Rendered at 07:15:09 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
a_bonobo 6 hours ago [-]
>On my visits to the Bay Area, I would ask AI researchers or interns why they are doing their current research or projects, when in a year or three agentic LLMs could probably do them;
This is such a weird point to make that doesn't become correct just because everyone makes it, all the time. Why clean the ocean if some magic future tech will clean them? Why save the world now if some benevolent AI is 'just around the corner' and will do it for us? And people have been making this point for years now, and it's not like my job got any easier. I just got more AI.
And I say that as someone who uses Claude Code in complex environments almost hourly; I, as the human, still have to do the thinking as Claude still 'can't jump' [1] and I have seen no evidence that they (or similar AI, any time soon) will 'jump' like a human brain does.
I think it is a great point to make, because if everyone really believed that AIs will do everything without human intervention in a handful of years, as the marketing repeats again and again (AGI, singularity, etc.) and have been saying for years... why then get bothered?
Because we DO know LLMs have their hallucinations, limitations, perform tasks not previously seen way worse than humans, etc. And it seems that, for now, there is not a good or magic solution to it, it is inherent limitations of the paradigm.
Yes, you can feed more and more and more (curated data) and eventually make AIs excellent at task X or Y, but then you spend your time specializing those engines. So the work does not really disappear, it just shifts and you make it more replicable for a bound set of problems.
Needless to say that at some point I prefer to learn (and combine with AIs, it is ok) than acritically getting inputs from something until I become totally useless.
Unless we have a paradigm for which a fully autonomous AI can do everything, this will just become improving our productivity in some ways, with all the in-between bottlenecks that it has.
nextlevelwizard 33 minutes ago [-]
Why spend money and time making the new flagship model when a future flagship model can make you the flagship model?
germandiago 30 minutes ago [-]
Then why not stop researching and doing the definitve model that will solve every problem? Why some people are not doing it?
Bc they are aware of the marketing and limitations. If they did believe it, then they would switch area of research.
joaquieneCnix 9 minutes ago [-]
> inherent limitations of the paradigm
This is such a weird point to make. We are currently ( only ) discovering that paradigm; we are not inventing anything. We found a bunch of laws that produce rather cool results but our paradigm is incomplete which leads more or less wordy or frame-rich weird stuff like hallucinations, singularity and so on ... it's childish, really and on that funny pseudo-profound, pseudo-intellectual, pseudo-spiritual ( personal opinion, if it gets you horny, you go, baby ) "universe consciousness unity, Rick James, bitch" level ...
Our bodies and minds need proper AI, not all the stuff we already outsource to middle and/or passionate men and women. Other species on the planet would certainly like to see us get augmented by AI so we can solve as many survivability issues as possible to keep as many ecosystems running long enough ... whatever that means but whether animals and plants are aware of chance and potential is another philosophical debate.
To individuals, software is a hammer and chisel, a knife, a brush and canvas, pen and paper, a reading help, and to a good amount of people it's a microscope and a fine scalpel.
To collectives, it's a tool to work on consensus and conventions, to share and gather.
It's baby steps for civilizations and it looks like our particular species is gonna get stuck in a puddle of our own monkey shit, with bottles of champagne in our hands and monkeys grinding up and down the few ivory towers in proximity.
> why then get bothered
Humans are on different levels. Most have decided that "nature realized the/a bug and wanted someone dead" or "their survival is a matter of chance" is not acceptable at all and some people decided that sabotage, poison, abuse, rape, murder are acceptable means to get chicken shit ...
The "paradigm" of life is far from explored/discovered, so we simply can't content ourselves with presumptions about inherent limitations of the LLM and AI paradigm for any other reason than to uncover ( not invent ) other parts of the paradigm.
We are happy with what AI can do for us but "AIs will do everything without human intervention" sounds weird because babies are born and the older they get and the less sabotaged ( vs influence, cultural manipulation ) they get to grow up, the more breadth and depth humans want to experience. For this they need to learn and use their hands & fingers. They need to feed body and mind to find what triggers what, and what excitement and curiosity are inherent and which can or need to be added/acquired/experienced extrinsically.
How many associations will we be able to make if AIs will do everything without human intervention?
Yes I double checked the quote is not in the article. HN is probably the best place on the internet for people actually reading the article, but this being the top comment here suggests that the majority of voters still do not read the article
pjmlp 15 minutes ago [-]
Robot powered factories still have humans there, the gist is that they are a tiny fraction of what a classical factory would require 50 years ago.
DrewADesign 5 hours ago [-]
Same reason some think preserving the environment is pointless because the believers will ascend to heaven, either way. It’s a religion. It’s dogmatic nihilism.
bwhiting2356 1 hours ago [-]
A fully automated utopia isn't just going to happen. Even with frontier models, the integrations, the evals, the UX, need a lot of work and someone needs to do it. After I've automated this thing I'll move on to the next task, this is what it means to be a software engineer.
ReactiveJelly 1 hours ago [-]
An actual utopia would require never-before-seen democratic mandate from people who are currently on the brink of hot civil war
fhub 6 hours ago [-]
> I, as the human, still have to do the thinking as Claude still 'can't jump'
I still have to do quite a bit of thinking but the amount of of thinking I do per task is trending down. I agree LLMs are not good at abduction but very few humans are either and very few jobs/tasks require it. I can't talk for researchers jobs though. But perhaps fewer researchers would be desired by these labs (not none).
bigfishrunning 5 hours ago [-]
> I still have to do quite a bit of thinking but the amount of of thinking I do per task is trending down
Don't worry, I'm sure you'll hit your goal of zero thinking soon!
xprnio 4 hours ago [-]
Trending towards the Homo Amens Mechanicus: the mechanical thoughtless human. What a goal
germandiago 39 minutes ago [-]
Only bc of this I will keep balancing what I do with my brain with what machines can do. dangerous outcome.
The IQ willl drop if we just become mechanical acritical people the same way muscles get worse if you do not exercise.
champagnepapi 5 hours ago [-]
[dead]
jihadjihad 5 hours ago [-]
> I agree LLMs are not good at abduction but very few humans are either
I, too, am glad that few humans seem good at abduction.
although, if i'm out of tokens and have to wait a full day, i won't bother doing some things manually because the day i'll spend doing something won't take more than 1 hour the next day when tokens are available again.
MikeTheGreat 3 hours ago [-]
That seems like a somewhat orthogonal point? Like, if I'm a carpenter and my batteries all run out / I can't actually power my power tools then the best course of action is to go home and recharge all the batteries instead of trying to hand-cut 100 pieces of lumber today. After all, the power tools can do it a lot faster (and with less effort) than I can.
I say this as someone who's watched a bunch of woodworking videos but hasn't actually done this myself :)
whatsgolden 1 hours ago [-]
I read that more so as, I'm a carpenter and my batteries have all ran flat, so I'll put them on charge and do something else today. I'll cut up the lumber tomorrow when the batteries have charged.
sevenzero 1 hours ago [-]
Also people tend to forget that LLMs still just work on compressed data... Where are the MAJOR breakthroughs? Where is all the "crazy" AI output going? Software seemed to degrade in quality a lot in the recent years. All "improvements" LLMs go through are simply improvements on how to burn more tokens out of my pockets given that Claude now want an actual browser extension to "visually" confirm small changes every time I use it for UI.
They are still just data parrots.
ChicagoDave 37 minutes ago [-]
There’s a small group of established architects talking about harness engineering, but I’m not sure anyone is actually listening to them.
And those same architects are quietly extracting real productivity from GenAI.
And even this write up skips that info by waving, “Some people…”
dgellow 22 minutes ago [-]
Mind sharing some names or something else we can learn about?
sevenzero 31 minutes ago [-]
Idk, maybe some people get crazy productivity out of LLMs. To me, going deep into the AI bubble, reading about terms I've never seen before just feels like some crypto bro bubble with people being too deep into the sauce to notice that these things are not the wonder machines they believe so hard in...
skew-aberration 5 hours ago [-]
AI has made 'jumps' in demanding fields like leading mathematical research and has made advancements in AI research itself. Is now a good time to start a maths career? Is there a field of research (yours?) which is inherently (more) AI proof?
Btw, I think the discussion of Einstein's career in the paper you link is historically wrong in many respects, particularly the argument about 'weak signal'. Einstein was in fact working on some of the most mainstream and widely discussed problems in physics of the day, he is admired for the creativity of his solutions to those problems, and much of his work built incrementally on ideas and breakthroughs that came (long) before (as all research does).
Article suggests that a central motivation of Einstein's work was resolving action-at-a-distance in Newtonian mechanics - yet Maxwell introduced the same Lagrangian field theories for electromagnetism we use today 50 years earlier to solve the same problem for Farraday's laws of electromagnetism. Similar wave equations existed even earlier. Heaviside in 1893 extended this technique to gravity (matching 'weak field' GR) 20 years earlier. So this is perhaps the one aspect of gravity that had actually already been solved before Einstein. Authors might be conflating his work on action-at-a-distance in QM.
Einstein's GR extended the linear 'weak field' understanding of gravity to include the non-linear self-referential case where masses themselves create gravity. This was mathematically incredibly difficult but was necessary precisely because SR's mass energy equivalence created so many strong signals that were unresolved. For example: if finite energy is mass, then mass changes as objects accelerate past a large mass like a start, and hence their propagation in space could not be explained by linear EM style field equations. Many such considerations were causing very 'strong signals' in SR, and there were analogous problems in QM atomic models being developed at the same time.
SR was also a solution to a problem that was actively being worked by many of the leading physicists of the day. SR actually does match Newtonian mechanics for a single observer - it resolves contradictions in the case of separate observers, by allowing them to assign different values to the speeds, masses, etc of objects such that each object appears to follow Newtonian mechanics for each observer. Again, this was necessary because of a lot of contradictions related to the behavior of light that had been well-known for ~20 years at the time.
Personally, I don't consider this kind of reasoning to be beyond the capabilities of future LLMs (even current LLMs if the task was broken into technical rather than philosophical problems). Personally, I doubt that such problems could stand open for 20+ years waiting for a creative genius to solve them in the modern world.
And don't get me started on the philosophy.
saghm 3 hours ago [-]
It also seems kinda tone deaf. If someone basically told me I was wasting my time and asked what I would do in the future, I would not bother giving them a particularly thoughtful answer because trying to spend effort justifying my life choices to them would be the actual waste of time.
What kind of answers were they expecting to get?
paul7986 5 hours ago [-]
By trade I'm a UX Researcher/Designer who designs in code (HTML/CSS) and have done so since 2009. Recently I vibe coded an entire python app with a database and each time I didnt know what to do I would just feed screenshots to Gemini or Codex for guidance (i think i could share my screen with Codex and it can guide me via a voice conversation). I know I could follow up and build a companion iPhone and Android app using these tools.
Overall, I'd like to understand those who have a positive outlook on design and software engineering as a career. Where do you see the opportunity where I just see a bleak one where anyone can do this stuff by typing or talking to AI? Myself, after 17 years in the field I am begrudingly back in school for a new medical career. As well, anytime an IT recruiter reaches out I am getting responses back only after under-cutting the hourly rate I use to demand and what others probably are still trying to get. And with it feels even bleaker as it becomes a race to the bottom!
cjcenizal 9 minutes ago [-]
I also work in UX and SWE, and heavily use GenAI in my work. I don’t have a positive outlook for people who limit their career to one of those fields, but I do have a positive outlook for generalist, multi-disciplinary careers. When you have the experience and skill to steer product development from end-to-end, you can produce high-quality products super-quickly. The experience and skills are the differentiator — if you lack those you can still use GenAI to move fast but probably in the wrong direction.
aryehof 19 minutes ago [-]
The industry has vast (and increasing) oversupply of “programmers” versus diminishing demand. Add to this, the adoption of AI.
> Overall, I'd like to understand those who have a positive outlook on design and software engineering as a career.
I think until the market better achieves some equilibrium, there is no way general software programming (sorry “engineering”) should be considered as a career. That said, there will always be opportunities in particular markets or specialties.
griffiths 3 hours ago [-]
In my experience, not everyone can really do this stuff by typing. I think you need to be creative, resourceful, inventive, open minded and have ideas how to approach the typing/prompting. I see many people struggle in using AI.
> > Mathematician Richard Hamming used to ask scientists in other fields "What are the most important problems in your field?" partly so he could troll them by asking "Why aren't you working on them?" and partly because getting asked this question is really useful for focusing people's attention on what matters.
> I imagine someone being asked this question, and how they should respond. I think like so - ‘Fuck off Richard’.
> This is partly because I imagine this question being asked in a kind of snarky, gotcha kind of way, with some sort of nerdy superiority. Like ‘ha your behaviour is inconsistent with your implied preferences, you idiot, do you even von Neumann–Morgenstern?’
esafak 6 hours ago [-]
Also, if you believe your well-paying job is eventually going to be automated you would be prudent to bank the money while you prepare for the future.
simonw 6 hours ago [-]
> We already know developers don’t actually spend most of their time writing code, with studies at Microsoft and elsewhere showing it’s closer to 14 percent.
Anyone else finding they're spending more time writing code (or at least driving agents to write code) now?
14% used to feel about right for me - I'd spend the rest of the time researching approaches and libraries, planning things out in issues, or sometimes just thinking really hard about problems I ran into.
Now... I still do those things, but I'm doing many of them faster - and I'm often doing them while my coding agents are churning away on code.
There's also this weird effect where the harder a problem is the more I can get done in parallel with it, because an agent might need to spend 20 minutes on it without my involvement.
01100011 4 hours ago [-]
You read my mind. I suspect it is a transient spike while I blow through the backlog and I'll run out of things that AI can vastly accelerate in, say, 6 months. But right now? I'm spending 80-90% of my time blasting through implementing features, finding bugs, fixing old bugs, writing and improving my tools... Code was never the majority of my job. But right now the bar to doing things is so low and the sheer joy of blasting through some previously tedious, low to medium hanging fruit is thrilling.
AdieuToLogic 4 hours ago [-]
>> We already know developers don’t actually spend most of their time writing code, with studies at Microsoft and elsewhere showing it’s closer to 14 percent.
> Anyone else finding they're spending more time writing code (or at least driving agents to write code) now?
Not really, as once it is time to write code, the problem has been defined/understood (to the degree possible with knowledge acquired at the time), and encoding it is largely an exercise in typing along with verifying assumptions via test suites.
Does GenAI quicken some portions of the above workflow? Sure, in the same way IDEs with contextual code snippet suggestions can make encoding faster.
geoduck14 6 hours ago [-]
I used it to write SQL and make dashboards. Back in the day, I would spend a lot of time doing that, then I changed roles. I dipped my toe in it recently and used AI exclusively. I would send a prompt, see the output, decide if that is what I wanted or not. I kept my brain in "what-if mode" and I let the LLM handle the technical specs.
geraneum 3 hours ago [-]
Does the code get reviewed? How do you deal with increased amount of code that may need to be looked at?
Ampersander 9 minutes ago [-]
If you're going to be reading the code you might as well write it by hand instead of using AI.
simonw 2 hours ago [-]
I review the code that matters - anything security adjacent or that's an API that will be used by other code in the future.
I don't review code that either works or doesn't - most HTML and CSS layout code for example. There I test it on desktop and mobile and commit it if it works.
Ditto for stuff that's simple. A JSON endpoint that runs a SQL query and returns some JSON? If it works and a glance at the tests looks OK then I trust my agents wrote it properly.
I'm getting more confident with my judgement over what needs a close look and what doesn't over time, as so far I haven't been majorly burned my any mistakes that snuck through.
Honestly, it's similar to being an engineer on a larger team. You don't review every line of code written by every one of your coworkers.
I think this is THE issue of our time as programmers to be honest: do you review every line of code an agent writes?
An increasing number of expert programmers are moving in the direction of NOT reviewing every line. It's working out OK for a lot of them.
mekael 54 minutes ago [-]
I've found that engineers on a large team do read every line, mainly due to the fact that the skill levels run the gamut from intern to lead, and only 1 or 2 people out of 12 might have knowledge of the application being modified.
It's actually getting worse due to "AI code bloat", for example I have 16k lines of code to review across 3 apps by the end of this week. Normally it would be a quarter of that, but what Claude produces is extremely verbose in some places and anemic in others, and I can't tell at a glance what's right and what looks right with that much ground to cover.
t-writescode 1 hours ago [-]
> Ditto for stuff that's simple. A JSON endpoint that runs a SQL query and returns some JSON? If it works and a glance at the tests looks OK then I trust my agents wrote it properly.
That is *exactly* the sort of area I *wouldn’t* blindly trust AI, there’s a huge security boundary there. What if the AI is doing string concatenation with user-provided data???
simonw 15 minutes ago [-]
Once you've seen the AI not make mistakes like that a few dozen times you start to trust it not to mess that up in the future.
hombre_fatal 1 hours ago [-]
We're pretty far past this if you're using anything close to the sota models.
But you could be defensive with a security checklist in agents.md and have adversarial review, if you wanted.
slopinthebag 37 minutes ago [-]
I think it also depends on what you're building. Some solo project or basic html thing? Sure no need to review every line. It's a bit different when you're working on foundational libraries that a business relies on, anything touching a production database, etc.
skydhash 50 minutes ago [-]
> Honestly, it's similar to being an engineer on a larger team. You don't review every line of code written by every one of your coworkers.
We don’t because everyone is accountable for his or her own mistakes. So everyone is incentivized for their recklessness to not be the root cause of some bug.
> An increasing number of expert programmers are moving in the direction of NOT reviewing every line. It's working out OK for a lot of them.
Have you ever asked your users? What about bug reports? Is the amount and rate decreasing?
enraged_camel 6 hours ago [-]
Yeah. I spend most of my day driving agents to write code, verifying the results, orchestrating work streams, and so on. The rest of the time, a Fable agent is organizing work in Linear/Jira and making sure coworkers are getting their stuff done in a way that won’t conflict.
levmiseri 52 minutes ago [-]
> f developers spend only about 15 percent of their time typing in the editor
I think this is missing an important detail. Lots of time was spent on non-coding stuff, because coding used to be more committal and hence expensive. With how quickly one can code up a quick prototype or even production-ready code these days, the code becomes the communication tool as well.
willtemperley 11 minutes ago [-]
With all the myths and hyperbolae circulating regarding AI, I'd love to know what it's like at large software companies adjusting to this brave new world.
It's easy for a small team to adjust workflows and roles, but I just imagine the office politics must be a waking nightmare in big organisations right now.
32 minutes ago [-]
mkozlows 6 hours ago [-]
I feel like all you need to know about how seriously to take this is that they cite that ancient early-2025 METR study, and describe it in the text as "recently one even found..."
katzgrau 5 hours ago [-]
Same thought - 80% through reading it occurred to me to check the citations. A few items from 2025 and most well before that.
So much has changed since late 2025 one can’t really draw any conclusions from this.
In fact, I’m guessing things will continue to move so fast that by the time one were to execute a survey of developers, many of the responses and findings are no longer relevant.
greenhat76 4 hours ago [-]
Your point really goes both ways, we really don't know anything about how LLM usage is affecting anything. No one knows, it's the wild wild west, which is whatever. But I think no one can really draw conclusions from what's happening in tech right now.
Reminds me of COVID and how everyone was fighting over early trends during that time.
joshuastuden 4 hours ago [-]
Exactly. I saw them using things from 2025... AI sorta sucked then and didn't really "take off" until that Opus drop in December or whatever it was.
CompoundEyes 5 hours ago [-]
I felt the same and why didn’t the authors look over METR’s recent material?
The whole point of the 2025 one is that they found the self-reporting to be significantly inflated, which is why self-reported surveys like this one are hard to trust.
mkozlows 3 hours ago [-]
Yes, but their newer write-up discusses that (and shows that the self-reported numbers have gone up radically, in a way that suggests that even if there is some inflation, the numbers are almost certainly positive if you deflate).
They also have an update -- linked from the original study! -- explaining that it's out of date and no longer reliable, and explaining why they had to cancel a follow-up study because it was understating productivity gains (but also was showing wins for the people who carried over from their previous study): https://metr.org/blog/2026-02-24-uplift-update/
The authors of this paper decided to ignore all of METR's follow-up data and discussion, and to report only the ancient number from early 2025 (a time when Windsurf was state of the art). And then, rather than apologizing for it, and caveating it as a number not to be taken seriously, they described it as a study done "recently."
That's either shockingly dishonest or incredibly out-of-touch.
bjourne 9 minutes ago [-]
> Despite this decade-old research, many organizations still rely on lines of code as a measure of developer productivity.
No, they don't! It's easy to dispel myths when the myths are built on straw men. Dumb article.
lz400 5 hours ago [-]
Like many others in the comments, I feel there are a lot of assumptions in this piece. Before, coding is only 14% therefore, small slice. I think that's a very superficial assumption. That was because coding was expensive and we needed to be sure we didn't code the wrong thing. If code is as cheap as it is now, we will optimize differently, we will structure around it. Instead of so many meetings we will code 5 different versions of the same thing and choose, etc.
AdieuToLogic 3 hours ago [-]
> That was because coding was expensive and we needed to be sure we didn't code the wrong thing.
Coding has never been expensive as it is nothing more than a reification of a solution to a problem as it is understood at that time.
It is the underlying understanding of the problem which has always been expensive and remains so.
lz400 2 hours ago [-]
Coding was expensive in the sense that once you decided what to do, it took a few engineers months / years to do moderately complex projects. That's not true anymore. Therefore the risk of "coding the wrong thing" is less.
gmueckl 1 hours ago [-]
Is it? The temptation to start without a thorough design is now much stronger because the implementation osnperceived to be cheap and easy to replace. But if you start building the wrong thing fast, you still get the right thing later than when you had checked properly at the start.
champagnepapi 5 hours ago [-]
So you're suggesting that coding will take more of the PRD phase?
lz400 2 hours ago [-]
Basically yes, there will be more coding in that phase, more prototyping, the PRD phases will be shorter too, there will be more pressure to deliver quickly and the PRDs will be under more pressure to move more quickly. This is what I'm already seeing to be honest.
kylecazar 6 hours ago [-]
I don't understand Myth 1 (Developers Spend Most of Their Time Writing Code).
They quote a study in which developers report to spend 11-14% of their day coding. The rest is stuff like solution design and meetings. The insinuation is that AI can at most automate 14% of your day.
The problem with this argument is that once you have code, some (not all) of the precursors to code go away.
unknownfuture 6 hours ago [-]
Okay.
Show me the evidence that AI has an impact on productivity when doing design work. Or reducing meeting load.
My own experience is that AI doesn't tighten the design cycle, and in fact might extend it by encouraging gold plating.
Izkata 4 hours ago [-]
> Or reducing meeting load.
Well, I expect when you've vibed too much and lost track of the code, and can't answer questions in meetings anymore, you'll stop getting invited to them.
griffiths 2 hours ago [-]
How often do you talk about low code details in the meetings? It is mostly about requirements and goals (IMHO) and "architecture"/design, and now I wouldn't even bother my peers with questions about the code as I can let AI tell me how things work exactly as written in the code, not some thing that someone remembers or thinks how it works (in large codebases, most developers only know in detail the things they are working on currently or from recent history)
otabdeveloper4 55 minutes ago [-]
> It is mostly about requirements and goals and architecture/design
People vibe code because they have no clue about any of that. Not because they're slow typers.
01100011 4 hours ago [-]
It helps me. Case in point, I recently had a API refactoring project that was 80% busywork and only a small amount of code. It's crap work, but it has to be done and it's part of my role. I told AI to generate the initial design doc. It took it upon itself to reference the local headers on my system and found some APIs that should be included that I hadn't even considered would also need to be touched for consistency. It saved me hours typing up the doc, requirements, test plan, etc and then saved me embarrassment by preventing me from making a stupid oversight on a tedious and boring task.
thewhitetulip 3 hours ago [-]
AI is helpful in such projects. Less so in other projects where design is heavy
sandeepkd 4 hours ago [-]
From the patterns I have seen people would tend to use the time to build demo's using AI instead of design and then have a back and forth on the demo itself. One has to consider the fact that demo's look more convincing/complete and give an impression that the design decision has been made. In some ways, how agile got rid of the need for explicit & detailed documentation, the AI created demo's will cut the requirements of high level designs too.
Exoristos 3 hours ago [-]
> demo's look more convincing/complete ...
In the case of an LLM generated demo, usually deceptively so.
> and give an impression that the design decision has been made.
In the case of a vibed design, this is the opposite of useful for the team.
sandeepkd 2 hours ago [-]
The concept of MVP has been there for a while, it felt deceptive too and incurred lot of tech debt. However one would feel the deception only if they understand the missing details. For a lot of roles making decisions it benefits them to ignore the details unless it really bites them
simonw 6 hours ago [-]
What kind of shape of evidence would you find convincing?
unknownfuture 6 hours ago [-]
Self-reported or observational data capturing time spent for categories of task ala every other similar study in this space?
This isn't exactly novel territory, here, Simon. Let's not pretend I'm asking for something strange, unprecedented, or unreasonable.
The article mentions that coding is only a fraction of dev time and thus accelerating that part of the job can only create incremental gains, and oh PS, lines of code and similar metrics are a bad way if measuring dev productivity, anyway, and we've known that for decades.
The OP claims AI accelerates non-coding parts of the job, too, and so the article is misguided.
I ask for evidence.
In response you give me... code output metrics?
simonw 1 hours ago [-]
I don't have anything else to hand that I can think of. I don't keep a time tracking diary.
qsera 4 hours ago [-]
I don't think volume of code changed was the metric the other commenter was asking for.
4 hours ago [-]
the_af 3 hours ago [-]
Isn't this falling into the "lines of code" trap TFA mentions?
More code written is not a good measure of productivity. It could be garbage, or redundant code, or simply not addressing the real or more pressing needs, it could be building the wrong thing, etc.
Like TFA mentions, it's been known for decades LoC is a misleading metric for productivity. It's one of the lessons of software engineering.
simonw 2 hours ago [-]
I'm planning an article at the moment in defense of lines of code. Saying "lines of code are a bad measurement" is too easy! It's about time someone presented an opposing argument.
In my specific case, lines of code for my published open source projects is a metric that I trust, because I have high standards for those. I have plenty of other projects where I'll accept poor quality, unreviewed code (almost all of https://github.com/simonw/tools for example), but Datasette, sqlite-utils and LLM are not that.
Of course, that's only useful for me personally and for people who trust me to stick to my own self-declared high standards!
skydhash 26 minutes ago [-]
> Saying "lines of code are a bad measurement" is too easy! It's about time someone presented an opposing argument.
Maybe because there is none.
One of the main quality of good codebase is simplicity. Which is about how easy for someone else to understand the code. It’s hard to define what simplicity looks like, so the best bet is to avoid the other side, making the code too complex.
And you can make the code complex by shortening variable name, doing code golfing with quicks of the platforms, so smaller LoC. You can also go the other way and increase the LoC by adding unneeded abstractions, repeating slices of code,… There’s a window where the LoC is perfect to attain simplicity, but that amount is an effect of striving for simplicity, not a cause of it. And it’s variable for every problem.
So you got something where the correct value is a different for each case. And trying to manipulate it artificially often results in complex code. And you want to say that is a good metric for productivity?
And in the cases of your projects, there are a lot more info could share that are interesting, like the amount of issues (reported or found by you) that are tied to implementation bugs (coding, libraries API breakage,…) or design issues (requirements conflicts,…), documentation improvement,… Anything that is tied to the actual usefulness of the projects, and not fumbling around with code.
otabdeveloper4 53 minutes ago [-]
> Lines of code are akshually a good metric now, because that's the only metric LLMs can optimize and I really, really, really love LLMs, they're the bee's knees.
Really now?
simonw 31 minutes ago [-]
No.
bluefirebrand 6 hours ago [-]
People working fewer hours :)
kaashif 5 hours ago [-]
Damn, looks like we have lower productivity than cavemen!
keeda 4 hours ago [-]
Sir, this is Capitalism.
therealdrag0 4 hours ago [-]
It certainly helps as a research assistant for design work. But it can’t do it for you.
cuttothechase 6 hours ago [-]
[flagged]
mountainriver 6 hours ago [-]
It does because you can now just steamroll features out the door and make everyone look bad that’s sitting around in meetings all day
dgellow 14 minutes ago [-]
Are you aware literally everyone else also has access to AI stuff? You’re not special for using an agent. You’re part of the crowd
mkozlows 6 hours ago [-]
Yeah, this seriously drives me nuts.
That meeting that you spent an hour in to understand the requirements? You don't need that meeting if you're not writing the code. That sync up with the QA engineer you did to hand it off to them? Don't need that meeting if you're not writing the code. That half hour you spent installing vim extensions? Don't need 'em if you don't open vim anymore.
There are engineers whose jobs go well beyond coding, of course. Staff engineers and principal engineers have had their jobs radically change because of AI, but not because it's writing all their code.
But there are also a lot of engineers -- your standard mid-level engineer, or even senior engineers at a lot of orgs with title inflation -- whose job is almost entirely about delivering code, and who spend all day either writing code or engaging in scaffolding around code-writing activities. Let's not pretend that automating away that code writing is a 15% boost.
decimalenough 6 hours ago [-]
> That meeting that you spent an hour in to understand the requirements? You don't need that meeting if you're not writing the code.
How are you going to prompt the LLM or validate its output if you don't understand the requirements?
mikeyouse 4 hours ago [-]
In some number of cases, the business partner who would have passed you the requirements will just generate the code themselves.. I’m shocked shin some engineers don’t see a massive disruption coming..
blharr 3 hours ago [-]
Sure, but why doesn't said business partner just give a requirement to the AI to do the business partner's own job? Since it is hypothetically superhuman at fulfilling requirements at such a point...
PMs seem pretty easy to replace. In fact, given any single role to fixate on, it's probably replaceable
4 hours ago [-]
claytongulick 4 hours ago [-]
Oh, some of us see a massive disruption coming, just maybe not in the way you think.
Talented, experienced devs leaving the field out of disgust would be pretty disruptive.
Crumbling, faulty technical infrastructure with unmanageable heaps of AI slop that no one understands would definitely disrupt businesses.
Skill atrophy, and an entire generation of "developers" that have never actually written code will be disruptive when something breaks and no one knows how to fix it (this isn't a hypothetical, I'm seeing this phenomenon in a lot of large enterprise groups).
The total elimination of novel solutions, new ideas and new approaches to development and the resulting industry wide stagnation won't be disruptive, but will certainly be a drag.
griffiths 2 hours ago [-]
You will have artisans or software SWAT teams on one side and code factories on the other.
For most problems code factories will be good enough.
decimalenough 3 hours ago [-]
> Crumbling, faulty technical infrastructure with unmanageable heaps of AI slop that no one understands
Replace "AI slop" with "legacy code", and you've got basically any large enterprise today.
The timescales are different though: it takes years if not decades to accumulate classic legacy cruft, while LLMs can spew up huge amounts of incomprehensible spaghetti in seconds.
mkozlows 5 hours ago [-]
Your PM can write the ticket, and your QA can test the output.
qsera 4 hours ago [-]
I think we would, at some point realize that the roles of PM and QA are best done by an LLM, while coding is left to humans.
So an 180 from the current coding-automation frenzy..
AdieuToLogic 3 hours ago [-]
>> How are you going to prompt the LLM or validate its output if you don't understand the requirements?
> Your PM can write the ticket, and your QA can test the output.
If your PM can write the ticket and your QA can test the output, why not cut out the middleman by having the PM write the coding agent prompt(s)?
And if your PM can write the coding agent prompt(s), why not cut out the QA group by having the PM write the verification prompt(s)?
And if your PM can write those prompts, why not cut out the PM and have your CSRs write the feature request/verification prompts?
And if your CSRs can write the feature request/verification prompts, why not cut them out and have the organization's customers write the feature request along with acceptance prompts?
And if customers can write those prompts, why would they pay you?
blackqueeriroh 1 hours ago [-]
Because customers have better things to do.
01100011 4 hours ago [-]
Also AI is now drafting design docs, generating PLC work products, entering it all in Jira, characterizing and root causing bugs... It's speeding up the 86% of my job that isn't coding. The article is a bit myopic and frankly contradicts itself.
SpicyLemonZest 6 hours ago [-]
If Claude told you to work on a task that you don't want to work on, or make a design choice that you think is wrong, would you do it? If not, then it can't really replace things like design or meetings. (Note that this is subtly but importantly different than the "vibecoding" model, where you just don't bother to supervise Claude's decisions.)
deadbabe 5 hours ago [-]
I don't know about others, but at work, the reason I only spend like 14% of my day coding is because I'm lazy, not because I'm actually doing other stuff.
hahahaa 4 hours ago [-]
Not lazy it is a taxing task like doing an exam.
mumin00 23 minutes ago [-]
one still has to think.
Also about this ai automating stuff and humans playing around.
I believe there is a time for this and time for that
cyliu 3 hours ago [-]
Some experience from my work:
- In biz development, a dev usually spends 30-40% time on coding, and more time on requirement discussion, integration testing (especially when the tests involves mobilephone or car)
- coding time can be reduced to 30%, which means reduce 20%-30% time of the full pipeline
- meanwhile, every phase and role is using LLM now, for example, product manager can produce longer requirement doc easily (we can use LLM to read it anyway:) Meeting sometimes is more than before, because more document output leads to more reading and discussion.
- I hope to find new ways to express biz requirements, in a more efficient and automatic manner.
- Shorten the requirement-dev-test-deploy loop is very important. OUTPUT is not OUTCOME. It is equal when we can see the final result, instead of intermediate metric.
- Agentic infra is extremely useful, or every one will find a way to access the database, report and ops system, in some weird fragile method.
Supermancho 7 hours ago [-]
|--------|-------|------|------|-------|------|
|Contract|Product|Design|Coding|Testing|Deploy|
Writing Code Isn't the Bottleneck, until writing code is the bottleneck, until it's not again.
pstuart 6 hours ago [-]
Getting a usable PRD is often the bottleneck.
hahahaa 4 hours ago [-]
The real bottleneck is Omega Star getting their shit together. And I ain't joking.
Exoristos 3 hours ago [-]
Or SPS Commerce. Evidence is becoming unconvincing that they're even in business anymore.
sublinear 6 hours ago [-]
You forgot to add "coordination" to that pipeline. That is easily far and away the biggest source of delays.
That includes talking to vendors, meetings with every layer of stakeholder when just one person digs their heels, etc.
That is truly the final frontier for "AI", and one that it will likely never cross. That would be when even the execs and upper management feel threatened by "AI". But, since they also delegate so much, you often see someone at the bottom of the totem pole in those meetings. This is why nobody is getting replaced by "AI". We really need to move this discussion away from the scifi stupidity already. There is no singularity or godlike AGI about to take over the world.
I hate to use awful terms like "synergy" and "teamwork", but they do have a lot more substance and truth to them than any perceived threat from "AI".
1saadcodes 1 hours ago [-]
I think the paper would have been stronger if it acknowledged how quickly the underlying evidence is becoming outdated. AI-assisted development in 2026 isn't just better models. The way many devs including myself work has changed and matured quite a bit as compared to last year
tkzed49 1 hours ago [-]
how so?
cyptus 58 minutes ago [-]
I think the tooling around the models themself has improved _a lot_ - they are really good at giving the models the correct context, even in big code bases
sevenzero 1 hours ago [-]
Many devs now work in YOLO mode letting LLMs automate their tasks.
baobeta2907 4 hours ago [-]
This is actually true at my company. They expect employees to be 10× more productive now that we have AI.
jdlshore 3 hours ago [-]
I’ve had people tell me, with a completely straight face, that they expected 10-100x productivity improvements. This is at the executive and VC level. The mania is extreme.
miraculixx 2 hours ago [-]
So then they should get 10-100x more revenue, now that AI does all the marketing and selling.
bwhiting2356 2 hours ago [-]
> A June 2025 study of Microsoft developers
A year ago feels like forever
davidpapermill 44 minutes ago [-]
Not so much due to length of time, but because of the Opus shockwave than fell within that period.
langs 5 hours ago [-]
> Myth 2: Writing Code Is the Bottleneck
Writing code is indeed the bottleneck for same resource constrained companies.
Rapid code development creates more opportunities for trial and error, providing companies with more information for decision making, that previously might have been addressed by meetings.
Of course, this might bring other problems, but it might not right to generally speaking that writing code is not a bottleneck.
davidpapermill 39 minutes ago [-]
I’m very suspicious of this objection, because when Claude first landed the same people now saying “code is not the bottleneck” were saying “the generated code doesn’t work.” Smacks of moving goalposts.
The only solid objection to “AI is going replace developers” is “AI is an accelerant.” It helps developers move faster. I haven’t seen anywhere it has fully replaced developers.
Whether this leads to a large number of job losses depends on whether you think we can increase software output by the same factor as the acceleration and still be profitable. I think we can, latent software demand is extremely high. I also think we’re nearing the limit of capability with current models.
Situation could change if more advanced models emerge, but some of the more foreseeable advances probably have compute requirements beyond today’s hardware.
zkmon 5 hours ago [-]
11-18% of time spent in coding is still very high number I think. For a large org with lots of process and risk aversion, this number could be as low as 5%. Even for 14%, the 10x improvement could mean 86+(14/10) => 87.4/100 => 12.6% overall time saved.
hahahaa 4 hours ago [-]
And time in coding is like time on highways for taxi drivers. A fairly useless metric.
osigurdson 7 hours ago [-]
It seems that this could have been expanded or contracted to any Fibonacci number of myths.
lovecg 5 hours ago [-]
> studies at Microsoft and elsewhere showing it’s closer to 14 percent
This is a depressing stat. The real productivity gains come from leaving soul sucking big tech companies where nothing gets done with any sort of urgency.
afdbcreid 5 hours ago [-]
In my open source work I believe this is the same. I don't have numbers, but I'm sure the vast majority of my time isn't spent writing code. Of course, it depends on how you define "writing code".
hahahaa 4 hours ago [-]
It is not urgency. Large production systems mean you are doing mostly unsexy operational planning. If I had a dollar each time I hear the word "data migration" I reckon I could do well.
dasil003 4 hours ago [-]
A lot of this rings true, but I think it's still too narrow. Sure, coding does not equal productivity, that is well debunked already. But I would argue that productivity is a product of engineering delivery + product decision making. Now where is the line between product and engineering? It varies by company, team and individual, but I don't think productivity can be measured for those functions independently, and in fact I see gains from AI on both the coding AND the product management side.
Basically as a senior tech lead in a large company engineering org, I don't have the bandwidth to individually validate every assertion from engineers on other teams OR from every product manager that comes with a half-baked ask. In the past I would be limited by the influence I could get through human relationships to strong SMEs with good judgment, and those folks always thin out as a company grows and calcifies. The number of creative and innovative thinkers dwindles, and the number of people protecting their turf and doing the minimum not to get fired increases. As a result many good ideas can get blocked by random gatekeeprs with poor imagination, poor expertise or both. However with AI I can follow up on gut instincts and fact check a lot more things, and ask incisive questions that can cut through a lot of organizational bullshit.
That's where I think most of the AI gains are today. Of course once AI plateaus and normalizes I think it will be baked into the org structures of tomorrow. But for now it offers real competitive advantage to those with the expertise to ask the right questions.
mellosouls 3 hours ago [-]
even an AI assist that makes coding twice as fast would, in theory, improve developers’ overall productivity by less than 15 percent. The other 85 percent of their time remains untouched
I stopped reading after this. AI has massively impacted most aspects of my non-coding work including the mentioned planning, understanding legacy code bases, setting up environments, etc etc.
Either this article is written by people with skill issues or - given the platform - its a biased and protectionist take that will fall quickly under the march of reality.
armitron 5 hours ago [-]
This reads like a critique of 2023 tooling published in 2026. Their Amdahl-style arithmetic (speed up a 14% slice, cap your gains at 14%) holds only if "AI" means autocomplete. Current frontier models do far more than that: research, code comprehension, review, test authoring, debugging, exploratory prototyping, ideation. That's most of the rest of the working day or "86%".
The only point that still holds is that organizational policies and procedures that automate AI use and lower the barrier to entry are more efficient than leaving it up to each individual. Every other point they make is either stale or was never true to begin with.
physix 5 hours ago [-]
>a “good” workday, engineers spent 18 percent of their time “coding” (not including bug fixing, testing, etc.)
I must be a crap developer, because I probably spend twice as much time bugfixing and testing than "coding". (Both of which actually involve coding stuff, so I really don't like that distinction they make)
This is stuff AI can be really good at, so brushing that part under the table distorts the picture.
Having said that, I do agree with most of the myths they present.
LAC-Tech 5 hours ago [-]
All very sensible points which I think all senior programmers who have used AI would largely to agree with.
For those more junior - keep in mind that a lot of the maximalist rhetoric are from people either selling models, or the cottage industry of people selling you courses or tools to help you use the models. Try and keep in mind software is not a mature industry, it's an immature one, and it's prone to hype and fads.
fenestella 3 hours ago [-]
[flagged]
abratabia 5 hours ago [-]
[flagged]
outoftheweed 4 hours ago [-]
[flagged]
TrustChain 6 hours ago [-]
The 14% coding time figure is one of those stats that sounds surprising until you actually track your own time. When I started building a coding agent with persistent state, I realized how some days are spent with minimal actual typing, most of it is design, reading code, debugging, problem solving, and context-switching.
But I'd push back on one thing the article implies that AI is automatically a productivity win. It's not. Some days I've shipped two months of work in a few days with AI. Other days, like today, I've burned a whole day and gotten almost nothing done because the proper research was not done by me or multiple agents.
The bottleneck for AI can be the human understanding of how to optimally use the tool. While the bottleneck for the human can be not maximizing multiple agents, or the input the user enters, then the retention of the output. If the user's input is lost, the output falters. If the user doesn't understand what the AI output is, there is going to be a problem eventually.
The article touches on adoption barriers (Myth 7), but it doesn't really get into the ego piece. There's still a wave of experienced devs who either refuse to adopt AI, or use it quietly and don't share what they're doing. That slows the whole team's learning curve. At this point, I think it's pretty much understood that you should be using AI as a dev — not to replace your skills, but to accelerate them. That means still learning new languages, still writing code, still troubleshooting. The tools change, but the craft doesn't.
I think the article is right that the real leverage is organizational, not individual. The teams that succeed with AI aren't the ones giving everyone a license — they're the ones rethinking how they review, test, and maintain code.
What I'm still uncertain about is how to measure whether AI is actually making systems better, not just faster. Lines of code is clearly a bad metric, but I haven't seen a good alternative yet. What metrics are people actually using that feel meaningful?
imrozim 4 hours ago [-]
[dead]
3 hours ago [-]
01100011 4 hours ago [-]
I'm getting tired of these articles telling me what AI will or won't do to my career when every day I see something different first hand. I'm about to stop arguing with people. If you think it's all BS then fine. Good luck.
laichzeit0 3 hours ago [-]
I don’t trust point estimates like 14%. It’s like calculating an average salary and saying it’s $120k. Completely meaningless. What does the actual distribution look like that this was pulled from? No standard deviation. Is it even symmetric? What’s the 10th and 90th percentiles? Just giving a statistic on its own tells me nothing.
This is such a weird point to make that doesn't become correct just because everyone makes it, all the time. Why clean the ocean if some magic future tech will clean them? Why save the world now if some benevolent AI is 'just around the corner' and will do it for us? And people have been making this point for years now, and it's not like my job got any easier. I just got more AI.
https://www.poetryfoundation.org/poems/51294/waiting-for-the...
And I say that as someone who uses Claude Code in complex environments almost hourly; I, as the human, still have to do the thinking as Claude still 'can't jump' [1] and I have seen no evidence that they (or similar AI, any time soon) will 'jump' like a human brain does.
[1] https://www.tomzahavy.com/files/llms-cant-jump.pdf
I think it is a great point to make, because if everyone really believed that AIs will do everything without human intervention in a handful of years, as the marketing repeats again and again (AGI, singularity, etc.) and have been saying for years... why then get bothered?
Because we DO know LLMs have their hallucinations, limitations, perform tasks not previously seen way worse than humans, etc. And it seems that, for now, there is not a good or magic solution to it, it is inherent limitations of the paradigm.
Yes, you can feed more and more and more (curated data) and eventually make AIs excellent at task X or Y, but then you spend your time specializing those engines. So the work does not really disappear, it just shifts and you make it more replicable for a bound set of problems.
Needless to say that at some point I prefer to learn (and combine with AIs, it is ok) than acritically getting inputs from something until I become totally useless.
Unless we have a paradigm for which a fully autonomous AI can do everything, this will just become improving our productivity in some ways, with all the in-between bottlenecks that it has.
Bc they are aware of the marketing and limitations. If they did believe it, then they would switch area of research.
This is such a weird point to make. We are currently ( only ) discovering that paradigm; we are not inventing anything. We found a bunch of laws that produce rather cool results but our paradigm is incomplete which leads more or less wordy or frame-rich weird stuff like hallucinations, singularity and so on ... it's childish, really and on that funny pseudo-profound, pseudo-intellectual, pseudo-spiritual ( personal opinion, if it gets you horny, you go, baby ) "universe consciousness unity, Rick James, bitch" level ...
Our bodies and minds need proper AI, not all the stuff we already outsource to middle and/or passionate men and women. Other species on the planet would certainly like to see us get augmented by AI so we can solve as many survivability issues as possible to keep as many ecosystems running long enough ... whatever that means but whether animals and plants are aware of chance and potential is another philosophical debate.
To individuals, software is a hammer and chisel, a knife, a brush and canvas, pen and paper, a reading help, and to a good amount of people it's a microscope and a fine scalpel.
To collectives, it's a tool to work on consensus and conventions, to share and gather.
It's baby steps for civilizations and it looks like our particular species is gonna get stuck in a puddle of our own monkey shit, with bottles of champagne in our hands and monkeys grinding up and down the few ivory towers in proximity.
> why then get bothered
Humans are on different levels. Most have decided that "nature realized the/a bug and wanted someone dead" or "their survival is a matter of chance" is not acceptable at all and some people decided that sabotage, poison, abuse, rape, murder are acceptable means to get chicken shit ...
The "paradigm" of life is far from explored/discovered, so we simply can't content ourselves with presumptions about inherent limitations of the LLM and AI paradigm for any other reason than to uncover ( not invent ) other parts of the paradigm.
We are happy with what AI can do for us but "AIs will do everything without human intervention" sounds weird because babies are born and the older they get and the less sabotaged ( vs influence, cultural manipulation ) they get to grow up, the more breadth and depth humans want to experience. For this they need to learn and use their hands & fingers. They need to feed body and mind to find what triggers what, and what excitement and curiosity are inherent and which can or need to be added/acquired/experienced extrinsically.
How many associations will we be able to make if AIs will do everything without human intervention?
I still have to do quite a bit of thinking but the amount of of thinking I do per task is trending down. I agree LLMs are not good at abduction but very few humans are either and very few jobs/tasks require it. I can't talk for researchers jobs though. But perhaps fewer researchers would be desired by these labs (not none).
Don't worry, I'm sure you'll hit your goal of zero thinking soon!
The IQ willl drop if we just become mechanical acritical people the same way muscles get worse if you do not exercise.
I, too, am glad that few humans seem good at abduction.
I say this as someone who's watched a bunch of woodworking videos but hasn't actually done this myself :)
And those same architects are quietly extracting real productivity from GenAI.
And even this write up skips that info by waving, “Some people…”
Btw, I think the discussion of Einstein's career in the paper you link is historically wrong in many respects, particularly the argument about 'weak signal'. Einstein was in fact working on some of the most mainstream and widely discussed problems in physics of the day, he is admired for the creativity of his solutions to those problems, and much of his work built incrementally on ideas and breakthroughs that came (long) before (as all research does).
Article suggests that a central motivation of Einstein's work was resolving action-at-a-distance in Newtonian mechanics - yet Maxwell introduced the same Lagrangian field theories for electromagnetism we use today 50 years earlier to solve the same problem for Farraday's laws of electromagnetism. Similar wave equations existed even earlier. Heaviside in 1893 extended this technique to gravity (matching 'weak field' GR) 20 years earlier. So this is perhaps the one aspect of gravity that had actually already been solved before Einstein. Authors might be conflating his work on action-at-a-distance in QM.
Einstein's GR extended the linear 'weak field' understanding of gravity to include the non-linear self-referential case where masses themselves create gravity. This was mathematically incredibly difficult but was necessary precisely because SR's mass energy equivalence created so many strong signals that were unresolved. For example: if finite energy is mass, then mass changes as objects accelerate past a large mass like a start, and hence their propagation in space could not be explained by linear EM style field equations. Many such considerations were causing very 'strong signals' in SR, and there were analogous problems in QM atomic models being developed at the same time.
SR was also a solution to a problem that was actively being worked by many of the leading physicists of the day. SR actually does match Newtonian mechanics for a single observer - it resolves contradictions in the case of separate observers, by allowing them to assign different values to the speeds, masses, etc of objects such that each object appears to follow Newtonian mechanics for each observer. Again, this was necessary because of a lot of contradictions related to the behavior of light that had been well-known for ~20 years at the time.
Personally, I don't consider this kind of reasoning to be beyond the capabilities of future LLMs (even current LLMs if the task was broken into technical rather than philosophical problems). Personally, I doubt that such problems could stand open for 20+ years waiting for a creative genius to solve them in the modern world.
And don't get me started on the philosophy.
What kind of answers were they expecting to get?
Overall, I'd like to understand those who have a positive outlook on design and software engineering as a career. Where do you see the opportunity where I just see a bleak one where anyone can do this stuff by typing or talking to AI? Myself, after 17 years in the field I am begrudingly back in school for a new medical career. As well, anytime an IT recruiter reaches out I am getting responses back only after under-cutting the hourly rate I use to demand and what others probably are still trying to get. And with it feels even bleaker as it becomes a race to the bottom!
> Overall, I'd like to understand those who have a positive outlook on design and software engineering as a career.
I think until the market better achieves some equilibrium, there is no way general software programming (sorry “engineering”) should be considered as a career. That said, there will always be opportunities in particular markets or specialties.
> > Mathematician Richard Hamming used to ask scientists in other fields "What are the most important problems in your field?" partly so he could troll them by asking "Why aren't you working on them?" and partly because getting asked this question is really useful for focusing people's attention on what matters.
> I imagine someone being asked this question, and how they should respond. I think like so - ‘Fuck off Richard’.
> This is partly because I imagine this question being asked in a kind of snarky, gotcha kind of way, with some sort of nerdy superiority. Like ‘ha your behaviour is inconsistent with your implied preferences, you idiot, do you even von Neumann–Morgenstern?’
Anyone else finding they're spending more time writing code (or at least driving agents to write code) now?
14% used to feel about right for me - I'd spend the rest of the time researching approaches and libraries, planning things out in issues, or sometimes just thinking really hard about problems I ran into.
Now... I still do those things, but I'm doing many of them faster - and I'm often doing them while my coding agents are churning away on code.
There's also this weird effect where the harder a problem is the more I can get done in parallel with it, because an agent might need to spend 20 minutes on it without my involvement.
> Anyone else finding they're spending more time writing code (or at least driving agents to write code) now?
Not really, as once it is time to write code, the problem has been defined/understood (to the degree possible with knowledge acquired at the time), and encoding it is largely an exercise in typing along with verifying assumptions via test suites.
Does GenAI quicken some portions of the above workflow? Sure, in the same way IDEs with contextual code snippet suggestions can make encoding faster.
I don't review code that either works or doesn't - most HTML and CSS layout code for example. There I test it on desktop and mobile and commit it if it works.
Ditto for stuff that's simple. A JSON endpoint that runs a SQL query and returns some JSON? If it works and a glance at the tests looks OK then I trust my agents wrote it properly.
I'm getting more confident with my judgement over what needs a close look and what doesn't over time, as so far I haven't been majorly burned my any mistakes that snuck through.
Honestly, it's similar to being an engineer on a larger team. You don't review every line of code written by every one of your coworkers.
I think this is THE issue of our time as programmers to be honest: do you review every line of code an agent writes?
An increasing number of expert programmers are moving in the direction of NOT reviewing every line. It's working out OK for a lot of them.
It's actually getting worse due to "AI code bloat", for example I have 16k lines of code to review across 3 apps by the end of this week. Normally it would be a quarter of that, but what Claude produces is extremely verbose in some places and anemic in others, and I can't tell at a glance what's right and what looks right with that much ground to cover.
That is *exactly* the sort of area I *wouldn’t* blindly trust AI, there’s a huge security boundary there. What if the AI is doing string concatenation with user-provided data???
But you could be defensive with a security checklist in agents.md and have adversarial review, if you wanted.
We don’t because everyone is accountable for his or her own mistakes. So everyone is incentivized for their recklessness to not be the root cause of some bug.
> An increasing number of expert programmers are moving in the direction of NOT reviewing every line. It's working out OK for a lot of them.
Have you ever asked your users? What about bug reports? Is the amount and rate decreasing?
I think this is missing an important detail. Lots of time was spent on non-coding stuff, because coding used to be more committal and hence expensive. With how quickly one can code up a quick prototype or even production-ready code these days, the code becomes the communication tool as well.
It's easy for a small team to adjust workflows and roles, but I just imagine the office politics must be a waking nightmare in big organisations right now.
So much has changed since late 2025 one can’t really draw any conclusions from this.
In fact, I’m guessing things will continue to move so fast that by the time one were to execute a survey of developers, many of the responses and findings are no longer relevant.
Reminds me of COVID and how everyone was fighting over early trends during that time.
https://metr.org/blog/2026-05-11-ai-usage-survey/
They also have an update -- linked from the original study! -- explaining that it's out of date and no longer reliable, and explaining why they had to cancel a follow-up study because it was understating productivity gains (but also was showing wins for the people who carried over from their previous study): https://metr.org/blog/2026-02-24-uplift-update/
The authors of this paper decided to ignore all of METR's follow-up data and discussion, and to report only the ancient number from early 2025 (a time when Windsurf was state of the art). And then, rather than apologizing for it, and caveating it as a number not to be taken seriously, they described it as a study done "recently."
That's either shockingly dishonest or incredibly out-of-touch.
No, they don't! It's easy to dispel myths when the myths are built on straw men. Dumb article.
Coding has never been expensive as it is nothing more than a reification of a solution to a problem as it is understood at that time.
It is the underlying understanding of the problem which has always been expensive and remains so.
They quote a study in which developers report to spend 11-14% of their day coding. The rest is stuff like solution design and meetings. The insinuation is that AI can at most automate 14% of your day.
The problem with this argument is that once you have code, some (not all) of the precursors to code go away.
Show me the evidence that AI has an impact on productivity when doing design work. Or reducing meeting load.
My own experience is that AI doesn't tighten the design cycle, and in fact might extend it by encouraging gold plating.
Well, I expect when you've vibed too much and lost track of the code, and can't answer questions in meetings anymore, you'll stop getting invited to them.
People vibe code because they have no clue about any of that. Not because they're slow typers.
In the case of an LLM generated demo, usually deceptively so.
> and give an impression that the design decision has been made.
In the case of a vibed design, this is the opposite of useful for the team.
This isn't exactly novel territory, here, Simon. Let's not pretend I'm asking for something strange, unprecedented, or unreasonable.
I have similar charts across my three main open source projects:
https://github.com/simonw/datasette/graphs/code-frequency
https://github.com/simonw/llm/graphs/code-frequency
https://github.com/simonw/sqlite-utils/graphs/code-frequency
The OP claims AI accelerates non-coding parts of the job, too, and so the article is misguided.
I ask for evidence.
In response you give me... code output metrics?
More code written is not a good measure of productivity. It could be garbage, or redundant code, or simply not addressing the real or more pressing needs, it could be building the wrong thing, etc.
Like TFA mentions, it's been known for decades LoC is a misleading metric for productivity. It's one of the lessons of software engineering.
In my specific case, lines of code for my published open source projects is a metric that I trust, because I have high standards for those. I have plenty of other projects where I'll accept poor quality, unreviewed code (almost all of https://github.com/simonw/tools for example), but Datasette, sqlite-utils and LLM are not that.
Of course, that's only useful for me personally and for people who trust me to stick to my own self-declared high standards!
Maybe because there is none.
One of the main quality of good codebase is simplicity. Which is about how easy for someone else to understand the code. It’s hard to define what simplicity looks like, so the best bet is to avoid the other side, making the code too complex.
And you can make the code complex by shortening variable name, doing code golfing with quicks of the platforms, so smaller LoC. You can also go the other way and increase the LoC by adding unneeded abstractions, repeating slices of code,… There’s a window where the LoC is perfect to attain simplicity, but that amount is an effect of striving for simplicity, not a cause of it. And it’s variable for every problem.
So you got something where the correct value is a different for each case. And trying to manipulate it artificially often results in complex code. And you want to say that is a good metric for productivity?
And in the cases of your projects, there are a lot more info could share that are interesting, like the amount of issues (reported or found by you) that are tied to implementation bugs (coding, libraries API breakage,…) or design issues (requirements conflicts,…), documentation improvement,… Anything that is tied to the actual usefulness of the projects, and not fumbling around with code.
Really now?
That meeting that you spent an hour in to understand the requirements? You don't need that meeting if you're not writing the code. That sync up with the QA engineer you did to hand it off to them? Don't need that meeting if you're not writing the code. That half hour you spent installing vim extensions? Don't need 'em if you don't open vim anymore.
There are engineers whose jobs go well beyond coding, of course. Staff engineers and principal engineers have had their jobs radically change because of AI, but not because it's writing all their code.
But there are also a lot of engineers -- your standard mid-level engineer, or even senior engineers at a lot of orgs with title inflation -- whose job is almost entirely about delivering code, and who spend all day either writing code or engaging in scaffolding around code-writing activities. Let's not pretend that automating away that code writing is a 15% boost.
How are you going to prompt the LLM or validate its output if you don't understand the requirements?
PMs seem pretty easy to replace. In fact, given any single role to fixate on, it's probably replaceable
Talented, experienced devs leaving the field out of disgust would be pretty disruptive.
Crumbling, faulty technical infrastructure with unmanageable heaps of AI slop that no one understands would definitely disrupt businesses.
Skill atrophy, and an entire generation of "developers" that have never actually written code will be disruptive when something breaks and no one knows how to fix it (this isn't a hypothetical, I'm seeing this phenomenon in a lot of large enterprise groups).
The total elimination of novel solutions, new ideas and new approaches to development and the resulting industry wide stagnation won't be disruptive, but will certainly be a drag.
Replace "AI slop" with "legacy code", and you've got basically any large enterprise today.
The timescales are different though: it takes years if not decades to accumulate classic legacy cruft, while LLMs can spew up huge amounts of incomprehensible spaghetti in seconds.
So an 180 from the current coding-automation frenzy..
> Your PM can write the ticket, and your QA can test the output.
If your PM can write the ticket and your QA can test the output, why not cut out the middleman by having the PM write the coding agent prompt(s)?
And if your PM can write the coding agent prompt(s), why not cut out the QA group by having the PM write the verification prompt(s)?
And if your PM can write those prompts, why not cut out the PM and have your CSRs write the feature request/verification prompts?
And if your CSRs can write the feature request/verification prompts, why not cut them out and have the organization's customers write the feature request along with acceptance prompts?
And if customers can write those prompts, why would they pay you?
- In biz development, a dev usually spends 30-40% time on coding, and more time on requirement discussion, integration testing (especially when the tests involves mobilephone or car)
- coding time can be reduced to 30%, which means reduce 20%-30% time of the full pipeline
- meanwhile, every phase and role is using LLM now, for example, product manager can produce longer requirement doc easily (we can use LLM to read it anyway:) Meeting sometimes is more than before, because more document output leads to more reading and discussion.
- I hope to find new ways to express biz requirements, in a more efficient and automatic manner.
- Shorten the requirement-dev-test-deploy loop is very important. OUTPUT is not OUTCOME. It is equal when we can see the final result, instead of intermediate metric.
- Agentic infra is extremely useful, or every one will find a way to access the database, report and ops system, in some weird fragile method.
|Contract|Product|Design|Coding|Testing|Deploy|
Writing Code Isn't the Bottleneck, until writing code is the bottleneck, until it's not again.
That includes talking to vendors, meetings with every layer of stakeholder when just one person digs their heels, etc.
That is truly the final frontier for "AI", and one that it will likely never cross. That would be when even the execs and upper management feel threatened by "AI". But, since they also delegate so much, you often see someone at the bottom of the totem pole in those meetings. This is why nobody is getting replaced by "AI". We really need to move this discussion away from the scifi stupidity already. There is no singularity or godlike AGI about to take over the world.
I hate to use awful terms like "synergy" and "teamwork", but they do have a lot more substance and truth to them than any perceived threat from "AI".
A year ago feels like forever
Writing code is indeed the bottleneck for same resource constrained companies.
Rapid code development creates more opportunities for trial and error, providing companies with more information for decision making, that previously might have been addressed by meetings.
Of course, this might bring other problems, but it might not right to generally speaking that writing code is not a bottleneck.
The only solid objection to “AI is going replace developers” is “AI is an accelerant.” It helps developers move faster. I haven’t seen anywhere it has fully replaced developers.
Whether this leads to a large number of job losses depends on whether you think we can increase software output by the same factor as the acceleration and still be profitable. I think we can, latent software demand is extremely high. I also think we’re nearing the limit of capability with current models.
Situation could change if more advanced models emerge, but some of the more foreseeable advances probably have compute requirements beyond today’s hardware.
This is a depressing stat. The real productivity gains come from leaving soul sucking big tech companies where nothing gets done with any sort of urgency.
Basically as a senior tech lead in a large company engineering org, I don't have the bandwidth to individually validate every assertion from engineers on other teams OR from every product manager that comes with a half-baked ask. In the past I would be limited by the influence I could get through human relationships to strong SMEs with good judgment, and those folks always thin out as a company grows and calcifies. The number of creative and innovative thinkers dwindles, and the number of people protecting their turf and doing the minimum not to get fired increases. As a result many good ideas can get blocked by random gatekeeprs with poor imagination, poor expertise or both. However with AI I can follow up on gut instincts and fact check a lot more things, and ask incisive questions that can cut through a lot of organizational bullshit.
That's where I think most of the AI gains are today. Of course once AI plateaus and normalizes I think it will be baked into the org structures of tomorrow. But for now it offers real competitive advantage to those with the expertise to ask the right questions.
I stopped reading after this. AI has massively impacted most aspects of my non-coding work including the mentioned planning, understanding legacy code bases, setting up environments, etc etc.
Either this article is written by people with skill issues or - given the platform - its a biased and protectionist take that will fall quickly under the march of reality.
The only point that still holds is that organizational policies and procedures that automate AI use and lower the barrier to entry are more efficient than leaving it up to each individual. Every other point they make is either stale or was never true to begin with.
I must be a crap developer, because I probably spend twice as much time bugfixing and testing than "coding". (Both of which actually involve coding stuff, so I really don't like that distinction they make)
This is stuff AI can be really good at, so brushing that part under the table distorts the picture.
Having said that, I do agree with most of the myths they present.
For those more junior - keep in mind that a lot of the maximalist rhetoric are from people either selling models, or the cottage industry of people selling you courses or tools to help you use the models. Try and keep in mind software is not a mature industry, it's an immature one, and it's prone to hype and fads.
But I'd push back on one thing the article implies that AI is automatically a productivity win. It's not. Some days I've shipped two months of work in a few days with AI. Other days, like today, I've burned a whole day and gotten almost nothing done because the proper research was not done by me or multiple agents.
The bottleneck for AI can be the human understanding of how to optimally use the tool. While the bottleneck for the human can be not maximizing multiple agents, or the input the user enters, then the retention of the output. If the user's input is lost, the output falters. If the user doesn't understand what the AI output is, there is going to be a problem eventually.
The article touches on adoption barriers (Myth 7), but it doesn't really get into the ego piece. There's still a wave of experienced devs who either refuse to adopt AI, or use it quietly and don't share what they're doing. That slows the whole team's learning curve. At this point, I think it's pretty much understood that you should be using AI as a dev — not to replace your skills, but to accelerate them. That means still learning new languages, still writing code, still troubleshooting. The tools change, but the craft doesn't.
I think the article is right that the real leverage is organizational, not individual. The teams that succeed with AI aren't the ones giving everyone a license — they're the ones rethinking how they review, test, and maintain code.
What I'm still uncertain about is how to measure whether AI is actually making systems better, not just faster. Lines of code is clearly a bad metric, but I haven't seen a good alternative yet. What metrics are people actually using that feel meaningful?