I keep being told this and the tools keep falling at the first hurdle. This morning I asked Claude to use a library to load a toml file in .net and print a value. It immediately explained how it was an easy file format to parse and didn’t need a library. I undid, went back to plan mode and it picked a library, added it and claimed it was done. Except the code didn’t compile.

Three iterations later of trying to get Claude to make it compile (it changed random lines around the clear problematic line) I fixed it by following the example in the readme, and told Claude.

I then asked Claude to parse the rest of the toml file, whereby it blew away the compile fix I had made..

This isn’t an isolated experience - I hit these fundamental blocking issues with pretty much every attempt to use these tools that isn’t “implement a web page”, and even when it does that it’s not long before it gets tangled up in something or other…

The real magic of LLMs comes when they iterate until completion until the code compiles and the test passes, and you don't even bother looking at it until then.

Each step is pretty stupid, but the ability to very quickly doggedly keep at it until success quite often produces great work.

If you don't have linters that are checking for valid syntax and approved coding style, if you don't have tests to ensure the LLM doesn't screw up the code, you don't have good CI, you're going to have a bad time.

LLMs are just like extremely bright but sloppy junior devs - if you think about putting the same guardrails in place for your project you would for that case, things tend to work very well - you're giving the LLM a chance to check its work and self correct.

It's the agentic loop that makes it work, not the single-shot output of an LLM.

There are techniques that can help deal with this but none of them work perfectly, and most of the time some direct oversight from me is required. And this really clips the potential productivity gains, because in order to effectively provide oversight you need to page in all the context of what's going on and how it ought to work, which is most of what the LLMs are in-theory helping you with.

LLMs are still very useful for certain tasks (bootstrapping in new unfamiliar domains, tedious plumbing or test fixture code), but the massive productivity gains people are claiming or alluding to still feel out of reach.

For instance, if you are working on a compiler and have a huge test database of code to compile that all has tests itself, "all sample code must compile and pass tests, ensuring your new optimizer code gets adequate branch coverage in the process" - the underlying task can be very difficult, but you have large amounts of test coverage that have a very good chance at catching errors there.

At the very least "LLM code compiles, and is formatted and documented according to lint rules" is pretty basic. If people are saying LLM code doesn't compile, then yes, you are using it very incorrectly, as you're not even beginning to engage the agentic loop at all, as compiling is the simplest step.

Sure, a lot of more complex cases require oversight or don't work.

But "the code didn't compile" is definitely in "you're holding it wrong" territority, and it's not even subtle.

But honestly I think sane code organization is the bigger hurdle, which is a lot harder to get right without manual oversight. Which of course leads to the temptation to give up on reviewing the code and just trusting whatever the LLM outputs. But I'm skeptical this is a viable approach. LLMs, like human devs, seem to need reasonably well-organized code to be able to work in a codebase, but I think the code they output often falls short of this standard.

(But yes agree that getting the LLM to iterate until CI passes is table-stakes.)

I think getting good code organization out of an LLM is one of the subtler things - I've learned quite a bit about what sort of things need to be specified, realizing that the LLM isn't actively learning my preferences particularly well, so there are some things about code organization I just have to be explicit about.

Which is more work, but less work than just writing the code myself to begin with.

If you read my post, you’d see that Claude code didn’t do that, I had to intervene in the agent loop and when I did it undid my fixes.

In fact, I find that with a strict feedback loop set up (i.e. a lot of lint rules, a strict type checker and fast unit tests), it will almost always generate what I want.

As someone else said, each step might be pretty stupid, but if you have a fast iteration loop, it can run until everything passes cleanly. My recommendation is to specify what really counts as "done" in your AGENTS.md/CLAUDE.md.

I didn't give my Agent any tools, it hallucinated code.

...well, yes. It's like evaluating a programmer's skill by having one-shot a program on paper with zero syntax errors.

And this was from two people who were 100% aligned on agentic AI coding. I've been using AI for years now and agentic AI for several months now. I was told that I bring out the worst in LLMs. Except... I was able to achieve better results, WITH LLMs, on OTHER frameworks. So like, ?

It may be easier to draw the boundary between "AI and non AI users", but as AI becomes prolific, the us vs them angle that people keep using won't apply anymore.

The age old "user versus tool" debate goes on, but it seems like gaslighting is popular these days. I classify it as gaslighting because I'm clearly a falsifiable test case, I'm even gung ho for LLMs, yet any kind of dissent is immediately warped into user error. It doesn't matter what you say or where you are on the spectrum, if you have 1 bad experience and speak up about it it's an issue. Guess that's not really an issue for the human condition though.

I don't know anything about DotNet, but I just fired up Claude Code in an empty directory and asked it to create an example dotnet program using the Tomlyn library, it chugged away and ~5 minutes later I did a "grep Deserialize *" in the project and it came up with exactly the line you wanted (in your comment here) it to produce: var model = TomlSerializer.Deserialize<TomlTable>(tomlContent)!;

The full results of what it produced are at https://github.com/linsomniac/tomlynexample

That includes the prompt I used, which is:

Please create a dotnet sample program that uses the library at https://github.com/xoofx/Tomlyn to parse the TOML file given on the command line. Please only use the Tomlyn library for parsing the TOML file. I don't have any dotnet tooling installed on my system, please let me know what is needed to compile this example when we get there. Please use an agent team consisting of a dotnet expert, a qa expert, a TOML expert a devils advocate and a dotnet on Linux expert.

I can't really comment on the code it produced (as I said, I don't use dotnet, I had to install dotnet on my system to try this), so I can't comment on the approach. 346 lines in Program.cs seems like a lot for an example TOML program, but I know Claude Code tends to do full error checking, etc, and it seems to have a lot of "pretty printing" code.

When I prompted, it went spelunking in the parent folder for a bunch of stuff, which I interrupted and told it to focus on its current folder. It came up with a sensible plan, which I let it execute. I then switched to a new tab and ran `dotnet run publishing/publish.cs` - like claude says. And the script fails at the first hurdle. The agentic loop came up with a plan, claimed it was done, but _didn't execute the verifiaction it told me it was going to execute_.

[0] https://gist.github.com/donalmacc/e91955be9792a983c375a098ba...

Legit this morning Claude was essentially unusable for me

I could explicitly state things it should adjust and it wouldn't do it.

Not even after specifying again, reverting everything eventually and reprompt from the beginning etc. Even super trivial frontend things like "extract [code] into a separate component"

After 30 minutes of that I relented and went on to read a book. After lunch I tried again and it's intelligence was back to normal

It's so uncanny to experience how much ita performance changes - I strongly suspect anthropic is doing something whenever it's intelligence drops so much, esp. Because it's always temporary - but repeatable across sessions if occuring... Until it's normal again

But ultimately just speculation, I'm just a user after all

And then the next step is to dynamically vary resources based on prediction of user stickiness. User is frustrated and thinking of trying competitor -> allocate full resources. User is profiled as prone to gambling and will tolerate intermittent rewards -> can safely forward requests to gimped models. User is an resolute AI skeptic and unlikely to ever preach the gospels of vibecoding -> no need to waste resources on him.

Its quite likely this is already happening buddy...

The 'random' degradation across all LLM-based services is obvious at this point.

Honestly, this is my experience. Every now and again it just completely self implodes and gives up, and I’m left to pick up the pieces. Look at the other replies who are making sure I’m using the agrntic loop/correct model/specific enough prompt - I don’t know what they’re doing but I would love to try the tools they’re using.

Maybe Anthropic is trying to cut costs a little and we are all just gaslighting ourselves into thinking its our problem.

Like when I was trying to find a physical store again with ChatGPT Pro 5.4 and asked it to prepare a list of candidates, but the shop just wasn't in the list, despite GPT claiming it to be exhaustive. When I then found it manually and asked GPT for advice on how I could improve my prompting in the future, it went full "aggressively agreeable" on me with "Excellent question! Now I can see exactly why my searches missed XY - this is a perfect learning opportunity. Here's what went wrong and what was missing: ..." and then 4 sections with 4 subsections each.

It's great to see the AI reflect on how it failed. But it's also kind of painful if you know that it'll forget all of this the moment the text is sent to me and that it will never ever learn from this mistake and do better in the future.

If 80% of the time they 10x my output, and the other 20% I can say "well they failed, I guess this one I have to do manually" - that's still an absolutely massive productivity boost.

Keep in mind that the AI is not reflecting, and it has no idea it made a mistake. It's just generating statistically-likely text for "apology" * "ai did not find results".

I wonder if it was getting blocked on searches or something, and just didn't tell you.

I try to give the model as little freedom as possible. That usually means it's not being used for novel work.

Then it crawls around for awhile, does some web searches, fetches docs from here and there, whatever. Sometimes it'll ask me some questions. And then it'll finally spit out a plan. I'll read through it and just give it a massive dump of issues, big and small, more questions I have, whatever. (I'll also often be spinning off new planning sessions for pre-work or ancillary tasks that I thought of while reviewing that plan). No structure or anything, just brain dump. Maybe two rounds of that, but usually just one. And then I'll either have it start building, or I'll have it stash in the linear agent so I can kick it off later.

But that's the hard part! You can only eke out moderate productivity gains by automating the tedium of actually writing out the code, because it's a small fraction of software engineering.

The code it needed to write was:

    var model = TomlSerializer.Deserialize<TomlTable>(toml)!;

I remember having to write code on paper for my CS exams, and they expected it to compile! It was hard but I mostly got there. definitely made a few small mistakes though

The only reason I wasted so much time was that given all of my other positive experiences, clearly I must have been prompting it incorrectly, and I should learn how to fix it.

After blowing four hours I just changed the colors by hand. I took away that it's almost completely unpredictable what will be easy and what will be hard for an LLM.

Friday afternoon I made a new directory and told Claude Code I wanted to make a Go proxy so I could have a request/callback HTTP API for a 3rd party service whose official API is only persistent websocket connections. I had it read the service’s API docs, engage in some back and forth to establish the architecture and library choices, and save out a phased implementation plan in plan mode. It implemented it in four phases with passing tests for each, then did live tests against the service in which it debugged its protocol mistakes using curl. Finally I had it do two rounds of code review with fresh context, and it fixed a race condition and made a few things cleaner. Total time, two hours.

I have noticed some people I work with have more trouble, and my vague intuition is it happens when they give Claude too much autonomy. It works better when you tell it what to do, rather than letting it decide. That can be at a pretty high level, though. Basically reduce the problem to a set of well-established subproblems that it’s familiar with. Same as you’d do with a junior developer, really.

Equating "junior developers" and "coding LLMs" is pretty lame. You handhold a junior developers so, eventually, you don't have to handhold anymore. The junior developer is expected to learn enough, and be trusted enough, to operate more autonomously. "Junior developers" don't exist solely to do your bidding. It may be valuable to recognize similarities between a first junior developer interaction and a first LLM interaction, but when every LLM interaction requires it to be handheld, the value of the iterative nature of having a junior developer work along side you is not at all equivalent.

I simply said the description of the problem should be broken down similar to the way you’d do it for a junior developer. As opposed to the way you’d express the problem to a more senior developer who can be trusted to figure out the right way to do it at a higher level.

What’s giving too much autonomy about

“Please load settings.toml using a library and print out the name key from the application table”? Even if it’s under specified, surely it should at least leave it _compiling_?

I’ve been posting comments like this monthly here, my experience has been consistently this with Claude, opencode, antigravity, cursor, and using gpt/opus/sonnet/gemini models (latest at time of testing). This morning was opus 4.6

Are you using Claude Code? Do yo have it configured so that you are not allowing it to run the build? Because I've observed that Claude Code is extremely good at making sure the code compiles, because it'll run a compile and address any compile errors as part of the work.

I just asked it to build a TOML example program in DotNet using Tomlyn, and when it was done I was able to run "./bin/Debug/net8.0/dotnettoml example.toml", it had already built it for me (I watched it run the build step as part of its work, as I mentioned it would do above).

> I’ve observed Claude code is extremely good at making sure the code compiles

My observation is that it’s fine until it’s absolutely not, and the agentic loop fails.

I don't know that it's useful to assign blame here.

It probably is to your benefit, if you are a coding professional, to understand why your results are so drastically different from what others are seeing. You started this thread saying "I keep getting told I'll be amazed at what it can do, but the tools keep failing at the first hurdle."

I'm telling you that something is wrong, that is why you are getting poor results. I don't know what is wrong, but I've given you an example prompt and an example output showing that Claude Code is able to produce the exact output you were looking for. This is why a lot of people are saying "you'll be amazed at what it can do", and it points to you having some issue.

I don't know if you are running an ancient version of Claude Code, if you are not using Opus 4.6, you are not using "high" effort (those are what I'm using to get the results I posted elsewhere in reply to your comment), but something is definitely wrong. Some of what may be wrong is that you don't have enough experience with the tooling, which I'd understand if you are getting poor results; you have little (immediate) incentive to get more proficient.

As I said, I was able to tell Claude Code to do something like the example you gave, and it did it and it built, without me asking, and produced a working program on the first try.

Oh - I’m blaming Claude not anyone else. I’ve tried again this evening and the same prompt (in the same directory on the same project) worked.

> i don’t know if you’re using an ancient version of Claude code,

I’m on a version from some time last week, and using opus 4.6

> This is why a lot of people are saying "you'll be amazed at what it can do", and it points to you having some issue.

If you look at my comments in these threads, I’ve had these issues and been posting about this for months. I’m still being told “ you’re using the wrong model or the wrong tool or you’re holding it wrong” but yet, here I am.

I’m using plan mode, clearly breaking down tasks and this happens to me basically every time I use the damn tool. Speaking to my team at work and friends in other workplaces, I hear the same thing. But yet we’re just using it wrong or doing something wrong,

Honestly, I genuinely think the people who are not having these experiences just… don’t notice that they are.

For some reason I'm getting downvoted for trying to help, but regardless, if you can (and want to) post a transcript somewhere of some of these sessions that aren't working out, maybe some of us who are having more success can take a look and see what's up.

Fresh prompt in a codebase with a claude.md I asked it to do a simple task. I've shared the prompt, plan, output in the linked gist.

We’ve gone from “I’m baffled at your experience” to well yeah it often fails” in two sentences here…

I also clearly said I didn’t allow it one output, I gave it the compile error message, it changed a different line, I told it it was at the affected line and to check the docs. Claude code then tried to query the DLL for the function, abandoned that and then did something else incorrect.

I’m literally asking it to install a package and copy the example from the readme

Amazon has demonstrated that it takes just as longer, or longer, to have senior devs review LLM output than it would to just have the senior devs do the programming in the first place. But now your senior devs are wasted on reviewing instead of developing or engineering. Amazon, Microsoft, Google, Salesforce, and Palantir have all suffered multiple losses in the tens of millions (or more) due to AI output issues. Now that Microsoft has finally realized how bad LLMs really are at generating useful output, they've begun removing AI functionality from Windows.

Product quality matters more than time to market. Especially in tech, the first-to-market is almost never the company that dominates, so it's truly bizarre that VCs are always so focused on their investments trying to be first to market instead of best to market.

If Competitor Y just fired 90% of their developers, I would have a toast with my entire human team. And a few months later, we'd own the market with our superior product.

At my company, we are maintaining our hiring plan (I'm the decision maker). We have never been more excited at our permission to win against the incumbents in our market. At the same time, I've never been more concerned about other startups giving us a real run. I think we will see a bit of an arms race for the best talent as a result.

Productivity without clear vision, strategy and user feedback loops is meaningless. But those startups that are able to harness the productivity gains to deliver more complete and polished solutions that solve real problems for their users will be unstoppable.

We've always seen big gains by taking a team of say 8 and splitting it into 2 teams of 4. I think the major difference is that now we will probably split teams of 4 into 2 teams of 2 with clearer remits. I don't want them to necessarily delivery more features. But I do want them to deliver features with far fewer caveats at a higher quality and then iterate more on those.

Humans that consume the software will become the bottlenecks of change!

They'd be unstoppable irrespective of LLMs. Why do you think Zuckerberg acquired Instagram? He literally tried copying it and failed. Instagram at the time was absolutely tiny in terms of pure labour, relative to Facebook.

Most people on hacker news are missing the point. Productivity gains for the sake of perceived productivity gains is not what creates economic value. Its not the equivalent of a factory all of a sudden becoming more productive in producing more of the same stuff. Not comparable at all.

Ruby on Rails and its imitators blew away tons of boilerplate. Despite some hype at the time about a productivity revolution, it didn’t _really_ change that much.

> , libraries, build-tools,

Ensure what you mean by this; what bearing do our friends the magic robots have on these?

> and refactoring

Again, IntelliJ did not really cause a productivity revolution by making refactoring trivial about 20 years ago. Also, refactoring is kind of a solved problem, due to IntelliJ et al; what’s an LLM getting you there that decent deterministic tooling doesn’t?

the ones i've used come with defaults that you can then customize. here are some of the better ones:

- https://guides.rubyonrails.org/command_line.html#generating-...

- https://hexdocs.pm/phoenix/Mix.Tasks.Phx.Gen.html

- https://laravel.com/docs/13.x/artisan#stub-customization

- https://learn.microsoft.com/en-us/aspnet/core/fundamentals/t...

> my experience has been these get left behind as the service implementations change

yeah i've definitely seen this, ultimately it comes down to your culture / ensuring time is invested in devex. an approach that helps avoid drift is generating directly from an _actual_ project instead of using something like yeoman, but that's quite involved

> it comes down to ensuring time is invested in devex

That’s actually my point - the orgs haven’t invested in devex buyt that didn’t matter because copilot could figure out what to do!

And of course, we didn't see a massive layoff after the introduction of say, StackOverflow, or DreamWeaver, or jQuery vs raw JS, Twitter Bootstrap, etc.

If your dev group is spending 90% of their time on these... well, you'd probably be right to fire someone. Not most of the developers but whoever put in place a system where so much time is spent on overhead/retrograde activities.

Something that's getting lost in the new, low cost of generating code is that code is a burden, not an asset. There's an ongoing maintenance and complexity cost. LLMs lower maintenance cost, but if you're generating 10x code you aren't getting ahead. Meanwhile, the cost of unmanaged complexity goes up exponentially. LLMs or no, you hit a wall if you don't manage it well.

My company has 20 years of accumulated tech debt, and the LLMs have been pretty amazing at helping us dig out from under a lot of that.

You make valid points, but I'm just going to chime in with adding code is not the only thing that these tools are good at.

Even the most enthusiastic AI boosters I've read don't seem to agree with this. From what I can tell, LLMs are still mostly useful at building in greenfield projects, and they are weak at maintaining brownfield projects

And I have more examples, where I, personally, was on the both sides of the fence: defined by my level of the same problem understanding, not by the tooling.

Dude that's everybody in charge. You're young, you build a system, you mold it to shifting customer needs, you strike gold, you assume greater responsibility.

You hire lots of people. A few years go by. Why can't these people get shit done? When I was in their shoes, I was flying, I did what was needed.

Maybe we hired second rate talent. Maybe they're slacking. Maybe we should lay them off, or squeeze them harder, or pray AI will magically improve the outcome.

Come on, look in the mirror.

If it is writing both the code and the tests then you're going to find that its tests are remarkable, they just work. At least until you deploy to a live state and start testing for yourself, then you'll notice that its mostly only testing the exact code that it wrote, its not confrontational or trying to find errors and it already assumes that its going to work. It won't ever come up with the majority of breaking cases that a developer will by itself, you will need to guide it. Also while fixing those the odds of introducing other breaking changes are decent, and after enough prompts you are going to lose coherency no matter what you do.

It definitely makes a lot of boilerplate code easier, but what you don't notice is that its just moving the difficult to find problems into hidden new areas. That fancy code that it wrote maybe doesn't take any building blocks, lower levels such as database optimization etc. into account. Even for a simple application a half-decent developer can create something that will run quite a bit faster. If you start bringing these problems to it then it might be able to optimize them, but the amount of time that's going to take is non-negligible.

It takes developers time to sit on code, learn it along with the problem space and how to tie them together effectively. If you take that away there is no learning, you're just the monkey copy-pasting the produced output from the black box and hoping that you get a result that works. Even worse is that every step you take doesn't bring you any closer to the solution, its pretty much random.

So what is it good for? It can both read, "understand", translate, write and explain things to a sufficient degree much faster than us humans. But if you are (at the moment) trusting it at anything past the method level for code then you're just shooting yourself in the foot, you're just not feeling the pain until later. In a day you can have it generate for example a whole website, backend, db etc. for your new business idea but that's not a "product", it might as well be a promotional video that you throw away once you've used it to impress the investors. For now that might still work, but people are already catching on and beginning to wise up.

I feel like most folks commenting uncritically here about second coming of Jesus must work in some code sweatshops, churning eshops or whatever is in vogue today quickly and moving on, never looking back, never maintaining and working on their codebases for decade+. Where I existed my whole career, speed of delivery was never the primary concern, quality of delivery (which everybody unanimously complaints one way or another with llms) was much more critical and thats where the money went. Maybe I just picked up right businesses, but then again I worked for energy company, insurance, telco, government, army, municipality, 2 banks and so on across 3 European states. Also, full working code delivery to production is a rather small part of any serious project, even 5x speed increase would move the needle just a tiny bit overall.

If I would be more junior, I would feel massive FOMO from reading all this (since I use it so far just as a quicker google/stackoverflow and some simpler refactoring, but oh boy is is sometimes hilariously and dangerously wrong). I am not, thus I couldn't care less. Node.js craze, 10x stronger, not seeing forest for the trees.

Any sensible experienced programmer will take into account possible future features spoken about around the office, when deciding how implement a feature. To me that’s the key reason I can’t delegate AI complex features that will grow. The same reason I inspect and clean the toilet when I’m done - out of respect to my colleagues

Another way of increasing profit is to simply reduce your headcount by 90% while keeping the same profit.*

Hence, I think some companies will keep downsizing. Some companies will hire. It depends a lot.

*Assuming 90% productivity increase.

Is it the same with tech? Facebook has 3 billion monthly active users. No amount of tech will bring that up to 6 billion. If you were to double the amount of time someone spends on Facebook, or double the ads they see or double the click through rate, what does that really mean?

Taking the example of Facebook. They are in social media, messaging, AI, VR/AR hardware and software, a few other things, meta universe whatever that was, now left with the name. Facebook isn't delivering or successful on all its ventures, it knows that, it keeps investing in other segments.

More productivity would mean at least diversifying, they have some of the best engineers, it would make no sense to not simply attempt to hit the jackpot by playing more machines.

What fewer people talks about is that the entire tech industry is tertiary services. Ads, entertainment, communication, etc. If/when hard industries take a hit, tertiary takes a hit. If it isn't clear to you that the overall economy has already started to take some irreversible dents, and that those will accelerate, know that the capital is well aware.

Or we can continue wishful thinking and seek comfort that monetary tightening is just temporary, investments will flow more into tangent ventures and growth is around the corner, the U.S still is and will remain the world's strongest economy.

Even if it succeeded, having more buttons everywhere - marketplace, reels, stories, blogs, articles, groups, etc. A lot of this make the whole experience more crowded.

Productivity here has an odd effect, sort of like larger roads causing more traffic jams.

I think most companies are making the right call by downsizing instead of staying same size. Let people go to where there is more potential for growth.

I have a vague memory of a hotel reservation "app" that was just putting lines in a Google Sheet where the founders grabbed the info, called the hotels themselves and reserved the room.

When they got too many requests to keep up, they slowly automated the bits they could one by one.

Smart organizations will not just deliver better products but likely start products that they were hesitant to start before because the cost of starting is a lot closer to zero. Smart engineering leadership will encourage developers into delivering value and not self-serving, endless iterations of tooling enhancements, etc.

If I was a CTO and my competitor Y fired 90% of their devs, I'd try to secure funding to hire their top talent and retain them. The vitriol alone could fuel some interesting creations and when competitor Y realizes things later, their top talent will have moved on.

>> Smart organizations will not just deliver better products but likely start products [...]

This is not the 90s anymore when low hanging fruit was everywhere ready to be picked. We have everything under the sun now and more.

The problem with bullshit apps is not that it took you 5 months to build. What you build now in 5 minutes it's still bullshit. Most of the remaining work is bullshit jobs. Spinning useless "features" and frameworks that nobody needs and shove them down the throat of customers that never asked for them. Now it's possible to dig holes and fill them back (do pointless work) at much improved pace thanks to AI.

Seriously, all managers I know (2, somehow) that have done this are now in sprint hell because they didn't realize an LLM cannot do the work their devs did. Bit of schadenfreude.

Not saying this can't be done, some projects are easy enough that you can throw an LLM at it

For greenfield development you don't need as many software engineers. Some developers (the top 10%) are still needed to guide AI and make architectural decisions, but the remaining 90% will work on the lifecycle management task mentioned above.

The productivity gains can be used to produce more software, and if you are able to sell the software you produce should result in a revenue boost. But if you produce more than you can sell then some people will be laid off.

Terafab is suddenly making so much sense!

I was working with it just yesterday and found it mixed up class attributes and dict keys when looking if something existed. As a result, it always thought the key was missing. Glancing through the code, it would have been easy to overlook, as it was checking for something... just doing it the wrong way. The code also appeared to still work in the larger context, just not as well or as efficiently as it was supposed to work. It was only my obsessive testing and attention to detail that caught it.

I find it much faster and easier to write things myself, using AI to help me along the way, rather than trying to delegate everything to the AI and hoping for the best. Writing code is much more interesting than reviewing code, and it puts me close enough to it that those kind of issues don't happen in the first place.

The companies gutting engineering right now are going to realise too late that AI amplifies what you already have. If you fire the people with institutional knowledge and domain expertise, you're left with an AI that generates confident, well-formatted mediocrity. Agree or not?

It's not a direct multiplier, because even non-programmers can get stuff done with it just from an idea.

And it's not exactly an exponent either, even though experienced programmers can get absolutely massive gains when they learn to worth with it.

"Replace tools, not jobs" is the best AI slogan I've found yet. Use the VC funded practically free massive language models to quickly write every single small utility, tool, connector and dashboard you've ever wanted.

You know, the ones you've always wanted to do, but could never justify the time spent on them compared to the utility. Specifically in corporate environments. Stuff that used to require a project manager, programmer and a QA person dedicated to some tiny app for a week or two can now be done in two days by a single person with zero coding experience and an idea to a degree where you can be evaluate whether it's worth continuing or not.

CTO is rewriting company platform (by himself with AI) and is convinced it's 100x productivity. But when you step back and look at the broader picture, he's rewriting what something like Rails, .NET, or Spring gave us 15-20 years ago? It's just in languages and code styles he is (only) familiar with. That's not 100x for the business, sorry...

I think companies that maintain the same product roadmap, will stagnate irrespective of whether they keep the engineers or fire 90% of them. The ones that will win, in my humble opinion, will be the ones that ask, "What can we build now, that we could not build before due to any reason (cost, expertise etc.)?" That is an entirely different question.

Just like financial markets, this phase is very volatile. But today’s price can look much different in 1, 2, 5, and 10 years.

Right now there’s a lot of opportunity to disrupt companies who are moving much slower since adoption can vary greatly.

But I think in a few years this meta will be overcrowded. The overall skill/productivity gap across companies will be reduced and the bar for productivity will be raised.

There’s room for taking profits now from the productivity gains but I don’t think it will last long.

If AI is really increasing productivity enough to reduce headcount right now, it won’t be in future. If all your competitors are using AI as effectively as you are, can you still do this?

The world’s demand for productivity is limitless. As of right now you still need someone to at least install Claude Code and run the binary.

The more grounded reality is that AI coding can be a productivity multiplier in the right hands, and a significant hindrance in the wrong hands.

Somewhere there exists a happy medium between vibe coding without ever looking at the code, and hand-writing every single line.

The main effect is that you build things a lot faster.

The secondary effect is that, since you can code a lot faster, you are no longer scared of large refactors, ports to other languages, and all that kind of tedious stuff.

That means that if you design something, ship it, and realize it's bad, you can completely change it.

The "oh we should have done it differently, but now it would take years to change, so we have to deal with legacy garbage" argument does really make sense anymore.

Like if HN had LLM agents, we wouldn't have had to wait for years for them to solve the "click here for more comments" problem. And iterating with users is a lot faster.

After combine harvester, we produced the same food with less people.

At the moment, it seems like hardware is the constraint. Companies don't have access to enough machines or tokens to keep all their devs occupied, so they let some go. Maybe that changes, maybe we already have too much software?

Personally I think we already had too much software before LLMs and even without them many devs would have found themselves jobless as startups selling to startups selling to startups failed and we realized (again) that food, shelter, security, education etc are 'real' industries, but software isn't one if it's not actively helping one of those.

Unfortunately this kind of software needs specialised domain knowledge to produce that AI doesn't have yet, but when (if) it arrives I hope we see strides forwards in hardware engineering productivity.

Companies without the same constraints are well equipped to keep who they've got, pivot them into managing/overseeing agents to scale, and build better products from the outset.

So this'll be a good opportunity for smaller companies (or not-for-profits like co-ops and credit unions) to eat the lunches of bigger companies that'll be slow to adapt.

Remember sometimes the most productive thing to have is not money or people but time with your ideas.

I see it's necessary ALL the time. The AI generated code can be used as scaffolding, but it's newer get close to real production quality. The expierence from small startup with team of 5 developers. I do review and approve all PRs, and none ever able to pass AI code review from the first iteration.

Seniors can adjust, but eg. junior frontend-only devs might be doomed in both situations, as they might not be able to contribute enough to business-critical features to justify their costs and most frontend-related tasks will be taken over by the "10x" seniors.

If it is a big company the answer is and will always be: whatever makes the stock price rise the most.

This is a crazy take. Even if said people are matching or exceeding the outcome of those using the technology?

I’m not in this group. But the closest analog to what you are saying is firing people for not using a specific IDE.

I think that it's more along the lines of "do you fire people" instead of just "do you fire devs". Fewer devs means less of a need for PMs, so they can be let go as well, and maybe with the rise of AI assisted design tools, you don't need as many UX people, so you let some of them go as well.

As for building better products, I feel like that's a completely different topic than using AI for productivity gains, but only because at the end of the day you need buy in from upper management in order to build the features/redo existing features/both that will make the product better. I should also mention I'm viewing this from the position of someone who works at an established company and not a startup, so it may differ.

I'm not sure what your circumstances are but even if it's not true for you, it's true for many other people.

People online with identical views to them all assure me that theyre all highly skilled though.

Meanwhile I've been experimenting using AI for shopping and all of them so far are horrendous. Cant handle basic queries without tripping over themselves.

But you can understand why all the 1700 and below chess players say it is good and it is making them better using it for eval?

Don't worry, AI will replace you one day, you are just smarter than most of us so you don't see it yet.

This kind of thinking is actually a big reason why execs are being misinformed into overestimating LLM abilities.

LLM coding agents alone are not good enough to replace any single developer. They only make a developer x% faster. That dev who is now x% faster may then allow you to lay off another dev. That is a subtle yet critical difference.

To adress your point, let's try another one analogy. Imagine secreterial assistants, discussing their risk of been replaced by computers in the 80s. They would think: someone still need to type those letters, sit next to that phone and make those appointments, I am safe. Computers won't replace me.

It is not that AI will do all of your tasks and replace you. It is that your role as a specialist in software development won't be necessary most of the time (someone will do that, and that person won't call themselves a programmer).

I tend to find that the volume of automation predictions inversely correlates to how real they are.

When capitalists actually have the automation tech they dont shout about it they just do it quietly and collect the profits.

When, say, Bezos is worried about his unionizing workforce and wants to intimidate - that's when the hot takes and splashy media articles about billions invested in automation "coming for yer jerb" you read about get published.

For me the main difference is now some people can explain what their code does. While some other only what it wants to achieve

Most people dont really understand coding but shopping is a far simpler task and so it's easier to see how and where it fails (i.e. with even mildly complex instructions).

Im very confident the experts in every field are not all that impressed by LLMs, relative of course, to those who were 'meh' in the first place. Experts meaning those who actually understand the content, not simply regurgitate or repeat it on demand.

I'd even go as far as to say there are many out there who have a feeling of disdain of the experts and want to see LLMs flourish because of this.

On the tech side I see it saving some time with stuff like mock data creation, writing boiler plate, etc. You still have to review it like it's a junior. You still have to think about the requirements and design to provide a detailed understanding to them (AI or junior).

I don't think either of these will provide 90% productivity gains. Maybe 25-50% depending on the job.

Because it does.

> I still don't see ANY proof that it doesn't generate a total unmaintainable unsecure mess, that since you didn't develop, you don't know how to fix.

I wouldn't know since it's been years since I've tried but I'd imagine that Claude Code would indeed generate a half-baked Next.js monstrosity if one-shot and left to its own devices. Being the learned software engineer I am, however, I provide it plenty of context about architecture and conventions in a bootstrapped codebase and it (mostly) obeys them. It still makes mistakes frequently but it's not an exaggeration to say that I can give it a list of fields with validation rules and query patterns and it'll build me CRUD pages in a fraction of the time it'd take me to do so.

I can also give it a list of sundry small improvements to make and it'll do the same, e.g. I can iterate on domain stuff while it fixes a bunch of tiny UX bugs. It's great.

Ah, is this what you do for a job? I actually have to accord accounting rules between countries, evaluate different conflicting laws, understand what the government wants in a report, check that everything sums up... But if my job was CRUD apps and thats it, yes I would fear for my job.

> if my job was CRUD apps and thats it,

I'm not sure whether you intended this but your post reads as presumptuous and disdainful. The whole point of these tools is to free people from having to do these parts of their job so that they can focus on the ones of greater substance. At no point did I say that was my entire job or that I was afraid of it disappearing - if that was your takeaway then I don't know what to say.

Sure it is not as fast to understand as code I wrote. But at least I mostly need to confirm it followed how it implemented what I asked. Not figuring out WHAT it even decided to implement in the first place.

And in my org, people move around projects quite a bit. Hasn’t been uncommon for me to jump in projects with 50k+ lines of code a few times a year to help implement a tricky feature, or help optimize things when it runs too slow. Lots of code to understand then. Depending on who wrote it, sometimes it is simple: one or two files to understand, clean code. Sometimes it is an interconnected mess and imho often way less organized that Ai generated code.

And same thing for the review process, lots of having to understand new code. At least with AI you are fed the changes a a slower pace.

not talking about toys or vibecoded crap no one uses.

Weirdly, people who have actually created functional one-man products don't seem to have the same problem, as they welcome the business.

"oooh Im afraid of doxxing myself", wtf? lmao!

We are very much in need of an actual way to measure real economic impact of AI-assisted coding, over both shorter and longer time horizons.

There's been an absolute rash of vibecoded startups. Are we seeing better success rates or sales across the industry?

That's the same false argument that the religious have offered for their beliefs and was debunked by Bertrand Russell's teapot argument: https://en.wikipedia.org/wiki/Russell%27s_teapot

I'm saying that I think this is an important enough question that I think we should seek real evidence in either direction, especially since apparently everyone already has a strong opinion (warranted or not.)

If you use it correctly, you can get better quality, more maintainable code than 75% of devs will turn in on a PR. The “one weird trick” seems to be to specify, specify, specify. First you use the LLM to help you write a spec (document, if it’s pre existing). Make sure the spec is correct and matches the user story and edge cases. The LLM is good at helping here too. Then break down separations of concerns, APIs, and interfaces. Have it build a dependency graph. After each step, have it reevaluate the entire stack to make sure it is clear, clean, and self consistent.

Every step of this is basically the AI doing the whole thing, just with guidance and feedback.

Once you’ve got the documentation needed to build an actual plan for implementation, have it do that. Each step, you go back as far as relevant to reevaluate. Compare the spec to the implementation plan, close the circle. Then have it write the bones, all the files and interfaces, without actual implementations. Then have it reevaluate the dependency graph and the plan and the file structure together. Then start implementing the plan, building testing jigs along the way.

You just build software the way you used to, but you use the LLM to do most of the work along the way. Every so often, you’ll run into something that doesn’t pass the smell test and you’ll give it a nudge in the right direction.

Think of it as a junior dev that graduated top of every class ever, and types 1000wpm.

Even after all of that, I’m turning out better code, better documentation, and better products, and doing what used to take 2 devs a month, in 3 or 4 days on my own.

On the app development side of our business, the productivity gain also strong. I can’t really speak to code quality there, but I can say we get updates in hours instead of days, and there are less bugs in the implementations. They say the code is better documented and easier to follow , because they’re not under pressure to ship hacky prototype code as if it were production.

On the current project, our team size is 1/2 the size it would have been last year, and we are moving about 4x as fast. What doesn’t seem to scale for us is size. If we doubled our team size I think the gains would be very small compared to the costs. Velocity seems to be throttled more by external factors.

I really don’t understand where people are coming from saying it doesn’t work. I’m not sure if it’s because they haven’t tried a real workflow, or maybe tried it at all, or they are definitely “holding it wrong.” It works. But you still need seasoned engineers to manage it and catch the occasional bad judgment or deviation from the intention.

If you just let it, it will definitely go off the rails and you’ll end up with a twisted mess that no one can debug. But use a system of writing the code incrementally through a specification - evaluation loop as you descend the abstraction from idea to implementation you’ll end up winning.

As a side note, and this is a little strange and I might be wrong because it’s hard to quantify and all vibes, but:

I have the AI keep a journal about its observations and general impressions, sort of the “meta” without the technical details. I frame this to it as a continuation of “awareness “ for new sessions.

I have a short set of “onboarding“ documents that describe the vision, ethos, and goals of the project. I have it read the journal and the onboarding docs at the beginning of each session.

I frame my work with the AI as working with it as a “collaborator” rather than a tool. At the end of the day, I remind it to update its journal of reflections about the days work. It’s total anthropomorphism, obviously, but it seems to inspire “trust” in the relationship, and it really seems to up-level the effort that the AI puts in. It kinda makes sense, LLMs being modelled on human activity.

FWIW, I’m not asserting anything here about the nature of machine intelligence, I’m targeting what seems to create the best result. Eventually we will have to grapple with this I imagine, but that’s not today.

When I have forgotten to warm-start the session, I find that I am rejecting much more of the work. I think this would be worth someone doing an actual study to see if it is real or some kind of irresistible cognitive bias.

I find that the work produced is much less prone to going off the rails or taking shortcuts when I have this in the context, and by reading the journal I get ideas on where and how to do a better job of steering and nudging to get better results. It’s like a review system for my prompting. The onboarding docs seem to help keep the model working towards the big picture? Idk.

This “system” with the journal and onboarding only seems to work with some models. GPT5 for example doesn’t seem to benefit from the journal and sometimes gets into a very creepy vibe. I think it might be optimized for creating some kind of “relationship” with the user.

I suspect you either already were or would’ve been great at leading real human developers not just AI agents. Directing an AI towards good results is shockingly similar to directing people. I think that’s a big thing separating those getting great results with AI from those claiming it simply does not work. Not everyone is good at doing high level panning, architecture, and directing others. But those that already had those skills basically just hit the ground running with AI.

There are many people working as software engineers who are just really great at writing code, but may be lacking in the other skills needed to effectively use AI. They’re the angry ones lamenting the loss of craft, and rightfully so, but their experience with AI doesn’t change the shift that’s happening.

My own experience...

I've tried approaching vibe coding in at least 3 different ways. At first I wrote a system that had specs (markdown files) where there is a 1 to 1 mapping between each spec to a matching python module. I only ever edited the spec, treating the code itself as an opaque thing that I ignore (though defined the intrefaces for). It kind of worked, though I realized how distinct the difference between a spec that communicates intent and a spec that specifies detail really is.

From this, I felt that maybe I need to stay closer to the code, but just use the LLM as a bicycle of the mind. So I tried "write the code itself, and integrate an LLM into emacs so that you can have a discussion with the LLM about individual code, but you use it for criticism and guidance, not to actually generate code". It also worked (though I never wrote anything more then small snippets of Elisp with it). I learned more doing things this way, though I have the nagging suspicion that I was actually moving slower than I theoretically could have. I think this is another valid way.

I'm currently experimenting with a 100% vibe coded project (https://boltread.com). I mostly just drive it through interaction on the terminal, with "specs" that kind of just act as intent (not specifications). I find the temptation to get out of the outside critic mode and into just looking at the code is quite strong. I have resisted it to date (I want to experiment with what it feels like to be a vibe coder who cannot program), to judge if I realistically need to be concerned about it. Just like LLM generated things in general, the project seems to get closer and closer to what I want, but it is like shaping mud, you can put detail into something, but it won't stay that way over time; its sharp detail will be reduced to smooth curves as you then switch to putting detail elsewhere. I am not 100% sure on how to deal with that issue.

My current thoughts is that we have failed to actually find a good way of switching from the "macro" (vibbed) to the "micro" (hand coded) view of LLM development. It's almost like we need modules (blast chambers?) for different parts of any software project. Where we can switch to doing things by hand (or at least with more intent) when necessary, and doing things by vibe when not. Striking the balance between those things that nets the greater output is quite challenging, and it may not even be that there is an optimal intersection, but simply that you are exchanging immediate change for future flexibility to the software?

Not hiring someone?