Eight more months of agents
crawshaw.io
crawshaw.io
Where are all the new houses? I admit I am not a bleeding edge seeker when it comes to software consumption, but surely a 10x increase in the industry output would be noticeable to anyone?
You can't saw faster than the wood arrives. Also the layout of the whole job site is now wrong and the council approvals were the actual bottleneck to how many houses could be built in the first place... :/
Coding speed was never really a bottleneck anywhere I have worked - it’s all the processes around it that take the most time and AI doesn’t help that much there.
AI reduced this from a 5-day process to a 4.9-day process
If I may hijack your analogy, it would be like if all the construction crews got really fast at their work, so much so that the city decided to go for an “iterative construction” strategy because, in isolation, the cost of one team trying different designs on-site until they hit on one they liked became very small compared to the cost of getting city planners and civil engineers involved up-front. But what wasn’t considered was the rework multiplier effect that comes into play when the people building the water, sewage, electricity, telephones, roads, etc. are all repeatedly tweaking designs with minimal coordination amongst each other. So then those tweaks keep inducing additional design tweaks and rework on adjacent contractors because none of these design changes happen in a vacuum. Next thing you know all the houses are built but now need to be rewired because the electricity panel is designed for a different mains voltage from the drop and also it’s in the wrong part of the house because of a late change from overhead lines in the alleys to underground lines below the street.
Many have observed that coding agents lack object permanence so keeping them on a coherent plan requires giving them such a thoroughly documented plan up front. It actually has me wondering if optimal coding agent usage at scale resembles something of a return to waterfall (probably in more of a Royce sense than the bogeyman agile evangelists derived from the original idea) where the humans on the team mostly spend their time banging out systems specifications and testing protocols, and iteration on the spec becomes somewhat more removed from implementing it than it is in typical practice nowadays.
People haven't noticed because the software industry was already mostly unoriginal slop, even prior to LLMs, and people are good at ignoring unoriginal slop.
https://www.reddit.com/r/pcgaming/comments/1pl7kg1/over_1900...
I created a platform for a virtual pub quiz for my team at my day job, built multiple pandingpages for events, debugged dark table to recognize my new camera (it was to new to be included in the camera.xml file, but the specs were known). I debugged quite a few parts of a legacy shitshow of an application, did a lot of infrastructure optimization and I also created a massive ton of content as a centaur in dialog with the help of Claude Code.
But I don't do "Show HN" posts. And I don't advertise my builds - because other than those named, most are one off things, that I throw away after this one problem was solved.
To me code became way more ephemeral.
But YMMV - and that is a good thing. I also believe that way less people than the hype bubble implies are actually really into hard core usage like Pete Steinberger or Armin Ronacher and the likes.
I use AI/agents in quite similar ways, and even rekindled multiple personal projects that had stalled. However, to borrow OPs parlance, these are not "houses" - more like sheds and tree-houses. They are fun and useful, but not moving the needle on housing stock supply, so to speak.
I don’t see AI helping with knowing what to build at all and I also don’t see AI finding novel approaches to anything.
Sure, I do think there is some unrealized potential somewhere in terms of relatively low value things nobody built before because it just wasn’t worth the time investment – but those things are necessarily relatively low value (or else it would have been worth it to build it) and as such also relatively limited.
Software has amazing economies of scale. So I don’t think the builder/tool analogy works at all. The economics don’t map. Since you only have to build software once and then it doesn’t matter how often you use it (yeah, a simplification) even pretty low value things have always been worth building. In other words: there is tons of software out there. That’s not the issue. The issue is: what it the right software and can it solve my problems?
At the same time I see people claiming 100x increases and how they produce 15k lines of code each day thanks to AI, but all I can wonder is how these people managed to find 100x work that needed to be done.
The problem with this that after doing this hard work someone can just copy easily your hard work and UI/UX taste. I think distribution will be very important in the future.
We might end up that in future that you have already in social media where influencers copy someones post/video and not giving credits to original author.
Or indeed, somebody might steal and launder your work by scooping them up into a training set for their model and letting it spit out sloppy versions of your thing.
So now need to think of different kind of ideas, something on line of games that may take multiple iteration to get perfected.
I mean this is how it's always been throughout history.
Creating something new is hard, copying something in terms of energy spent, is far easier. This is software or physical objects that don't require massive amounts of expensive technology to reproduce.
Small and mid sized companies are getting custom software now.
Small software is able to be packed with extra features instead of bare minimum.
To be honest, I think the surrounding paragraph lumps together all anti-AI sentiments.
For example, there is a big difference between "all AI output is slop" (which is objectively false) and "AI enables sloppy people to do sloppy work" (which is objectively true), and there's a whole spectrum.
What bugs me personally is not at all my own usage of these tools, but the increase in workload caused by other people using these tools to drown me in nonsensical garbage. In recent months, the extra workload has far exceeded my own productivity gains.
For the non-technical, imagine a hypochondriac using chatgpt to generate hundreds of pages of "health analysis" that they then hand to their doctor and expect a thorough read and opinion of, vs. the doctor using chatgpt for sparring on a particular issue.
https://en.wikipedia.org/wiki/Brandolini%27s_law
>The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it.
[0] https://www.marble.onl/posts/this_cost_170.html
[1] https://www.anthropic.com/engineering/building-c-compiler
Headline features aren't much faster. You still need to gather requirements, design a good architecture, talk with stakeholders, test your implementation, gather feedback, etc. Speeding up the actual coding can only move the needle so much.
I've got a couple Claude Code skills set up where I just copy/paste a Slack link into it and it links people relevant docs, gives them relevant troubleshooting from our logs, and a hook on the slack tools appends a Claude signature to make sure they know they weren't worth my time.
That said, there's this weird quicksand people around me get in where they just spend weeks and weeks on their AI tools and don't actually do much of anything? Like bro you burned your 5 hour CC Enterprise limit all week and committed...nothing?
rather than new stuff for everyone to use, the future could easily be everyone building their own bespoke tools for their own problems.
> You have to turn off the sandbox, which means you have to provide your own sandbox. I have tried just about everything and I highly recommend: use a fresh VM.
> I am extremely out of touch with anti-LLM arguments
'Just pay out the arse and run models without a sandbox or in some annoying VM just to see them fail. Wait, some people are against this?'
So why not just wait out this insane initial phase, and if anything is left standing afterwards and proves itself, just learn that.
They provide value today so I'm using them today.
I beg to differ. There are a whole lot of folks with astonishingly incomplete understanding about all the facts here who are going to continue to make things very, very complicated. Disagreement is meaningless when the relevant parties are not working from the same assumption of basic knowledge.
There’s a lot of unwillingness to even attempt to try the tools.
There are people I work with who are deep in the AI ecosystem and it's obvious what tools they're using It would not be uncharitable in any way to characterize their work as pure slop that doesn't work, buggy, untested adequately, etc.
The moment I start to feel behind I'll gladly start adopting agentic AI tools, but as things stand now, I'm not seeing any pressing need.
Comments like these make me feel like I'm being gaslit.
If this stuff was self-evidently as useful as it's being made out to be, there would be no point in constantly trying to pressure, coax and cajole people into it. You don't need to spook people into using things that are useful, they'll do it when it makes sense.
The actual use-case of LLMs is dwarfed by the massive investment bubble it has become, and it's all riding on future gains that are so hugely inflated they will leave a crater that makes the dotcom bubble look like a pothole.
One dude with an LLM should be able to write a browser fully capable of browsing the modern web or an OS from scratch in a year, right?
Chrome took at least a thousand man years i.e. 100 people working for 10 years.
I'm lowballing here: it's likely way, way more.
If ai gives 10x speedup, to reproduce Chrome as it is today would require 1 person working for 100 years, 10 people working for 10 years or 100 people working for 1 year.
Clearly, unrealistic bar to meet.
If you want a concrete example: https://github.com/antirez/flux2.c
Creator of Redis started this project 3 weeks ago and use Claude Code to vibe code this.
It works, it's fast and the code quality is as high as I've ever seen a C code base. Easily 1% percentile of quality.
Look at this one-shotted working implementation of jpeg decoder: https://github.com/antirez/flux2.c/commit/a14b0ff5c3b74c7660...
Now, it takes a skilled person to guide Claude Code to generate this but I have zero doubts that this was done at least 5x-10x faster than Antirez writing the same code by hand.
"Being left in the dust" would also mean it's impossible for new people / graduates to ever catch up. I don't think it is. Even though I learned react a few years after it was in vogue (my company bet on the wrong horse), I quickly got up to speed and am just as productive now as someone that started a bit earlier.
There was a land rush to create apps. Basic stuff like the flash light, todo lists, etc, were created and found a huge audience. Development studios were established, people became very successful out of it.
I think the same thing will happen here. There is a first mover advantage. The future is not yet evenly distributed.
You can still start as an iOS developer today, but the opportunity is different.
The introduction of the App Store did not increase developer productivity per se. If anything, it decreased developer productivity, because unless you were already already a Mac developer, you had to learn a programming language you've never used, Objective-C, (now it's largely Swift, but that's still mainly used only on Apple platforms) and a brand new Apple-specific API, so a lot of your previous programming expertise became obsolete on a new platform. What the App Store did that was valuable to developers was open up a new market and bring a bunch of new potential customers, iPhone users, indeed relatively wealthy customers willing to spend money on software.
What new market is brought by LLMs? They can produce as much source code as you like, but how exactly do you monetize that massive amount of source code? If anything, the value of source code and software products will drop as more is able to be produced rapidly.
The only new market I see is actually the developer tool market for LLM fans, essentially a circular market of LLM developers marketing to other LLM developers.
As far as the developer job market is concerned, it's painfully clear that companies are in a mass layoff mood. Whether that's due to LLMs, or whether LLMs are just the cover story, the result is the same. Developer compensation is not on the rise, unless you happen to be recruited by one of the LLM vendors themselves.
My impression is that from the developer perspective, LLMs are a scheme to transfer massive amounts of wealth from developers to the LLM vendors. And you can bet the prices for access to LLMs will go up, up, up over time as developers become hooked and demand increases. To me, the whole "OpenClaw" hype looks like a crowd of gamblers at a casino, putting coins in slot machines. One thing is for certain: the house always wins.
I think it will make prototyping and MVP more accessible to a wider range of people than before. This goes all the way from people who don't know how to code up to people who know very well how to code, but don't have the free time/energy to pursue every idea.
Project activation energy decreases. I think this is a net positive, as it allows more and different things to be started. I'm sure some think it's a net negative for the same reasons. If you're a developer selling the same knowledge and capacity you sold ten years ago things will change. But that was always the case.
My comparison to iOS was about the market opportunity, and the opportunity for entrepreneurship. It's not magic, not yet anyway. This is the time to go start a company, or build every weird idea that you were never going to get around to.
There are so many opportunities to create software and companies, we're not running out of those just because it's faster to generate some of the code.
Returning to the iOS analogy, though, there was only a short period of time in history when a random developer with a flashlight or fart app could become successful in the App Store. Nowadays, such a new app would flop, if Apple even allowed it, as you admitted: "You can still start as an iOS developer today, but the opportunity is different." The software market in general is not new. There are already a huge number of competitors. Thus, when you say, "This is the time to go start a company, or build every weird idea that you were never going to get around to," it's unclear why this would be the case. Perhaps the barrier to entry for competitors has been lowered, yet the competition is as fierce as ever (unlike in the early App Store).
In any case, there's a huge difference between "the barrier to entry has been lowered" and "those who don't use LLMs will be left in the dust". I think the latter is ridiculous.
Where are the original flashlight and fart app developers now? Hopefully they made enough money to last a lifetime, otherwise they're back in the same boat as everyone else.
Yeah, it’s a bit incendiary, I just wanted to turn it into a more useful conversation.
I also think it overstates the case, but I do think it’s an opportunity.
It’s not just that the barrier to entry has been lowered (which it has) but that someone with a lot of existing skill can leverage that. Not everyone can bring that to the table, and not everyone who can is doing so. That’s the current advantage (in my opinion, of course).
All that said, I thought the Vision Pro was going to usher in a new era of computing, so I’m not much of a prognosticator.
> I also think it overstates the case
I think it's a mistake to defend and/or "reinterpret" the hype, which is not helping to promote the technology to people who aren't bandwagoners. If anything, it drives them away. It's a red flag.
I wish you would just say to the previous commenter, hey, you appear to be exaggerating, and that's not a good idea.
The App Store reshuffled the deck. Some people recognized that and took advantage of the decalcification. Some of them did well.
You've recognized some implications of the reshuffle that's currently underway. Maybe you're right that there's a bias toward the LLM vendors. But among all of it, is there a niche you can exploit?
What do you get from it? Say you produce more, do you get a higher salary?
What I have seen so far is the opposite: if you don't produce more, you risk getting fired.
I am not denying that LLMs make me more productive. Just saying that they don't make me more wealthy. On the other hand, they use a ton of energy at a time where we as a society should probably know better. The way I see it, we are killing the Earth because we produce too much. LLMs help us produce more, why should we be happy?
Not all that hard to learn, but waiting for things to settle down assumes things are going to settle down. Are they? When?
These are leaked implementation details that the labs are forcing us to know because these are weak, early products and they’re still exploring the design space. The median user doesn’t want to and shouldn’t have to care about details like this.
Future products in this space won’t have them and future users won’t be left in the dust by not learning them today.
Python programmers aren’t left behind by not knowing malloc and free.
"Anti-LLM sentiment" within software development is nearly non-existent. The biggest kind of push-back to LLMs that we see on HN and elsewhere, is effectively just pragmatic skepticism around the effectiveness/utility/ROI of LLMs when employed for specific use-cases. Which isn't "anti-LLM sentiment" any more than skepticism around the ability of junior programmers to complete complex projects is "anti-junior-programmer sentiment."
The difference between the perspectives you find in the creative professions vs in software dev, don't come down to "not getting" or "not understanding"; they really are a question of relative exposure to these pro-LLM vs anti-LLM ideas. Software dev and the creative professions are acting as entirely separate filter-bubbles of conversation here. You can end up entirely on the outside of one or the other of them by accident, and so end up entirely without exposure to one or the other set of ideas/beliefs/memes.
(If you're curious, my own SO actually has this filter-bubble effect from the opposite end, so I can describe what that looks like. She only hears the negative sentiment coming from the creatives she follows, while also having to dodge endless AI slop flooding all the marketplaces and recommendation feeds she previously used to discover new media to consume. And her job is one you do with your hands and specialized domain knowledge; so none of her coworkers use AI for literally anything. [Industry magazines in her field say "AI is revolutionizing her industry" — but they mean ML, not generative AI.] She has no questions that ChatGPT could answer for her. She doesn't have any friends who are productively co-working with AI. She is 100% out-of-touch with pro-LLM sentiment.)
For software the situation is different. Being opposed to LLM-generated software is just batshit crazy at this point. The value that LLMs provide to the process makes learning to use them, objectively, an absolute must; otherwise you are simply wasting time and money. Eric S. Raymond put it something like "If you call yourself a software engineer, you have no excuse not to be using these tools. Get your thumb out of your ass and learn."
I can say “learn how to use vim makeprg feature so that you can jump directly to errors reported by the build and tool” and it’s very clear where the ROI. But all the AI hypers are selling are hope, prayers, and rituals.
I'm curious about what industry you are in and the tech stack you are using?
It may take some human intervention, but the productivity results are pretty consistent: tasks that used to take weeks now take hours or days. This puts in reach the ability to try things you wouldn't countenance otherwise due to the effort and tedium involved. You'd have to be a damn fool not to take advantage of the added velocity. This is why what we do is called "engineering", not a handicraft.
> This puts in reach the ability to try things you wouldn't countenance otherwise due to the effort and tedium involved.
If you’re talking about prototypes, a whiteboard is way cheaper and less time consuming than an agent.
Strong disagree right there. I remember talking to a (developer) coworker a few months ago who seemed like the biggest AI proponent on our team. When we were one-on-one during a lunch though, he revealed that he really doesn't like AI that much at all, he's just afraid to speak up against it. I'm in a few Discord channels with a lot of highly skilled (senior and principal programmers) who mostly work in game development (or adjacent), and most of them either mock LLMs or have a lot of derision for it. Hacker News is kind of a weird pro-AI bubble, most other places are not nearly as keen on this stuff.
This is certainly untrue. I want to say "obviously", which means that maybe I am misunderstanding you. Below are some examples of negative sentiments programmers have - can you explain why you are not counting these?
NOTE: I am not presenting these as an "LLMs are bad" argument. My own feelings go both ways. There is a lot that's great about LLMs, and I don't necessarily agree with every word I've written below - some of it is just my paraphrasing of what other people say. I'm only listing examples of what drives existing anti-LLM sentiment in programmers.
1. Job loss, loss of income, or threat thereof
These two are exacerbated by the pace of change, since so many people already spent their lives and money establishing themselves in the career and can't realistically pivot without becoming miserable - this is the same story for every large, fast change - though arguably this one is very large and very fast even by those standards. Lots of tech leadership is focusing even more than they already were on cheap contractors, and/or pushing employees for unrealistic productivity increases. I.e. it's exacerbating the "fast > good" problem, and a lot of leadership is also overestimating how far it reduces the barrier to creating things, as opposed to mostly just speeding up a person's existing capabilities. Some leadership is also using the apparent loss of job security as leverage beyond salary suppression (even less proportion of remote work allowed, more surveillance, worse office conditions, etc).
2. Happiness loss (in regards to the job itself, not all the other stuff in this list)
This is regarding people who enjoy writing/designing programs but don't enjoy directing LLMs; or who don't enjoy debugging the types of mistakes LLMs tend to make, as opposed to the types of mistakes that human devs tend to make. For these people, it's like their job was forcibly changed to a different, almost unrelated job, which can be miserable depending on why you were good at - or why you enjoyed - the old job.
3. Uncertainty/skepticism
I'm pushing back on your dismissal of this one as "not anti-LLM sentiment" - the comparison doesn't make sense. If I was forced to only review junior dev code instead of ever writing my own code or reviewing experienced dev code, I would be unhappy. And I love teaching juniors! And even if we ignore the subset of cases where it doesn't do a good job or assume it will soon be senior-level for every use case, this still overlaps with the above problem: The mistakes it makes are not like the mistakes a human makes. For some people, it's more unnatural/stressful to keep your eyes peeled for the kinds of mistakes it makes. For these people, it's a shift away from objective, detail-oriented, controlled, concrete thinking; away from the feeling of making something with your hands; and toward a more wishy-washy creation experience that can create a feeling of lack of control.
4. Expertise loss
A lot of positive outcomes with LLMs come from being already experienced. Some argue this will be eroded - both for new devs and existing experienced devs.
5. The training data ownership/morality angle
> A lot of positive outcomes with LLMs come from being already experienced. Some argue this will be eroded - both for new devs and existing experienced devs.
This is true, but the pace of progress is so mind blowing the experts we have now might just be enough until the whole industry becomes obsolete (10-20 years assuming the lower bound of the trend line holds?)
I like this. What's more, while AI-generated art has a characteristic sameyness to it, the human-produced art stands out in its originality. It has character and soul. Even if it's bad! AI slop has made the human-created stuff seem even more striking by comparison. The market for human art isn't going anywhere, just like the audience for human-played chess went nowhere after Deep Blue. I think people will pay a premium for it, just to distinguish themselves from the slop. The same is true of writing and especially music. I know of no one who likes listening to AI-generated music. Even Sabrina Carpenter would raise less objection.
The same, I'm afraid, cannot be said for software—because there is little value for human expression in the code itself. Code is—almost entirely—strictly utilitarian. So we are now at an inflection point where LLMs can generate and validate code that's nearly as good, if not better, than what we can produce on our own. And to not make use of them is about as silly as Mel Kaye still punching in instruction opcodes in hex into the RPC-4000, while his colleagues make use of these fancy new things called "compilers". They're off building unimaginably more complex software than they could before, but hey, he gets his pick of locations on the rotating memory drum!
I'm one of the nonexistent anti-LLMers when it comes to software. I hate talking to a clanker, whose training data set I don't even have access to let alone the ability to understand how my input affects its output, just to do what I do normally with the neural net I've carried around in my skull and trained extensively for this very purpose. I like working directly with code. Code is not just a product for me; it is a medium of thought and expression. It is a formalized notation of a process that I can use to understand and shape that process.
But with the right agentic loops, LLMs can just do more, faster. There's really no point in resisting. The marginal value of what I do has just dropped to zero.
I see it all the time in professional and personal circles. For one, you are shifting the goalpost on what is “anti-llm”, two, people are talking about the negative social, political and environmental impacts.
What is your source here?
The anti-LLM arguments aren't just "hand tools are more pure." I would even say that isn't even a majority argument. There are plenty more arguments to make about environmental and economic sustainability, correctness, safety, intellectual property rights, and whether there are actual productivity gains distinguishable from placebo.
It's one of the reasons why "I am enjoying programming again" is such a frustrating genre of blog post right now. Like, I'm soooo glad we could fire up some old coal plants so you could have a little treat, Brian from Middle Management.
But secondly, there's an entire field of LLM-assisted coding that's being almost entirely neglected and that's code autocomplete models. Fundamentally they're the same technology as agents and should be doing the same thing: indexing your code in the background, filtering the context, etc, but there's much less attention and it does feel like the models are stagnating.
I find that very unfortunate. Compare the two workflows:
With a normal coding agent, you write your prompt, then you have to at least a full minute for the result (generally more, depending on the task), breaking your flow and forcing you to task-switch. Then it gives you a giant mass of code and of course 99% of the time you just approve and test it because it's a slog to read through what it did. If it doesn't work as intended, you get angry at the model, retry your prompt, spending a larger amount of tokens the longer your chat history.
But with LLM-powered auto-complete, when you want, say, a function to do X, you write your comment describing it first, just like you should if you were writing it yourself. You instantly see a small section of code and if it's not what you want, you can alter your comment. Even if it's not 100% correct, multi-line autocomplete is great because you approve it line by line and can stop when it gets to the incorrect parts, and you're not forced to task switch and you don't lose your concentration, that great sense of "flow".
Fundamentally it's not that different from agentic coding - except instead of prompting in a chatbox, you write comments in the files directly. But I much prefer the quick feedback loop, the ability to ignore outputs you don't want, and the fact that I don't feel like I'm losing track of what my code is doing.
But if you try some penny-saving cheap model like Sonnet [..bad things..]. [Better] pay through the nose for Opus.
Opus in Cursor is just so much more responsive and easy to talk to, compared to Opus in Claude.
Cursor has this "Auto" mode which feels like it has very liberal limits (amortized cost I guess) that I'm also trying to use more, but -- I don't really like to flip a coin and if it lands up head then waste half hour discovering the LLM made a mess the LLM and try again forcing the model.
Perhaps in March I'll bite the bullet and take this authors advice.
You can enjoy it while it lasts, OpenAI is being very liberal with their limits because of CC eating their lunch rn.
I was spending unholy amounts of money and tokens (subsidized cloud credits tho) forcing Opus for everything but I’m very happy with this new setup. I’ve also experimented with OpenCode and their Zen subscription to test Kimi K2.5 an similar models and they also seem like a very good alternative for some tasks.
What I cannot stand tho is using sonnet directly (it’s fine as a subagent), I’ve found it to be hard to control and doesn’t follow detailed instructions.
I’m an avid cursor user (with opus), and have been trying alternatives recently. Codex has been an immense letdown. I think I was too spoiled by cursor’s UX and internal planning prompt.
It’s incredibly slow, produces terribly verbose and over-complicated code (unless I use high or xhigh, which are even slower), and missed a lot of details. Python/django and react frontend.
For the first time I felt like I could relate to those people who say it doesn’t make them faster,” because they have to keep fixing the agent’s shot, never felt that with opus 4.5 and 4.6 and cursor
I mean does it matter what code it's producing? If it renders and functions just use it. I think it's better to take the L on verbose code and optimizing the really ugly bits by hand in a few minutes than be kneecapped every 5 hour by limits and constant pleas to shift to Sonnet.
Perhaps AI time is the inverse of Valve time.
The first sentence out of my mouth was a system prompt
This vscode extension makes it almost as easy to point codex to something as when doing it in cursor:
https://github.com/suzukenz/vscode-copy-selection-with-line-...
I think we are going to start hearing stories of people going into thousands in CC debt because they were essentially gambling with token usage thinking they would hit some startup jackpot.
It's 90 percent the same thing as Claude but with flat-rate costs.
Startup is a gamble with or without the LLM costs.
I have been coding for 20 years, I have a good feel for how much time I would have spent without LLM assistance. And if LLMs vanish from the face of the earth tomorrow, I still saved myself that time.
"Using anything other than the frontier models is actively harmful" - so how come I'm getting solid results from Copilot and Haiku/Flash? Observe, Orient, Decide, Act, Review, Modify, Repeat. Loops with fancy heuristics, optimized prompts, and decent tools, have good results with most models released in the past year.
We're at the point where copilot is irrelevant. Your way of working is irrelevant. Because that's not how you interact with coding AIs anymore, you're chatting with them about the code outside the IDE.
Yes.
> It's hard to communicate the difference the last 6 months has seen.
No, it isn't. The hypebeast discovered Claude code, but hasn't yet realized that the "let the model burn tokens with access to a shell" part is the key innovation, not the model itself.
I can (and do) use GH Copilot's "agent" mode with older generation models, and it's fine. There's no step function of improvement from one model to another, though there are always specific situations where one outperforms. My current go-to model for "sit and spin" mode is actually Grok, and I will splurge for tokens when that doesn't work. Tools and skills and blahblahblah are nice to have (and in fact, part of GH Copilot now), but not at all core to the process.
Just this month I've burned through 80% of my Copilot quota of Claude Opus 4.6 in a couple of days to get it to help me with a silly hobby project: https://github.com/ncruces/dbldbl
It did help. The project had been sitting for 3 years without trig and hyperbolic trig, and in a couple days of spare time I'm adding it. Some of it through rubber ducking chat and/or algorithmic papers review (give me formulas, I'll do it), some through agent mode (give me code).
But if you review the PR written in agent mode, the model still lies to my face, in trivial but hard to verify ways. Like adding tests that say cosh(1) is this number at that OEIS link, and both the number and the OEIS link are wrong, but obviously tests pass because it's a lie.
I'm not trying to bash the tech. I use it at work in limited but helpful ways, and use hobby stuff like this as a testbed precisely to try to figure out what they're good at in a low stakes setting.
But you trust the plausibly looking output of these things at your own peril.
If you check the docs, smaller, faster, older models are recommended for 'lightweight' coding. There's several reasons for this. 1) a smaller model doesn't have as good deep reasoning, so it works okay for a simple ask. 2) small context, small task, small model can produce better results than big context, big task, big model. The lost-in-the-middle problem is still unsolved, leading to mistakes that get worse with big context, and longer runs exacerbate issues. So small context/task that ends and starts a new loop (with planning & learning) ends up working really well and quickly.
There's a difference between tasks and problem-solving, though. For difficult problems, you want a frontier reasoning model.
I took a look at the result and its maybe half of stuff missing completely, rest is cryptic. I know that codebase by heart since I created it. From my 20+ years of experience correcting all this would take way more effort than manual rewrite from scratch by a senior. Suffice to say thats not what upper management wants to hear, llm adoption often became one of their yearly targets to be evaluated against. So we have a hammer and looking for nails to bend and crook.
Suffice to say this effort led nowhere since we have other high priority goals, for now. Smaller things here & there, why not. Bigger efforts, so far sawed-off 2-barrel shotgun loaded with buckshot right into both feet.
I used claude code to port rust pdb parsing library to typescript.
My SumatraPDF is a large C++ app and I wanted visibility into where does the size of functions / data go, layout of classes. So I wanted to build a tool to dump info out of a PDB. But I have been diagnosed with extreme case of Rustophobiatis so I just can't touch rust code. Hence, the port to typescript.
With my assistance it did the work in an afternoon and did it well. The code worked. I ran it against large PDB from SumatraPDF and it matched the output of other tools.
In a way porting from one language to another is extreme case of refactoring and Claude did it very well.
I think that in general (your experience notwithstanding) Claude Caude is excellent at refactorings.
Here are 3 refactorings from SumatraPDF where I asked claude code to simplify code written by a human:
https://github.com/sumatrapdfreader/sumatrapdf/commit/a472d3... https://github.com/sumatrapdfreader/sumatrapdf/commit/5624aa... https://github.com/sumatrapdfreader/sumatrapdf/commit/a40bc9...
I hope you agree the code written by Claude is better than the code written by a human.
Granted, those are small changes but I think it generalizes into bigger changes. I have few refactorings in mind I wanted to do for a long time and maybe with Claude they will finally be feasible (they were not feasible before only because I don't have infinite amount of time to do everything I want to do).
Translating a vibe is something the Ur-LLMS (GPT3 etc) were very good at so it’s not entirely surprising that the current state of the art is to be found in things of a “translate thing X that already exists into context Y” nature.
If that is true, why should one invest in learning now rather than waiting for 8 months to learn whatever is the frontier model then?
But if you do want to use LLMs for coding now, not using the best models just doesn't make sense.
I think you (and others) might be misunderstanding his statement a bit. He's not saying that using an old model is harmful in the sense that it outputs bad code -- he's saying it's harmful because some of the lessons you learn will be out of date and not apply to the latest models.
So yes, if you use current frontier models, you'll need to recalibrate and unlearn a few things when the next generation comes out. But in the meantime, you will have gotten 8 months (or however long it takes) of value out of the current generation.
Using agents that interact with APIs represents people being able to own their user experience more. Why not craft a frontend that behaves exactly the the way YOU want it to, tailor made for YOUR work, abstracting the set of products you are using and focusing only on the actual relevant bits of the work you are doing? Maybe a downside might be that there is more explicit metering of use in these products instead of the per-user licensing that is common today. But the upside is there is so much less scope for engagement-hacking, dark patterns, useless upselling, and so on.
OK, but: that's an economic situation.
> so much less scope for engagement-hacking, dark patterns, useless upselling, and so on.
Right, so there's less profit in it.
To me it seems this will make the market more adversarial, not less. Increasing amounts of effort will be expended to prevent LLMs interacting with your software or web pages. Or in some cases exploit the user's agentic LLM to make a bad decision on their behalf.
it's basically SEO all over again but worse, because the attack surface is the user's own decision-making proxy. at least with google you could see the search results and decide yourself. when your agent just picks a vendor for you based on what it "found," the incentive to manipulate that process is enormous.
we're going to need something like a trust layer between agents and the services they interact with. otherwise it's just an arms race between agent-facing dark patterns and whatever defenses the model providers build in.
I mean, services _could_ make it harder to use LLMs to interact with them, but if agents are popular enough they might see customers start to revolt over it.
We're already seeing this with search. Ask an LLM "what tools do X" and the answer depends heavily on structured data, citation patterns, and how well your docs/content map to the LLM's training. Companies with great API docs but zero presence in the training data just won't exist to these agents.
So it's not just "API docs = product" -- it's more like "machine-legible presence = existence." Which is a weird new SEO-like discipline that barely has a name yet.
I think this is a neglected area that will see a lot of development in the near future. I think that even if development on AI models stopped today - if no new model was ever trained again - there are still decades of innovation ahead of us in harnessing the models we already have.
Consider ChatGPT: the first release relied entirely on its training data to answer questions. Today, it typically does a few Google searches and summarizes the results. The model has improved, but so has the way we use it.
How I program with agents - https://news.ycombinator.com/item?id=44221655 - June 2025 (295 comments)
I see this a lot here
Copyright law, education, just the sheer scale of things changing because of LLMs are some things off the top of my head why "power tools vs carpentry" is a bad analogy.
Sure, replace me with AI, but I better get royalties on my public contributions. I like many other developers have kids and other responsibilities to pay for.
We did not share our work publicly to be replaced. The same way I did not lend my neighbour my car so he could run me over, that was implicit.
Writing code has never been the limiting factor, it's everything else that goes into it.
Like, I don't mind that there's a bunch of weekend warriors out here building shoddy gazebos and sheds with their brand new overpriced tools, incorrecting each other on the best way to do things. We had that with the bitcoin and NFT bros already.
What I do roll my eyes at is when the bros start talking about how they're totally going to build bridges and planes and it's gonna be soooo easy to get to new places, just slap down a bridge.
Uh huh. Y'all do not understand what building those actually entails lol.
The 'fear' is about losing ones livelihood and getting locked out of homeownership and financial security. its not complicated. life is actually largely determined by your access to capital, despite whatever fresh coping strategy the afflicted (and the afflicting) like to peddle.
the quality of life versus capital availability is very non-linear. there is a step-change around the $500k mark where you reach 'orbital velocity', where as long as you dont suffer severe misfortune or make mistakes, you will start accelerating upwards (albeit very slowly.)
under that line, you are constantly having to fight 'gravity'.
basically everyone in tech is openly or quietly aiming to get there, and LLMs have made that trek ever more precarious than before.
Not a plug but really that’s exactly why we’re building sandboxes for agents with local laptop quality. Starting with remote xcode+sim sandboxes for iOS, high mem sandbox with Android Emulator on GPU accel for Android.
No machine allocation but composable sandboxes that make up a developer persona’s laptop.
If interested, a quick demo here https://www.loom.com/share/c0c618ed756d46d39f0e20c7feec996d
muvaf[at]limrun[dot]com
> That was a net benefit to the world, that we all don't have to work to eat.
I’m pretty sure most all of us are still working to have food to eat and shelter for ourselves and our families.
Also, while the on-going industrial and technological revolution has certainly brought benefits, it’s an open question as to whether it will turn out to be a net benefit. There’s a large-scale tragedy of the commons experiment playing out and it’s hard to say what the result will be.
It might be just me but this reads as very tone deaf. From my perspective, CEOs are seething at the mouth to make as many developers redundant as possible, not being shy about this desire. (I don't see this at all as inevitable, but tech leaders have made their position clear)
Like, imagine the smugness of some 18th century "CEO" telling an artisan, despite the fact that he'l be resigned to working in horrific conditions at a factory, to not worry and think of all the mass produced consumer goods he may enjoy one day.
It's not at all a stretch of the imagination that current tech workers may be in a very precarious situation. All the slopware in the world wouldn't console them.
While the idea of programmers working two hours a day and spending the rest of it with their family seems sunny, that's absolutely not how business is going to treat it.
Thought experiment... CEO has a team of 8 engineers. They do some experiments with AI, and they discover that their engineers are 2x more effective on average . What does the CEO do?
a) Change the workweek to 4 hours a day so that all the engineers have better work/life balance since the same amount of work is being done.
b) Fire half the engineers, make the 4 remaining guys pick up the slack, rinse and repeat until there's one guy left?
Like, come on. There's pushback on this stuff not because the technology is bad, (although it's overhyped), but because the no sane person trusts our current economic system to provide anything resembling humane treatment of workers. The super rich are perfectly fine seeing half the population become unemployed, as far as I can tell, as long as their stock numbers go up.
Though at the same time I also think a lot of the CEO-types (at least in the pure software world) who believe they are going to capture the value of this productivity shift are also in for a rude awakening because if AI doesn't stall out, its only a matter of time from when their engineers are replaceable to when their company doesn't need to exist at all anymore.
"AI won't replace you. The guy who's about to get fired but has more to lose is going to replace you."
But if AI keeps getting better at code, it will produce entire in-silico simulation workflows to test new drugs or even to design synthetic life (which, again, could make us all die, or worse). Yet there is a tiny, tiny chance we will use it to fix some of the darkest aspects of human existence. I will take that.
We have a lot of actual problems to deal with that aren't telling ghost stories about sand. Focus on those.
Much of my learning still requires experimentation - including lots of token volume so hitting limits is a problem.
And secondly I’m looking for workflows that build the thing without needing to be at the absolute edge of the LLM capability. Thats where fragility and unpredictability live. Where a new model with slightly different personality is released and it breaks everything. I’d rather have flow that is simple and idiot proof that doesn’t fall apart at the first sign of non-bleeding edge tokens. That means skipping the gains from something opus could one shot ofc but that’s acceptable to me
I don't think it is the best way to look at it. I think that now every team has the power to build and maintain an internal agent (tool + UX) to manager software products. I don't necessarily think that chat-only is enough except for small projects, so teams will build agent that gives them access to the level of abstraction that works best.
It's a data point but this weekend (e.g. in 2 days) I build a desktop + web agent that is able to help me reason on system design and code. Built with Codex powered by the Codex SDK. It is high quality. I've been a software engineer and director of engineering for 10 years. I'm blown away.
I have yet to do this and see any other year. Was there someone who bought a ton of accounts in 2011 to farm them out? A data breach? Was 2011 just a very big year for new users? (My own account is from 2011)
Calling it bot is a bit dismissive though. It's an agent!
If so, send a DM on twitter to @edfixyz with your phone number and I will call you immediately. Or give me your twitter handle.
I'm tired of that BS - when people don't like what you write they call you a bot.
I like Claude Code too btw.
The crazy thing here is that I wrote the initial comment myself!
Assuming you’re not a bot. It’s nothing to do with you having a good experience, it’s the way you wrote about that experience that sounds like a product placement.
I asked OpenAIs very own ChatGPT 5.2 powered by OpenAI to tell you why it sounds like a product placement:
“ Because it hits a bunch of “native ad / testimonial” tells at once: • Brand-name density in a tiny space. “Built with Codex powered by the Codex SDK” repeats the same brand in two adjacent phrases, like copy that’s trying to lodge a name in your head rather than naturally describe a build. • Overly polished value signals. “High quality” is a generic superlative with no concrete evidence (features, metrics, constraints, tradeoffs). Ads often lean on verdict words instead of specifics. • Credential + astonishment combo. “I’ve been a software engineer and director of engineering for 10 years” is classic authority framing, immediately followed by “I’m blown away.” That’s a common testimonial structure: I’m hard to impress → I’m impressed. • Time-compressed “miracle build” narrative. “This weekend (in 2 days) I build a desktop + web agent…” reads like the “you can do it fast/easily now” story arc you see in promos. Not impossible—just a familiar marketing shape. • “It’s a data point” language. That phrase feels like social-proof seeding: “don’t treat this as hype, just one datapoint,” which paradoxically makes it feel more like deliberate persuasion. • No friction or downsides. Real engineer excitement usually includes at least one caveat (bugs, rough edges, limitations, cost, setup pain). The total absence makes it sound curated. • Benefit phrased like positioning. “Able to help me reason on system design and code” is basically a product pitch line (target user + problem + outcome) rather than a personal anecdote (“it helped me untangle X design and refactor Y”).”
It's always the CTO types who get most enthusiastic.
Not sure which camp I'm in, but I enjoyed the imagery.
Wow I know that feel.
I'm here using LLM for daily work and even hobbies in very conservative manners and didn't think much of it.
Now when I have casual discussions with other folks, especially non-tech people, the visceral hatred I get for even mentioning AI and the fact that I use it is insane. There's like an entire sub group of people who are so out of touch with these tools they think they're the devil like the anti-GMO crazies and the PETA psychos.
I agree with this and I think it's funny to see people publish best practices for working with AI that are like, "Write a clear spec. Have a style guide. Use automated tests."
I'm not convinced it's 100% true because I think there are code patterns that AI handles better than humans and vice versa. But I think it's true enough to use as a guiding philosophy.
My conclusion as well. It feels paradoxical, maybe because on some level I still think of an LLM as some weird gadget, not a coworker. Context ephemerality is more or less the only veritable difference from a human programmer, I'd say. And, even then, context introduction with LLMs is a speedrun of how you'd do it with new human members of a project. Awesome times we live in.
As the author says, there's nothing wrong with the idea of the IDE. Of course you want to be using the best, most powerful tools!
AI showed us that our current-gen text-editor-first IDEs are massively underserving the needs of the public, yes, but it didn't really solve that problem. We still need better IDEs! What has changed is that we now understand how badly we need them. (source: I am an IDE author)
Just a question? What IDE feature is obsolete now? Ability to navigate the code? Integration with database, Docker, JIRA, Github (like having PR comments available, listed, etc), Git? Working with remote files? Building the project?
Yes, I can ask copilot to build my project and verify tests results, but it will eat a lot of tokens and added value is almost none.
The added value is that it can iterate autonomously and finish tasks that it can't one-shot in its first code edit. Which is basically all tasks that I assign to Copilot.
The added value is that I get to review fully-baked PRs that meet some bar of quality. Just like I don't review human PRs if they don't pass CI.
Fully agree on IDEs, though. I absolutely still need an IDE to iterate on PRs, review them, and tweak them manually. I find VSCode+Copilot to be very good for this workflow. I'm not into vibe coding.
Visual C++ 6 was incredible! My favourite IDE of all time too.
We have built two of them now, and clearly the state of the art here can be improved. But it is hard to push too much on this while the models keep improving.
the hard part isn't the loop itself — it's everything around failure recovery.
when a browser agent misclicks, loads a page that renders differently than expected, or hits a CAPTCHA mid-flow, the 9-line loop just retries blindly. the real harness innovation is going to be in structured state checkpointing so the agent can backtrack to the last known-good state instead of restarting the whole task. that's where the gap between "works in a demo" and "works on the 50th run" lives.
┌────────────────────────────┐
│ User │
└──────────────┬─────────────┘
│
▼
┌────────────────────────────┐
│ Agent Harness │
│ (software interface) │
└──────┬──────────────┬──────┘
│ │
▼ ▼
┌────────────┐ ┌────────────┐
│ Models │ │ Tools │
└────────────┘ └────────────┘
That's why. I was using Claude the other day to greenfield a side project and it wanted to do some important logic on the frontend that would have allowed unauthenticated users to write into my database.
It was easy to spot for me, because I've been writing software for years, and it only took a single prompt to fix. But a vibe coder wouldn't have caught it and hackers would've pwned their webapp.
no, we don't
no, I'm a competent engineer
maybe you've not worked with any
I've worked with many competent engineers and have built things people couldn't even google help for before AI existed, and that surpassed mine and my teams expectations both solo and in a team setting, none of them were done in one sitting, which is what you're suggesting. Everything is planned out, and done piecemeal.
For the record, I can one shot an AI model to do all of those things, with all the detail they need and get similar output as if I gave a human all those tasks, I know because I've built the exact tooling to loop AI around the same processes competent developers use, and it still can do all of it in record time.
Bullshit you can lol. If it's that trivial, create an instagram right now and post the code.
You’ve moved the goalposts so far that you’re now talking about a different game altogether.
Any competent tech company will have canned ways to do all of those things that have already been reviewed and vetted
Anyways, yes, if I know I'm gonna need it? Because every framework has reasonable defaults or libraries for all of those things, and if you're in a corporate environment, you have vetted ways of doing them
1. import middleware.whatever
2. configure it
3. done
Like, you don't write these things unless you need custom behavior.
Features come after I have tested authn and authz.
The first feature runs with full security.
Then yeah, it makes sense.
I also find it wild how we’re sleepwalking into this, but I’m also part of the problem and using these things too.
(1) Tooling to enable better evaluation of generated code and its adherence to conventions and norms (2) Process to impose requirements on the creation/exposure of PRDs/prompts/traces (3) Management to guide devs in the use of the above and to implement concrete rewards and consequences
Some organizations will be exposed as being deficient in some or all of these areas, and they will struggle. Better organizations will adapt.
+1. I've tried many times, and failed, to replicate the joy of using that toolchain.
My clipart folder of that kid with the lolipop continues to stay relevant
The jury's still out on that one, because climate change is an existential risk.
>ah, they're so dumb, they don't get it, the anti-LLM people
This is one of the reasons I see AI failing in the short term. If I call you an idiot, are you more or less likely to be open minded and try what I'm selling? AI isn't making money, 95% of companies are failing with AI
https://fortune.com/2025/08/18/mit-report-95-percent-generat...
I mean, your AIs might be a lot more powerful if it was generating money, but that's not happening. I guess being condescending to the 95% of potential buyers isn't really working out.
Not obvious
> To me that statement is as obvious as "water is wet".
Well... is water *wet* or does it *wet things*? So not obvious either.
I'm really dubious when reading posts posing some things as obvious or trivial. In general they are not.
Water is not wet. Water makes things wet. Perhaps the inaccuracy of that statement should be taken as a hint that the other statements that you hold on the same level are worthy of reconsideration.
Or less.
And I don't think it's collar color they're going to be checking against.
So I guess I'm saying I agree that this is powerful and dangerous. These are language models, so they're more effective against humans and their languages. And self-preservation, empathy, humanity do not play a role as there is nobody in there to be offended at the notion of intentionally killing more than 9/10 of humanity… for some definitions of humanity, ones I'm sympathetic to.
First, we currently have 4 frontier labs, and a bunch of 2nd tier ones following. The fact that we don't have just oAI or just Anthropic or just Google is good in the general sense, I would say. The 4 labs racing each other and trading SotA status for ~a few weeks is good for the end consumer. They keep each other honest and keep the prices down. Imagine if Anthropic could charge 60$ /MTok or oAI could charge 120$ /MTok for their gpt4 style models. They can't in good part because of the competition.
Second, there's a bunch of labs / companies that have released and are continuing to release open mdoels. That's as close to "intelligence on tap" as you can get. And those models are ~6-12 months behind the SotA models, depending on your usecase. Even though the labs have largely different incentives to do so, a lot of them are still releasing open models. Hopefully that continues to hold. So not all control will be in the hands of big tech, even if the "best" will still be theirs. At some point "good enough" is fine.
There's also the thing about geopolitics being involved in this. So far we've seen the EU jumping the gun on regulation, and we're kinda sorta paying for it. Everyone is still confused about what can or cannot be done in the EU. The US seems to be waiting to see what happens, and China will do whatever they do. The worst thing that can happen is that at some point the big players (Anthropic is the main driver) push for regulatory capture. That would really suck. Thankfully atm there's this lingering thinking that "if we do it, the others won't so we'll be on the back foot". Hopefully this holds, at least until the "good enough" from above is out :)
The AI labs started down this path using the Manhattan Project as a metaphor and guess what? It's a good metaphor and we should embrace most of the wider implications of that (though I'd love to avoid all the MAD/cold war bullshit this time).