 20 years, and I’m impressed at its results for software development.
This all is reminiscent of Moore’s Law; even though we keep running into the physical limits of CPU clock speeds, transistor size, etc. we keep finding clever ways to work around those limits.
Of course I’m not saying we should; these models are, after all, models of intelligence, not wisdom.
Edit: fix apostrophe splice
Being reasonable about the tech is kind of a pain, here. At least we only got one flavor of campist, so anyone seeing outright “boosters” is jumping at shadows. I just think the chatbot that can code is neat. Maybe we can do stuff with it. Maybe it doesn’t need to spy on everyone forever.
We accidentally invented p-zombies and they’re already more intelligent than script kiddies. Once the grifters move on we can see what that’s good for.
Agreed. The bubble has to pop eventually… Or not, and we really are marching towards our own obsolescence as a species.
I mean, AGI is inevitable, but it’s never gonna come from these dinguses. They can’t even look past LLMs far enough to pursue text diffusion.
To imagine we cannot possibly build a mind, or that it cannot possibly improve that same effort, is baffling. It changes the shape of the universe.
Just because it’s possible does not mean it’s inevitable. It’s incredibly optimistic to think that we can get our shit together enough to pull it off before we destroy all our productive capacity through hubris.
Nothing’s inevitable. And as for “building a mind”, while it depends on precisely what you mean by “mind”, it’s totally possible that only a biological brain can produce minds as we understand the word “mind”. Building AGI doesn’t necessarily mean building a mind. And since thoughts seem to be properties of “matter”, and there seem to be rules about which configurations of matter produce mind, we don’t necessarily know that there are other configurations that can produce minds. We might produce something else equally interesting which still is not a mind.
it’s totally possible that only a biological brain can produce minds as we understand the word “mind”.
Bollocks. Thought is a process, like math. Nothing meat does with signals is impossible in other substrates.
At the utmost extreme: surely we can simulate physics at whatever level is necessary for virtual brains to function. Physical neurons are not gonna rely on quantum chromodynamics. Mere chemistry will probably suffice.
And hand-waving things that are like-minds-but sounds like Chinese Room nonsense.
I’d be curious why you think LLMs are dead ending? Is it that you think the Jepa models are likely to find success and win out or do you think the LLMs in generally are just hitting their peaks?
Your point on power usage is interesting, although I think that is mostly on training not usage correct?
I think they’re a dead-end mostly because of the exponential cost vs. performance. The decreasing returns are obvious, and the companies are trying to adapt by raising token prices, but that will not be enough with the current user numbers (or even double or triple, if we believe the analysts). I think that, at least with these large LLM companies, we’re actually beyond the point of economic equilibrium with this technology, at current energy and water prices.
And yes, training is more expensive than usage. That’s probably the reason why Anthropic suggested a pause in LLM development (training), supposedly because of the fear that AI could become Skynet, but really because they are getting an IPO soon and if people see their current balance numbers, the IPO would fail and the bubble would probably pop. Which really proves my point a little: the economics of these companies “improving their LLMs” (training) don’t make sense at current energy prices.
Give it time. My software career is also affected. At the rate they’re spending money at an order of magnitude higher than they’re making. They’ve also all borrowed money from each other. It’s going to collapse in a big heap. Hopefully before it sucks in mum and dad investors.
It’s going to collapse in a big heap
Just like the housing market which the government bailed out at our expense. And housing prices continue to outpace salaries. There’s a strong possibility that this is a permanently-forced inefficiency.
I’m not so sure. The AI hate in the general population is real. The fact that all these companies immediately claimed to replace all human labor within months, made them the enemy of the people, regardless how false these predictions were.
Banks are not liked either, but they are not hated in the same way. At least in the general public a mortgage seems to be the responsibility of both parties. “So the banks speculated, we still need houses and mortgages.”
I’m not so sure you can explain away the AI bubble that easily. Occupy Wallstreet will be a small affair in comparison.
It’s gonna be about a decade more.
I’ll be retired by then.
If you’re in the US, social security isn’t going to be funding anyone by then either. So, I hope you have enough money to live off of and pay out of pocket medical. 🙁
That’s the plan, else I wouldn’t be retiring.
I’ll tag you a ‘rich bitch’ so we can have proper interactions in the future. Enjoy your retirement, truly.
I’m not sure what that’s supposed to mean.
Try to argue with the LLM evangelists, the inevitable brain damage it causes will let you get on disability.
On one hand this is not the first hype cycle, on the other hand the other hype cycles didn’t all fade. The inevitable bullshit brought on by vibecoding and shit like that is eventually going to be some kind of problem. People may or may not just ignore that problem, like most other issues in tech.
I’m hoping the insanity calms down in the next year or so. I don’t wanna have to give up weed to be able to start using my fallback option of my CDL but I’ll be damned if I’m gonna take a job where I’m required to use LLM bullshit. not that I can even find that after 6 months of applying and not even a single recruiter followup, let alone any interviews.
I’m still employed and I see myself employed (at least in that company) for a foreseeable future.
Hmm, well, ok. Moving on then
The world is changing. It happens from time to time. In this case the change is a particularly big one and it’s still ongoing, so I can’t make any predictions about where it’s going to end. But I can be pretty confident that it’s not going to magically change back. So my best advice is to try out the new tools, see whether you can adapt to them and use them to improve your own productivity in new ways, and if not then as a fallback start looking at other directions to take your career.
Harsh, perhaps, but the world does as the world does.
Author of the blogpost seems to be fully engaged with that, but their worry is that they no longer have a professional edge over the competition because LLMs let less experienced people do all the stuff they spent years developing expertise in:
Of course, I’m still employable because someone has to review the code and steer the robot. But I’m just another off-the-shelf engineer now. I have no domain expertise that another Sr. engineer steering an LLM cannot match. All my finance and payment domain expertise, all the debugging intuition and distributed system knowledge earned through hours of sweat and tears, is now promptable.
We were taught that generalists and specialists will always have their roles. But now the market is shaping everyone into becoming a generalist. That’s not a bad thing per se, until you look under the economics of supply and demand: if everyone is a generalist, the price of a generalist falls if there’s no demand to match. And we all know the demand is drying up.
Well frankly i think the field of eco terrorism is about to expand greatly so maybe start preparing for that.
Not sure I agree with the “three pillar” framing in this one. I think it’s more like he was able to use the LLMs effectively because he had already built those years of experience. Someone new trying to vibe code their way into designing robust distributed systems is going to need a similar amount of time to build the correct intuitions, because the LLM doesn’t give a shit one way or the other, it’ll just do what’s in the prompt.
On the other hand, commodotizing domains that fit a repeatable pattern isn’t so bad, is it? At work I sometimes have to remind people that we’re generally just building CRUD apps to interact with data, which isn’t exactly at the bleeding edge of software. I like the idea of software engineers getting to be more portable, but unfortunately the industry still thinks LLMs are going to replace most jobs, which is obviously not possible.
It’s a tool, for me atm it’s like spellcheck and google had a baby
Learn to code.
Read the post.
Actually, reading stuff can be a bit of a high bar for some folks, unfortunately.
Do you have a couple sentence summary for me?
