I get the point having read much the same from Tesla (and fans) regarding self driving cars that still haven't done half the things that Musk said was just around the corner pending regulators a decade ago and repeatedly since then.
And myself I keep making comparisons between AI and the progress in 90s video games where every minor improvement got called "photo realistic" and then forgotten with the next game engine: https://archive.org/details/nextgen-issue-26
So I'm not gonna say "this is it" when the software quality really matters, and I absolutely won't speak to progress (or lack of it) outside of software.
But I will say "you can look around and easily see small businesses using AI to generate posters, quite a lot of small business software and websites are in the same category: the mistakes are real but increasingly don't matter".
>the mistakes are real but increasingly don't matter...
I think it would start to matter once again. People will get fed up of AI posters and art. I think they already are...and once some threshold is crossed, the business won't dare to use AI generated assets/designs.
Turns out humans are much better at recognizing patterns in stuff that is generated ONLY using patterns from human generated content.
> People will get fed up of AI posters and art. I think they already are
Agreed, but will this look like a meme/fashion cycle? If so, re-prompt each year with a different look. Yes, there are still issues here, a friend found an image he was amazed was AI generated, but to me it was obviously so, so I showed him a screenshot of ChatGPT making something just it and included my prompt:
create image: hand drawing of cute springer spaniel puppy looking sideways, various geometric shapes drawn in layer behind and in front of the puppy, all done in style of 7 year old using crayons with mediocre colouring-in skills
As I said to them:
yeah, the line thickness feels AI, to me, the bad colouring-in scribbles feel like just the art style it was propmpted with
it's like: it gets the big picture of the composition, and it knows how to colour in badly, but it doesn't know how to draw a dog as badly as the colouring in
> Turns out humans are much better at recognizing patterns in stuff that is generated ONLY using patterns from human generated content.
We're better at recognising patterns full stop. All biological brains are, and needed to be better than the current state of the art in machine learning because if a living organism was as poor at learning patterns as the SotA in machine learning, the organism would starve to death before being able to pick up anything and eat it.
AI also has a second disadvantage, because there are so few models: the laziest of ChatGPT "thinkpiece" blog posts being everywhere is hard to miss, and 5000 fake bloggers all prompting the same model with "find biggest news story of today and write a blog post about it in a way that maximises my ad revenue" will get 5000 almost identical posts. This will remain true while each instance of the most commonly used AI fail to talk to each other in a way that at least mimics them collectively getting bored with writing the same thing 5000 times, it does not depend on e.g. quality.
It's the exact opposite. Humans will hedonically adapt to anything. If they see enough AI posters pretty soon they will start to not care, at least the majority, while some minority will disagree. We've seen this pattern many times, from data privacy and people still using social media, to using non free software.
>If they see enough AI posters pretty soon they will start to not care..
Wrong analogy. A marketing item wants you to look at it. It takes advantage of an involuntary impulse. A repetitive boring AI artwork will not trigger the impulse to look it. So it is not about "caring"....
Right now GenAI content a bad thing, cringe, a sign of low-value or lack of attention to detail. I don't want to play even a free game if I think it was made by someone else prompting an AI, and that's despite liking the output when I do the same for myself.
If AI output is normalised and becomes simply "boring" or "mundane", those negatives must have also gone away.
After that (assuming there is an "after that", I don't want to bet either way), people would, as per your argument, need to un-boring them.
>I don't want to play even a free game if I think it was made by someone else prompting an AI
I have bad news for you if you think the video game industry isn't widely embracing generative AI. Companies try not to say it because of the current backlash involved but I truly believe Tim Sweeney isn't wrong[1]. I think in 5 years it will be a quaint idea to be AI-vegan and it'll be similar to how people resisted smartphones (I was one of them) until they became inevitable.
I am aware of its use. The point is: I *am* aware, even if the creator doesn't say so.
5 years is an eternity with the current rate of change of AI. Its development is already turning into an international geopolitical issue, even though the implications for mere task-level economics have yet to settle.
Even if it wasn't, 2031 (ish) happens to be roughly when a lot of long-term exponential growth trends all happen to reach points that suggest the assumptions behind them have to break, like more than 100% of electricity being made by PV etc.; the only thing I am confident of about AI is that even its current trend line on METR were to continue it becomes physically unmeasurable sooner than that. This also means nobody will pay Tim Sweeney so much as a single Zimbabwe Dollar in royalties for games written in the Unreal Engine.
Half the game logic I played on Kongregate back in the heyday of Flash games is now one-to-few shot prompting on Claude even if you don't pay for it, and even if you don't pay for ChatGPT you can get assets at that scale pretty quickly. I know 'cause I've tried. It's just, like every blog post whose headings contain emoji, every time there's a 6-7 word paragraph a little too bombastic about how important something is, every time I read or hear "you're absolutely right" or "delve" or "nuance", and yes hear even when the lips those words pass through happen to be human…
…the way most people prompt these models, I can tell. Most people don't bother to ask it to speak like a posh victorian, or draw poster art like it's the 60s, and they don't check and reject when the shadow on the foreground of the moon base doesn't match the Earth in the sky. And because humans are lazy and greedy, so even if the visual models gets as good as the finest artists and the language models as good as Nobel laureates, we're going to learn the style of whatever the default output of those models is when prompted by lazy and greedy humans. And we'll hate it, because it's a cliché, and we avoid those like the plague.
I'm talking about AI menus and such. When I first saw them they were a novelty, and then I didn't really care about them. In other words, they were normalized, just as all AI will be. The amount of people who have strong opinions one way or another will diminish, just as it has for people who have opinions about, say, the Internet. It's so normalized that it's just background noise, and really only technologists like those on HN have issues with the Internet.
I digress with your last quote, I feel like some people who are aware of AI are starting to develop a quasi allergic reaction to slop, and while the mistakes might not matter for most of the population, for others it does and will take notice.
I’m probably illustrating your point but as a FSD fan it really got ”good enough” recently with version 14. The tipping point was suddenly, much more often than not, it can drive end to end from start (my garage) to finish (parked at destination) with no interventions. I can text and watch videos on my phone and as long as I glance up once a minute, it doesn’t complain.
Handling highway driving with lane changes was great when it got there years ago, but just in the last year or so it has gone from a nice to have to “from now on I will never buy a car that can’t do this”.
AI has hit some milestones for replacing work as well. There’s still many more to go and maybe some of them will never get hit (much like I don’t think a coast to coast drive with zero interventions during winter conditions is ever going to happen) but there are points at which it forever meaningfully changes some field of work. I think it’s there for writing code.
> I’m probably illustrating your point but as a FSD fan
Half-and-half. I'm not denying that self driving cars (and LLMs) are improving, I'm comparing it against the standards set by the biggest proponents. But yes, I have heard basically the same thing you just wrote for the previous several major releases of FSD.
Where we agree is that, while you are a fan, you do explicitly give as an example of something you think it will never do, something very close to what Musk has promised:
"Ultimately you'll be able to summon your car anywhere … your car can get to you. I think that within two years, you'll be able to summon your car from across the country. It will meet you wherever your phone is … and it will just automatically charge itself along the entire journey."
(That said, I think Tesla's FSD will never get there, not that it's impossible. The way Musk is behaving, there's going to be a financial scheme named after him in whatever passes for a textbook in 20 years, and it won't be the positive kind of example).
What matters is the benefit it affords to drivers, not what "biggest proponents" say. Elon got overly excited in 2016, but for my day to day usage, I don't need it to self-charge on a drive across the country to get me. I need it to safely drive my family on my daily errands or weekend trips, which I now consider solved.
I'm mentally prepared for the next US administration to exact retribution on his companies, and I expect FSD will be neutered after that. Hopefully other car companies are able to catch up. I'm more of a self-driving fan than a Tesla fan, so as long as the thing works as well as what I have now, I'll be fine with it.
Yes. He was before and remains so to this day, but he did so then, too.
> I need it to safely drive my family on my daily errands or weekend trips, which I now consider solved.
This is probably unwise. As with LLMs, the statistics suggest a spikiness in the intelligence, with it being mostly good but also sometimes still making some very odd mistakes that humans would essentially never make.
As with your other comment, you know that if it gets into a crash you're responsible; while you consider this a win, I suggest waiting until the company you buy the car from (in this case Tesla) puts their money where their mouth is on quality and takes liability for crashes due to the AI upon themselves.
> Hopefully other car companies are able to catch up.
From the stats I've seen, they're much closer to the goal, relatively smoother/less spiky all-round driving intelligence. They may not be as impressive at their best, but when they fail the failure modes are themselves much safer.
The data shows that FSD (supervised) is safer than human alone driving. Given that I value the well-being of my family, FSD (supervised) it is as often as possible. Separately from the data, it is just so nice to hit a button and sit back and relax for the entirety of the drive.
The usual HN nits at this point are that the data is unreliable/biased and that the way I’m using it isn’t supervised enough. I’ll admit to the latter. In my experience the mistakes tend to be navigational but I’ve seen a few of the “nearly drove through the lane closed gate” videos so I definitely keep a closer watch when there is complicated highway stuff going on. I also have my foot at the ready for when I go through the gate to my community, similar failure mode and about 1% of the time it forgets to wait for the gate to close and reopen.
I look forward to the other car companies catching up, competition and more options are good. Elon is a loose cannon, so I need alternatives if Tesla ceases to be.
I urge caution when examining statistics, as even with the best will in the world… well, the Datasaurus dozen was created for good reason: https://en.wikipedia.org/wiki/Datasaurus_dozen
The lesson from the Datasaurus should be to ask detailed questions, e.g.:
• exposure bias: is the AI used more or less in certain conditions, on certain kinds of roads
• sharpshooter: shouldn't need to ask this one of a public company, but given regular headlines: how many crashes happen a few seconds after it switches itself off?
• driver demographics: most severe accidents by humans are due to impairment or being a new driver, is the AI actually better than an experienced-and-not-impaired driver?
• the nature of the failure modes: as a cyclist I've been hit by a car that stopped at a junction and pulled out into me, this was not fun but no serious injury; this is very different to any of the highway speed fatalities.
> Separately from the data, it is just so nice to hit a button and sit back and relax for the entirety of the drive.
No doubt. Myself, this is why I have been happy to live in a city with a good public transport network and use that instead of driving.
> I look forward to the other car companies catching up
My point is that they are in some important senses already ahead. Less optimised for the headline, more optimised for the real dream.
Good points about the stats. I know the answers to some of those questions, but at some point it's so hard to get apples to apples comparisons for these things. So I did a "paranoid super-supervised" usage of FSD for a year or two to learn its failure modes and get used to its reliability or lack thereof. Then combined with the (allegedly truthful) FSD stats, I came to my own conclusion which is that FSD is ready for my everyday needs. (nit: FSD can change version to version and potentially regress. some temporary paranoia is justified after major changes)
Usable public transport would be nice. It was unusable for various reasons where I used to live (Seattle/Bellevue) and is pretty much non-existent where I currently live (Socal / Orange County), but everything else about the location is perfect for me, so I'm willing to give that up.
> they are in some important senses already ahead. Less optimised for the headline, more optimised for the real dream.
Nice! I look forward to trying them out when they are available and working here. Rivian claims they'll achieve it next year, and one of their form factors (R1S) is perfect for my family so they'll be my next option if things go downhill with my Teslas.
I would deeply implore you to not put to much faith in this technology as there are many very dangerous edge cases, including as I personally experienced, jumping red lights after coming to a full stop: https://news.ycombinator.com/item?id=48928102
The fact of the matter is that it drives reliably end to end. I don't care who's responsible or what it's called. I only care about the benefit it provides my family, which is immense. If it gets into a crash I know I'm responsible, but given that I consider it a better driver than I am now, this is an improvement over before.
People's right to make stupid decisions end when they reach public roads and/or borrow children to make a point. That's why we have things like driver's licences.
Yes, I'll continue stupidly looking at data and putting the safety of my family in the hands of the safer option. If only I could be smart enough to use emotion-based politics to make those decisions instead.
I just spoke to a fried who is a headhunter and who's been trying to automate his processes for a while (he likes to fiddle and certainly has skills, but he's not an engineer). He kept trying, but it just wasn't good enough.
Now he said with GPT Work and Sol, it worked, but the key point is: all of it suddenly worked.
The problem was one of reliability, of handling edge cases. All previous attempts / model-harness-combinations were too brittle and needed too much observation and fiddling - cheaper to do it yourself.
Now he says "I don't know why I would ever hire a recruiter [the folks doing the cold outreach] again. I can focus on the candidate screening and acquiring projects, everything else is fully automated".
This doesn't come from an engineer or an AI lab, but a technically inclined power user, and I think this is where things get interesting.
You heard it from someone with no experience developing software. A lot of AI hype comes from that, even (or especially) from people in the actual business of developing software - a surprising amount of managers and adjacent or supporting roles in the field actually have very little clue about software development.
It's cool that 'regular' people can now create solutions to many small problems, and automate stuff - genuinely a step forward. Like Excel, only vastly better. But for bigger projects, real software engineers know that what LLMs do today is only a tiny part of development. And it solves it in a way that might well make the rest of the lifecycle a lot harder. It's like that saying about tools that make easy things easier and hard things impossible.
IMO it doesn’t replace software engineers, but it certainly changes the job. The problem lies with VPs and executives that rose to their positions through managing phones and printers and shit who don’t appreciate all of the work that isn’t “write code”, and those people are the majority outside of big tech in my experience.
Currently in a re-org justified by AI, AI changing the roles people will need to play. It’s disturbing how much content in the materials about the new org structure and roles and whatnot is clearly ai generated and contradictory. We’re laying off about half of 500 people.
At my employer (a reasonably well known company) every internal communication appears to be AI generated, from emails to slide decks.
It gets better: The internal AI gateway chat thingy where you can ask questions has AI autocomplete that pops up after your write more than 10 characters. What the fuck!?
Who are these "real" software engineers and what are they working on? Among software developers I know, the biggest users of LLMs are the seniors. I don't think any of them would agree that its usage is a tiny part of development.
If you're a high-end executive, people would not approach you directly but try to find a connection to you (which also serve as references).
For all other roles the first few messages are similar: this is <role X> with key challenges a,b,c. You seem to be a good fit because of d,e,f. Would you be open to explore this? This requires relocating to <place>.
Traditionally, this was done by entry-level people. In either case, this isn't the person who will jump on a call with you.
"Desperate" and "clueless" seem very strong words here in reference to someone who gets actively approached from a recruiter.
Then it just becomes a new baseline (everyone have access to the same LLMs), and recruiting moves up the philosophical ladder where human can add more value. What will it be? I don't know, I'm not a recruiter.
Again I've heard this since 2022 when gpt3.5 came out.
This is like microprocessors in the 80s. Sure they double in capability every 18 months but the start is so pathetic it will be 30 years before they are good enough for everyday tasks.
CPUs in the 90s were amazing! They were over-specced for "everyday tasks." Our problem is that we overbuilt CPUs too much, so software is now written with ten unnecessary layers of abstraction because there's no real reason to simplify.
3gb of ram aren't enough everyday tasks at any sane resolution. That we still don't have 1200 ppi desktop monitors is as stupid as using black and white screens in 2000.
I am not an expert in book printing from the 1800s. That being said, I don't think they had tolerances anywhere near 1200ppi. You'd need extremely smooth paper with almost no grain for that level of resolution to be possible at all, and you would need to distribute the ink so that it did not smear or bleed at all. Older books I've seen don't seem to possess those qualities.
It seems to be true this time though; I have observed it myself and heard it from several experienced developers I personally know and respect. It feels like some threshold was crossed with Opus 4.5 and Gpt 5.3, where the models are now able to reliably solve certain classes of problems that were previously unreliable.
Time will tell of course, and it’s early, but inflection points do exist with progress.
Thing is. You can find an extremely similar paragraph written about Claude 4.x or some equivalent gpt. And simultaneously, many people expressing their frustration and the shortcomings of <insert any model>
“But it’s different this time” - several people, several times over the last couple of years.
This is not at all a dig at you, I’m very sorry if it reads that way. My point is these things only get truly better in anecdotes. The ways in which they fail is yet to change. Just yesterday I had gpt 5.3 generate completely awful code for the Cinema 4D Python API. Also an anecdote. But for all of the people saying they are truly intelligent and truly reason, they still make obvious mistakes, write around problems, fail entirely at architectural decisions, fail at random, generate FAR too much code.
And no amount of harnesses, methodologies, loops make much of a difference. If you listen to people on the internet they say it’s all working. You listen to people on the job and they mostly say it’s creating tech debt and a review bottleneck. Also burnout, so much burnout.
I think LLMs are mediocre. I think it’s fine they’re mediocre. You can work with low expectations. But the hype cycles are so tiresome.
I don't buy into the huge LLM hype but I certainly think late 2025 was the inflection point. Up until then, I thought LLMs universally sucked at code. GPT-3.5, GPT-4, o1, o3, Claude 3.5, Sonnet 4, the whole bunch. Each iteration got marginally better, hallucinated APIs less and so on, but my overall evaluation of them all was that they wrote crap code and were unsuitable for anything other than one-off scripts.
Then the incremental improvements did, in my experience, cross some kind of threshold in late 2025 where the things became useful. It is of course anecdotal and personal judgment. But I asked LLMs to implement a small feature in my codebase (my usual test) and finally it produced code I was happy with. They've also been able to locate and diagnose a problem based on logs. In my view it's now a markedly different level of capability than we had a year ago, though I would call the previous two years equally useless.
You can find an extremely similar paragraph written about Claude 4.x or some equivalent gpt
Yes, as a product gradually improves there will always be many people for whom version X didn't work well for them and version X+1 does. It turns out that Opus 4.5 and GPT 5.1 were larger than average improvements that cross that threshold for a significant number of people.
My point is these things only get truly better in anecdotes. The ways in which they fail is yet to change.
If your claim is that there's no substantive difference between Sonnet 3.5 and Fable, then we live in very different worlds.
LLM capabilities are spiky. They're amazing at some things, and poor at others. Over time, the set of things they're amazing at has grown, while the set of things they're poor at has shrunk. If you think LLMs are just "mediocre" without any nuance, that's a sign you haven't spent the time to evaluate them in order to make an informed opinion.
> You can find an extremely similar paragraph written about Claude 4.x or some equivalent gpt
Even 3.7. I remember when it came out and people were claiming that that was now the model that was going to replace engineers. Cue Fable years later and people still claim that this one is the one.
Not shape, but even with napkin math, you can work out the conclusion. Take email vs snail mail. Online ordering vs relying on catalog. The spreadsheet, the word processor, the various CAD software.
The key thing is that it's easy to contrast the old way vs the new way and the evidence become obvious.
The thing with LLM tooling is that they're not reliable. I can do fine with risks, but only when there's a way to manage it so that if the disaster happens, it's practically a black swan event.
Typing more code or solving one task has never been the core problem. The core problem has always been to encode a whole system into the computer AND then provide a control interface for it. It requires both an understanding of the system you want to encode (especially how it behaves over time) and empathy to know what would be the best control interface for the users.
That understanding does not rely on the amount of code, and the best control is found through communication.
An understanding of the system could be the following: A conference schedule consisting of events (time, place, speaker, description,...) stored or presented in some format. The interface would be: A web app with a mobile first UI that presents the information in an accessible manner (highlighting, filtering, exports,...).
A relatively quick (I haven't tested it), would have been to open the web inspector and extract the data using the dom API (requires knowledge of the dom api and a desktop browser), put the data into some json or a tsv file, then write a php script or a python script and then serve that. The interface could have been built with the standard elements of some css framework (bulma?).
Not saying the above is better. But the thing is that is doable from even a raspberry pi. And more it's repeatable and extensible. And the individual piece of knowledge are reusable in different situation.
Con: "Where's the evidence that this actually is improving productivity?"
Pro: "The evidence hasn't shown up in statistics yet; it's too new!"
Con: "And won't this destroy maintainability?"
Pro: "Show me the maintainability disasters caused by AI."
Con: "I can't yet; it's too new!"
Both sides are playing the "it's too new" card when asked for actual evidence to prove their claims. In fairness, it actually is too new for there to be much statistically-valid data, especially if the inflection point was November 2025. So both sides are trumpeting their position, neither with actual trustworthy data.
Everybody has their anecdote. Nobody has data yet.
Okay thanks. But what I'm saying is all of these anecdotes should add up to something measurable if there's something to the idea that an inflection point was reached in November 2025, right? There's a lot of marketing and hype and motivated reasoning going on, so that's why I don't trust broadly-reported notions.
Exactly, "productivity" and similarly "efficiency" are meaningless weasel word unless they're qualified and measured against something. So which measurable things? Let's start with just anything really because then we could ground the conversation in something real rather than just feelings of being more productive, which can be deceptive.
AnimalMuppet in their reply to my OP comment makes the point that these things haven't been around long enough to measure end-to-end productivity gains and come to a conclusion either way, and I agree with that. But we can still at least be measuring something.
For example, in my case AI has allowed me to write 10x more LOC than I usually would in a similar amount of time. But having to review it all, I've also deployed 1/4 the number of releases I normally would in the same period. By one measure I'm more productive, by another I'm less productive.
People could claim to be more productive by skipping the review. But in that case did AI make you more productive or did you lower standards? People could say they're using AI to do the review but is the impact of that being measured and has that caused more or fewer bugs? If more bugs, has the time to fix those been factored into overall productivity? In my experience it's common for people to eagerly count immediate productivity gains and discount long-term productivity sinks.
For this reason I think case studies are the best convincing thing, because they properly contextualize the usage and consider a longer-term window. They're also backwards looking instead of in-the-moment, so have the benefit of hindsight. But they're harder to come by and we probably won't see any meaningful case studies for a thing that people say happened in November.
But the very least people can be doing is just defining what they mean when they say "productivity" because otherwise everyone is talking past one another.
I really don't think that's how it works. Smart, experienced developers who thought coding agents were junk for most of 2025 and think they're useful now in 2026 are not saying that because they got "worn down over time".
It tends to be when the training data wanders into their area of expertise temporarily and they go “OMG, they hype is real. I was so wrong” and then a few releases later they’re on the train and furious that the skills in their domain space have not just stopped improving, but regressed. Cue someone else in a different part of the world starting the same cycle.
Meanwhile the guy who leaned in a year ago and gave up reading the output is beginning to see work grind to a halt and throwing more agents at it is increasingly not working.
You can see these tropes all over social media near constantly.
But in all seriousness - you can even bring up Pope himself. Don't care. Show me data, show me the leaps our software made with all this 100x productivity boost. Show me a myriad of better LLVM projects, new usable kernels, and so on. Show me sharp decline in bugs and defects in existing projects.
This is entirely my point. Observing reality, you can see how little impact any of this is truly having. Decision making, communication and alignment is not helped at all by LLMs, it’s just leaving developers with more laundry.
I didn’t start using Claude Code until late 2025. Prior to that I would use ChatGPT to give me snippets of code but I was still doing most of the actual code writing. Coworkers told me in late 2025 about how they hadn’t written a line of code in “months” and just use Claude Code/agentic “whatever” so I tried out Claude Code and was pleasantly surprised. It is passable to have entire apps written by LLMs (I’ve made several that I otherwise never would have had the time to create by hand), but I wouldn’t say maintainable or easily extendable. It’s hard to be specific, but there’s something about LLM code that doesn’t look “natural”, and I’m not talking about the excessive use of comments in code. The code itself is unnatural. Functional, but unnatural. I wouldn’t want to suddenly lose LLMs and have to read through and understand and continue enhancing a codebase created by an LLM.
For me it feels a lot like generated images or video. I've made lots of things now, but those that are 100% LLM written "work" but are uncanny, weird and the details are wrong everywhere you care to look in detail.
I think how I'd put it is like, it's a deterministic common denominator thing iterated on, but it's not really intuitive, it's not really "what makes sense" but rather "what would x look like", similar to how you point out it is when it comes to image generation
it's useful for scaffolding but after that I'm not sure how you could rely on it without being in the loop and directing how the code should be like
This has been my experience has well. For code generation, you have to really constrain these LLMs on your coding style, design goals, test cases, and overall expectations. I mainly use these tools to help my understanding of the code and to generate code for very specific problems. Even after all that setup and careful review I'd say it's still a net big speed up for certain software engineering tasks.
Jason Turner gave an excellent talk at last year's CppCon explain how he thinks tools can be used to make generative AI coding assistance safer and more productive. https://www.youtube.com/watch?v=xCuRUjxT5L8
> November was, for me and many others in tech, a great surprise. Before, A.I. coding tools were often useful, but halting and clumsy. Now, the bot can run for a full hour and make whole, designed websites and apps that may be flawed, but credible. I spent an entire session of therapy talking about it.
> The real annoying thing about Opus 4.6/Codex 5.3 is that it’s impossible to publicly say “Opus 4.5 (and the models that came after it) are an order of magnitude better than coding LLMs released just months before it” without sounding like an AI hype booster clickbaiting, but it’s the counterintuitive truth to my personal frustration. [...] A year ago, I was one of those skeptics who was very suspicious of the agentic hype.
> Six months ago, in an episode of the Lex Fridman podcast, David shared how he doesn’t use AI tools to write code: he types out all his code. But things have changed a lot since then.
> In this episode, we discuss his approach to building software, how it’s changed in the last six months, and why he now takes an agent-first approach, and how he barely writes any code by hand.
> There are other questions around AI (like what the economy of it will actually look like in the end), but "is it useful" is no longer one of those questions. Anybody who doubts that clearly hasn't actually used it.
> Shock! Shock! I learned yesterday that an open problem I'd been working on for several weeks had just been solved by Claude Opus 4.6 - Anthropic's hybrid reasoning model that had been released three weeks earlier! It seems that I'll have to revise my opinions about "generative AI" one of these days. What a joy it is to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction and creative problem solving.
Okay so you have no names of experienced developers saying that coding agents are now useful when they weren't before. The only two you have actually saying that are... not working as software engineers.
> 2025 was a weird year for me. If you had asked me exactly a year ago, I would have said I thought LLMs were amusing toys but inappropriate for real software development. [...]
Today, I would say that about 90% of my code is authored by Claude Code.
> I don't think you are arguing in good faith here.
Funny coming from a guy using Linus saying "There's no doubt AI is useful today" and interpreting that as "Yeah last year AI coding agents were garbage but now they're really useful".
There's no denying AI is useful. I use AI, my project at work is all about using LLMs. They're even useful for coding, when I don't give a fuck about the quality of the output (which is almost never in a work scenario).
But to claim that there has been some magical shift late last year is simply laughable. Anyone working as a software engineer and claims that is either lying to themselves or simply bad.
Oh and don't try with the "you're not prompting it right".
I was getting useful coding work done with GPT 3.5. I think devs saying "the models are finally good enough" this year are just trying to save face from their own previous irrational denials.
Useful, yes, sometimes, but it wasn't fully automated "Here's our JIRA board URL, fix everything that's rated 1-3 story points and in the current sprint".
Nobody was saying coding agents started working in 2023 or 2024, because the category was defined by Claude Code which was first released in February 2025.
I would say that Aider is what defined coding agents. That was at least multiple months before Claude code. I remember seeing a coworker use aider for a hackathon project Adeline November-December 2024 , and it was already decent and pretty close to the DX we consider coding agents to have
Devin and Cline were the original true "agents" IIRC. Aider was the original terminal "agent" but it was't a true agent, it was sort of a hybrid agentic chat, since it had limited recursion ability (3 turns by default) and you had to configure it carefully to make it consistently take multiple turns in a row.
Claude 4.5 was it (nov 2025?), without a doubt. It went from frequent hallucinations to highly usable with much less garbage output. If you were making demos of AI tools around this time your demo/pitch/product was saved and you probably looked like a genius.
Since so many people are doubting you here, here is a post from ~a year ago that's pulling the same "LLMs 6 months ago were crap, now they're awesome" shtick: https://fly.io/blog/youre-all-nuts/. There's more posts along this vein being put out from 2024 on or so.
The idea that "yeah, you said technology had improved in the past, and now you're saying it has improved again" is a gotcha just seems incoherent to me.
The context for some of these posts are yet-another-null-result coming out, saying that AI isn't having the expected objective result, with the posts defending AI by saying that it totally changed just after the cut-off the research. Evolutionary step changes are definitely a thing, but the problem isn't that AI is having such step changes, it's that you're saying that the most recent step change is the one that shows the improvement, and every study that comes out saying "no, that wasn't," the response to that is, "no, there's a new change that is totally the key inflection point!"
Or, put differently: if a study comes out next year saying they don't see major impacts from AI in 2026, will you admit your viewpoint as being wrong, or is your response going to be "no, there was a massive inflection point in September 2026 that completely invalidates the paper"?
I think people are, rightly or wrongly, identifying goalpost moving re: coding agents are only just this year working well enough to impact work. I'm reminded of a quote I read last year from one of my favorite writers, Noah Smith:
> The debate over whether AI is taking people’s jobs may or may not last forever. If AI takes a lot of people’s jobs, the debate will end because one side will have clearly won. But if AI doesn’t take a lot of people’s jobs, then the debate will never be resolved, because there will be a bunch of people who will still go around saying that it’s about to take everyone’s job. Sometimes those people will find some subset of workers whose employment prospects are looking weaker than others, and claim that this is the beginning of the great AI job destruction wave. And who will be able to prove them wrong?
> re: coding agents are only just this year working well enough to impact work
Or, more likely, since software touches every single industry of man, you're seeing AI slowly able to handle different types of work. People in the types of work that early models struggled with are now enabled. The goal post didn't change for the people in each group. It's not one entity.
The last people to say that their work will be impacted are those that work in areas with the most novel ideas/innovation, and/or working on things where libraries/examples don't exist.
For example, I work in test/manufacturing, mostly with robots. The whole industry is proprietary. There are basically no open source libraries/code for this stuff, so claude is still pretty terrible at it, but, with Opus 5, it is able to now do some of the work!
I don't think we disagree on the capabilities of the latest agents, but my point is that the software industry has already been through multiple rounds of "the [models,agents,bots] only just became good enough to impact work in the last few months," yet this and several other studies have shown that the clankers still haven't taken any jobs. If the November timeline doesn't show an impact, then it must be because we only just got access to Fable and Solus in July. And if the Fable and Solus timeline don't show an impact, then it's because of this other thing that happened where models really took off – the data just needs to catch up.
Everyone has an anecdote about someone being laid off because of AI, and everyone has anxiety about the next model being the big one that will really cause the industry to implode, but there's never any hard labor impact. It's always still coming, still waiting for the next model, next study, next jobs report. As Smith said, people who believe that AI is going to take jobs will never be convinced that it's not coming eventually.
You miss my point. There is no single software industry. There are people solving completely unrelated problems with software. Some of those problem spaces could be services with old models. Some only with new. It's a rise in tide, with little towns completely drowning, and the ones up the hill still safe for now.
A year ago claude could NOT write code for the robot controller we use that was usable. Now it can. The water is now ankle deep for where I am.
I keep running into this as well. Someone shared a link to a study from last year telling me that AI doesn't make programmers as productive as they think. The study was obviously done months prior to it's release. So evaluating the state AI even further back.
But this didn't seem to concern them. The study said X, therefore it applies to today.
I'm just not sure it's worth it to argue with others about it at this point. Not that I'm 100% all behind AI coding, but I'm just shocked people are still this resistant.
> I'm just not sure it's worth it to argue with others about it at this point.
This is the correct response. Let closed-minded people do their thing. Makes them less competitive against you. There's nothing for you to gain by trying to help them understand what they are missing.
A year ago it was capable but required much more hand-holding. I thought it was as good as it would ever get and was still happy and productive, but only because I didn't know how much better it would get.
I'm just now seeing this because I periodically read all your comments, but, as the author of this post: what was your point here? This isn't a "schtick" from me; I wrote it once, after observing exactly the one thing the post observes (that critics of coding LLMs were premising their argument on things that were no longer true, or in some cases had never been true), and the subsequent year largely bore those observations out.
I didn't then create a side hustle for repeatedly re-observing that thing.
The quality of the output is so variable. It depends on the model, “effort level”, prompting, probably even the programming language/app functionality, and libraries involved. For example, I find LLMs are best at making simple web apps. These web apps, while simple, would still take a senior engineer perhaps a week or two to create, but LLMs can spit them out inside of an hour. Conversely, LLMs struggle with things like Docker or local model stuff. Parallelization of code is a mixed bag. In these areas I think it often would have been faster for me to write the thing by hand.
I'll give a recent example: the past month I had been using Codex cloud and it was working great writing Rust modules. Suddenly last week they made a few changes and everything went to shit, couldn't get a single good result out of it. Same text box, same prompting, same model, totally different results because they changed the cloud tool's resource limits.
So whereas before my debugging time would have been spent in Rust docs looking up traits and such, these days I feel more like an AI therapist trying to figure out why it's not feeling up to task on any particular day. It can be anything from regular service outages to geopolitics that on any given day my workflow is fucked up.
In my before-AI workflow, I was never restricted from compiling Rust code because of concerns that Cargo is a national security threat. So you really have to broadly scope the notion of "reliability" with these AI tools; it's much larger than whether it can give a good output but whether it can do so consistently enough to depend on.
"unless you're content with producing low quality work." - With the right guiding hand, it is a productivity multiplier without compromising quality. As a fully autonomous developer, it is a disaster.
> With the right guiding hand, it is a productivity multiplier without compromising quality
This just reads like another variation of “it’s the user not the tool,” which is just endless runway for always blaming people and never acknowledging the limitations of LLM’s.
I’d be curious to hear how the recipients of your work enabled by the “productivity multiplier” feel about the quality.
You can't play an entire orchestra's sheet music on a single guitar either, but your playing ability still matters a lot.
I would say that as of July 2026, with the right scaffolding, you can get reasonably good output out of a LLM, or better a combination of LLMs. For example, it pays off to prepare an implementation plan with one LLM and then let another LLM check it for flaws, then again. After several iterations like this, you will have a plan better than whatever you could come up with yourself.
It often is the user and not the tool. LLMs are complicated, have nontrivial failure modes, and the user needs to steer them carefully. They might be the most complicated tools on the planet right now.
Anecdotally, the recipients of my work have become visibly more happy in the last months. LLMs are great at diagnosing subtle problems which tend to appear at Friday night only, and this is the sort of problem that bugs actual people the most.
> I would say that as of July 2026, with the right scaffolding, you can get reasonably good output out of a LLM, or better a combination of LLMs.
Totally agree, I don’t think I said or implied otherwise.
And yes can it can be the user and often even is, but when it comes to any LLM conversation I’ve been a part of it seems people think the only answer is “you’re using it wrong.” Evangelists swear it’s a 100x multiplier and anything counter to that means you’re either a Luddite who is blinded by politics or are too dumb to use the tool.
I don't think it is "you are using it wrong". It is just that originally mediocre programmers still generate mediocre output (or worse sometimes) faster and originally good programmers generate good quality output faster. LLMs have not changed that yet.
How are junior devs becoming qualified “guiding hands” these days? If the expert with LLM assistance is multiplied, what’s a company’s incentive to pay for a junior, and how would they train to get good in these conditions?
I agree on this one. Junior developers are not yet good guiding hands yet.
But for the motivated junior developers LLMs are a great onboarding and learning tool. A good senior engineer must still guide, and there are very large gaps on how it's being done
Juniors are still useful, but not as pure "coders". The juniors you want to hire now can wear product, engineering and QA hats, and are self starters with good attention to detail and AI skills. These types of people are actually a bargain.
I produce code that is significantly higher quality with the assistance of coding agents, because I no longer succumb to the temptation to cut corners due to lack of time.
One example: everything I do is properly tested and documented now, even the most trivial of changes. Previously I would have weighed those tradeoffs and sometimes decided not to bother with the tests because they weren't worth the time.
> because I no longer succumb to the temptation to cut corners due to lack of time
No offense, but that says much more about the way you approach programming than about the quality of LLM outputs.
In my experience, LLMs are the ultimate corner-cutting tool. With LLMs, I now succumb to the temptation to cut corners, build something I haven't properly researched and don't fully understand, prioritize shipping quantity over quality.
Without LLMs, I have to understand the domain and the tools and ultimately my full solution (with all its warts and limitations). When I really care about the project and consider it "my baby", LLMs are out of the picture.
Software engineering is about tradeoffs. The single most important tradeoff is always time: this feature would be better if I did X, but doing X would take me Y hours, is it worth the time or should I spend that time on something else?
Not having to type in and then iterate on the code manually has a material effect on those time calculations.
I absolutely agree that you need to understand the domain and code and tools. That's what I want to spend my human time on - not typing in the code and sweating over every line of syntax.
I agree, good software engineering is largely about making good trade-offs.
However, I completely disagree with your characterization of LLM coding as saving you "typing in the code and sweating over every line of syntax".
Claude will happily architect complex layers of OOP, multithreading, SIMD (a "personal favorite" of both Claude and ChatGPT).
If you don't immerse yourself in the details, you will never know if you made the correct engineering trade-offs, let alone whether the various "optimizations" and clever solutions by the LLM actually work.
I think you're referring more to product strategy trade-offs than engineering trade-offs here; so yeah, getting something you don't understand out the door is often the fastest path to sales. This is also how we get so much low quality and unnecessary software out there (it was true before LLM coding, just amplified now).
I think a reasonable tradeoff, in theory, would be: use the LLM to help you explore the space of solution, then dive into the details and choose the path that feels right. However, I highly doubt most people do that. It's also easy to let the LLM's first solution influence your understanding of the problem, so that you will never truly explore novel solutions that fall outside of the most likely token sequence prediction.
Oh and LLMs when properly constrained can also do pretty good "auto-complete on steroids", which can definitely save you typing a lot of boilerplate. If that's how you use them, then your characterization is correct.
> I think a reasonable tradeoff, in theory, would be: use the LLM to help you explore the space of solution, then dive into the details and choose the path that feels right.
Oh absolutely. One of my favorite prompting patterns is "suggest options for X", followed by serval rounds back and forth to discuss tradeoffs and maybe introduce new ideas into the mix.
I also use LLMs to build exploratory prototypes. I'm working on a project today which I've had different coding agents build a total of four times already. Those explorations gave me the confidence to dictate a finished solution that I'm sure will work well against all sorts of future cases.
Your perspective is one that assumes a single entity is using AI, rather than disparate groups of people working on disparate problem spaces/topics, each with different "intelligence" thresholds for them to say "good enough". The goal post isn't changing, it's that there are multiple problem spaces with completely different, fixed, goalposts.
And even then, within a single group, you'll have multiple thresholds, of "this really helps make my coding more productive" to "I no longer type code, just review" to eventually "I'm no longer employed".
Here's the rub. They produce low quality work according to your rubric. If that was the universal rubric, they would already have RL'd against it, and you'd like the work they produce.
I wouldn't say it like that exactly, though you're not wrong either. People just heavily discount the value of future efficiency gains due to nuanced architectural/stylistic choices. I think this is in part because people have a limited expectation that the code will survive long enough for it to matter, or that they'll be at the company long enough to reap the rewards of the up front work.
The timescale is well established: Late '25 was the start of agentic ai when capabilities of model + api + scaffold reached autonomous state. Any study comapring events before that timeframe is comparing apples with oranges.
I haven't. Around the start of 2026 is pretty widely mentioned as when they went from "this is broken slop" to "huh this is actually 90% what I would have written", which matches my experience.
The capex is still "early" for sure (i.e. data centers are still being planned and built out, we're hardware/energy constrained).
If model scaling holds out, we're "early-ish" in terms of the reliability and performance of these systems, just based on utilization of the compute from the planned capex. If we hit hard diminishing returns and we don't find architectural/data workarounds, that would put a wrinkle in things, but I suspect that the AI we have now is capable of helping us find those workarounds and keep things moving.
2026? 5? 4? 3?
Heard this one way too many times.