The Post-Developer Era
joshwcomeau.com
joshwcomeau.com
The reality is AI will change how software is built, but it's still just a tool that requires the same type of precise thinking that software developers do. You can remove all need to understand syntax, but the need for translating vague desires from laypeople into a precise specification will remain—there's nothing on the AI horizon that will solve for this, and savvy tech leaders know this. So why are we hearing this narrative from tech leaders who should know better? Two reasons:
First is that the majority of investor gains in the US stock market have been fueled heavily by AI hype. Every public tech CEO is getting questions from analysts about what their AI strategy, and they have to tell a compelling story. There's very little space or patience for nuance here because no one really knows how it will play out, but in the meantime investors are making decisions based on what is said. It's no surprise that the majority of execs are just jumping on the bandwagon so they have a story to keep their stock price propped up.
Second, and perhaps more importantly, regardless of AI, software teams across the industry are just too big. Headcount for tech companies has ballooned over the last couple decades due to the web, smart phone revolution and ZIRP. With this type of environment the FAANGs of the world were hoarding talent just to be ready to capitalize on whatever is next. But the ugly truth is that a lot of the juice has already been squeezed, and the actual business needs don't justify that headcount over the long-term. AI is convenient cover story for RIFs that they would have done anyway, this just ties it with a nice bow for investors.
What makes you think LLMs will never be able to do that?
Have you tried any of the various DeepResearch products? The workflow is that you make a request for a research project; and then it asks various clarifying questions to narrow down the specific question, acceptable sources, etc.; and only then does it do all the research and collate a report for you.
AI may or may not, this is to be seen, but LLMs of today certainly won't.
Sounds like the work of mustering up instructions... like programming...
So how do these LLM's completely remove us from having to do this work of mustering up instructions? Seems to me someone still has to instruct LLMs on what to do, and that the only way this reality will cease to exist entirely, is if humanity stopped desiring computers to do what it is they want. I don't think that's happening anytime soon.
However, maybe fewer programmers will be needed, but then again, the same was said of Fortran and COBOL and look at where we are today, more programmers than ever...
> Sounds like the work of mustering up instructions... like programming...
Again, try Deep Research. You make a vague request, and it works with you to make it specific enough that it can deliver a product with some confidence that it will meet your requirements. Like a product manager, business analyst, requirements engineer, or whatever they call it these days.
Exactly. Plus we kind of want to believe it. The "extrapolate to infinity" bias writ large. It's seductive. Step 1: AI that genuinely does some amazing things. Step 2: handwave, but look out Step 3: Super Intelligence and AI that does it all. "This changes everything" etc. And there are just enough green shoots to go all in on this idea (and I mean cult-level all in).
In practice it plays out much closer to the author's sentiment. A useful tool. Perhaps even paradigm defining.
Yup, now instead of blindly using tab to autocomplete I have to check the "AI" is not fucking up before pressing the key.
LLMs are very good at exactly that. What they aren't good at (I'll add the yet as an opinion) is larger systems thinking, having the context of multiple teams, multiple systems, infra, business priorities, security, etc.
This is also the only place where LLMs have had a tangible impact wrt product offerings with some path of utility, so it must be sold this way.
The broader public (my experience, unprovoked, in conversations with the non-technical) are even aware of this--a neighbor of mine mentioned "prompting for code" to me the other day, while "AI" was a topic we discussed.
Programmers have been well-compensated and I suspect there's some sort of public dissatisfaction with the perception of "day in the life" types making loads of comp to drink free lattes or whatever; no-one will cry for us.
While, there's a billion and 6 "AI Startups" "revolutionizing healthcare/insurance/whatever with AI" but with nothing that the public has seen at any scale that can even be sold as a plausible demo.
Image/music gen and chatbots writing code are basically all of it, and the former isn't even often sold as a profitable path.
At several times I was like « why did you do that this way? This is so contrived ». And I should have known better but of course the answer started with « … I know but Copilot did this or that… ». Of course no test to validate properly the implementation.
The sentiment expressed in the article were developers won’t even bother to validate the output of coding assistants is real. And that’s one of the biggest shame to this current hype: quality was already decreasing for the past 10 years, signs indicate this will only go downhill from here
A HUGE portion of work in f500 companies that are non-software firms is outsourced--these companies spend loads of cash on consulting companies (yes, those ones) that often produce results to which the description of sub-par would be a compliment. There's been various cases made public, but more often is routine and poor quality mundane work that's way over budget and barely meets requirements.
If anything, the prevalence and adoption of LLM-generated code will increase the quality in a lot of places.
If you've never had to wade through something that one of Those Companies wrote, you have no idea how bad it gets and how frequently this is the case.
We just saw DOGE-related cuts to Accenture and Deloitte--these companies have huge contracts all over the place, not just public sector.
There's a massive amount of crap out there.
Most companies do not care, do not understand "quality" outside of some cargo-culted notion of "clean", have only adopted modern practices as some sort of ritual* of Things You Do without any assessment on what tangibly works or doesn't, have no understanding of how to attract skilled professionals, nor assess them, nor nurture them. Often their definition of skilled is a resume that contains the proper "experience" with whatever Java/Angular/React/we-love-containers thing they use. With the right number of "years" near it.
*a recent example--company has GH Actions with CI/CD for every commit on all branches but routinely suffers delays due to misconfigurations, runner issues, and other headaches. Code Reviews for PRs exist but are basically pedantry that just slow down deployments ("change this to switch over if-else") while missing crucial bugs, randomly updating major dependencies cause vague security warning that breaks interfaces and other APIs. All this and more--and this is one of the better examples I can think of, in terms of "quality."
I'm saying this as I worked at a Pretty Good consulting company that's business model changed to sub-outsourcing some of the labor to some of These Companies and often had to fix, assist, push back on, work extra to clean up, the various decisions made in these cases.
I also "inherited" and worked on fixes for things left behind by some of these places.
Good thing is there often was low-hanging fruit. Extra and overprovisioned cloud resources wasting $$ and the like, absurd architectures with extra dbs and queues, and all sorts of nightmares.
Your real problem is a people problem: hiring, management, feedback, etc.
Happy to be an independant and not an employee to not put up with this crap
Do implementation with minimal (not none, tbf) AI support, ask copilot if there are any obvious issues with and otherwise check my work.
This workflow has helped me catch quite a few gotchas I would have otherwise missed.
Coding assistants are really helpful for validating output, I've had much more mixed results trying to use it to generate novel outputs.
Having an LLM write the code for me? Blecch, it doesn't do it right.
Have an LLM make suggestions about my code? That's fine. If some of them are asinine I just get to laugh and feel smart while ignoring them. But if 1/5 of the suggestions are actually good? That's a win.
But if 1/5 of the questions I ask an LLM are correct, that's a waste of time. Funny how the accuracy of the model matters a different amount depending on the task at hand!
Less experienced developers are the primary vector of propagation for this « low quality » output, with seniors trying to educate and review the mess (if time permits)
If the market gave your company a PE ratio of 20+ and you're flush with cash, it is borderline fiduciary negligence to be slashing projects and doing layoffs. Your shareholders didn't invest in you for the capital preservation portfolio-management skills of your finance department.
> That's what makes you an investor -- you're offering the company the value of your shares by exposing yourself to "inflation" whenever the company needs to raise some money.
I beg to differ. What makes me an investor is buying a share of future company profits, directly via dividends or indirectly via a future sale of shares. Many of the companies I invested in over the last 30 years were (over time) reducing share counts via buybacks, not issuing shares like candy.
Except that you are -- every time that the company prints shares, they've taken some value from you to fund their operations. If companies weren't allowed to issue shares from thin air, there wouldn't be much point to the stock market.
That said, I do think he understates how much the value of certain developer skills is shifting. For example, pixel-perfect CSS or esoteric algorithmic showmanship used to signal craftsmanship — now they’re rapidly becoming commoditized. With today’s tools, I’d rather paste a screenshot into an LLM and tweak the output than spend hours on handcrafted layouts for a single corporate device.
Why are LLMs in these context being viewed as more than really powerful auto-complete/intellisense? I mean, correct me if I'm wrong, but aren't LLMs still years away from auto-generating a complete complex codebase that works without fixing a lot of bugs/assumptions?
If I may focus on this particular line without invalidating the rest of the content -- why aren't they going to benchmark the code?
That said, at least we have reached the phase where AI tools are commercialized, so that's another +1 I guess.
In the near future we'll probably see a lot more subject matter experts creating their own tools instead of requiring a dedicated person to translate their requirements.
1. Millions of layoffs across industries due to AI with some form of questionable UBI (not sure if this works)
2. 100BN in profits. (Microsoft / OpenAI definition)
3. Abundance in slopware. (VC's definition)
4. Raise more money to reach AGI / ASI.
5. Any job that a human can do which is economically significant.
6. Safe AI (Researchers definition).
7. All the above that AI could possibly do better.
I am sure there must be a industry aligned and concrete definition that everyone can agree on rather the goal post moving definitions.
Please do not anthropomorphize the AI, it's not a he.
Surely the level of intellectual difficulty in software engineering is similar to practicing law, medicine, banking… ?
Know if it actually is your tool.
AI is at best very helpful.
It’s a very very long way from making programmers obsolete.
However, LLMs might already be good enough to replace teams of 10 software developers with a domain expert and an expert developer armed with good LLMs.
That's enough for a fundamental, very violent change to the software industry.
Sadly most shops greatly overlook this reality, stuck in 1991, and continue to add redundant layers and staff thinking it makes them 'agile'.
In fact, this was totally and easily explicable after the dominant political party tried to convince Americans that their current leader was not suffering from severe cognitive decline and was "sharp as a tack" (a Biden admin talking point) and that NPCs lying to Americans about Trump being "easy to beat" by Kamala and also lying to them about Trump calling Charlottesville protestors "very fine people" and thinking that those fake attempts to make Trump look like he was endorsing Nazis wouldn't backfire explosively.
So, no, it was not "inexplicable" at all but it was rather Whiplash Effect initiated by media narratives originating in the Democrat ecosystem. And don't forget: Trump, Elon and Rogan are all ex-Democrats. I wonder, too, if the author was one of those people who deluded himself into thinking "Kamala is a great candidate" against all the evidence.
So the "Russiagate" hoax failed, "Kamala is awesome" failed, elites orienting themselves around appeasing far left pressure groups failed, smug contempt for middle America failed (astonishingly) and yet it is "inexplicable" why Orange Man Bad won the election!
Same as it ever was. No further explanation needed. Not a new phenomenon.
ed: "No, it must be the fault of the left." must be a reassuring thought, but let me occams razor that for you: when these people were younger, they were selfish, and as they weren't yet the status quo, they were progressive. When these people got older, they were still selfish, and as they were the status quo, they were conservatives. Simple.
+100 to that