I now started to us AI to help review my juniors PRs, because I couldn't keep up with the amount of code they ship. It started poorly, but now I have my method: I first read the code and flag the lines I'm not sure about, then ask any frontier model (I like Claude here for analysis, even if I don't use it for the rest) to explain the PR and to put effort on the parts I flagged (basically explain in detail the code, not only the PR), and to search through the libraries. Sometimes it notices something I would have missed (like missing an 'order_by' or off by one errors, because the underlying lib wasn't coded like the original AI pretended it was).
I also changed the way I do review because it has been more than a year and the juniors/new hire are still lost, wether on domain knowledge for the older new hire, or just capabilities for the juniors, and discussing with other departments, it's the same for like 95% of them. Now, rather than correcting the PR or adding a request for change, I add a whole unit/functional test to the PR and let that as an exercise to pass the test. They can use AI but I tell them to try to find what part of the code doesn't work before generating the fix, hopefully they'll take ownership of the code if I keep doing that.
It's a different mindset. Has nothing to do with curiosity.
For me and you, "fixing" something means finding which assumption was violated and redesigning the solution in light of that. Sometimes this makes the solution smaller.
For a lot of people, "fixing" means "adding code to make it work". This always makes the solution larger.
even before AI very few developers cared about the actual craft. I'd say only about 3%. For a lot of people it's just a job, no matter if it's a corperate dev or a startup dev.
"because the underlying lib wasn't coded like the original AI pretended it was"
So, the models ARE dumb. They just are very good at finding patterns in their training data. I mean, when they code minecraft clones, it is not because they can cook up how to write minecraft clones, rather, their training data includes a lot of minecraft-like games code, and they just reuse that.
I think it depends on what area of the world you work. We recenty had one of our plant managers build a web portal to keep track of some of our operational tech, with features you wouldn't find in standard products. I've been turning it into a container app that can actually deploy safely into our cloud infrastrcture. It's quite frankly better quality than what most external software companies have provided us with in the past.
Don't get me wrong. It's not great. It would never pass any of our policies for things that actually operate stuff on the power grid, but as an administrative tool that can live in total isolation from the vital networks. It's perfectly fine. It's also not like we would have hired the best software companies to build it otherwise. We'd hire some low-level cheap consultant house who would then likely get cheap student labour to build it. With that in mind though, the AI is much better than what the realistic alternative would be.
Money wise it's also cheaper. It's been roughly €1000 + the time it's taken us both. If I had known they were doing it, I would have rolled out the developer cowork app/skills/whateveryoucallconfigurationsthesedays to them. This would have avoided their AI building it to be depoyed on a VM rather than in our managed k8s in our Azure. It would also have written the code a little different, used UV and maybe django rather than flask. But hey. For what it is, it's like a 90% cost saving compared to buying what would've been a less maintainable and lower quality system.
I think perhaps the greater issue will be finding people who want to extract the gold from the heap of shit and getting it to run in production. I don't personally mind, but it's not like any of my colleagues would've wanted the task.
By coincidence, we've been building software for five years that other plant managers are now considering using to keep track of some operational metrics. And I don't think we should put LLMs into our development cycle, simply because the data our software is built for can have critical meaning in some situations.
The insecurity in a vibe-coded web portal isn't that someone hacks it with XSS, it's that after the next vibe-coded release, some X quietly becomes −Y somewhere no one expects.
From this perspective, having no software at all might be better, or as in your case, safer.
I don't think you should keep AI away from software development for security reasons. We're in the European economic area and are completely NIS2 compliant and we use LLM's to aid in our software development for high risk systems. None of that is vibe coded though, so it's not like I disagree with you either.
As you point out this portal isn't that, but what protects us is the processes around compliance. This can't grow from X to Y because not even the CEO has the authority to overwrite our compliance gates. The EU is a tremendous help in this area since personal liability changed things completely.
I don't think they are opposite. We still don't let AI run anything we do. That being said, things have changed. 4 months ago I wouldn't touch the M365 Copilot thing if you paid me to, now it's basically the only AI I use professionally. Cowork changed the way we work, and I've previously spoken about how I think Microsoft sort of won the AI enterprise "war" by selling the tools and not the mode. If you'd asked me a year ago I would've been very anti AI in general. I got proven wrong. I do think it's going to be interesting to see where the EU goes with things though. Currently you can't use all features of Cowork as an example, because they are illegal in the EU. So you can't let it run your browser session for you as an example. I wonder how long that will last.
I still doubt we will ever give an AI access to run code on our systems directly though. In isolation, sure, but other than that.
This is a side note, but my personal favorite part of Cowork is that I can roll out our compliance policy to every developer as a Microsoft Teams app (no, that makes no sense to me either). So when they try to install some package that isn't pre-approved their Cowork agent won't let them and will instead explain how they might get approval. If they then continue to reference it, Cowork will even alert us.
What you're telling me now is that in the short period of 4 months, an American corporation took an employee of a company the EU relies on for critical matters from "wouldn't touch it if you paid me" to outsourcing thinking.
With the models we got at out disposal, you can easily deploy a kickass internal tool, it's fine even for mvp's unless you're not handling sensitive user data, shooting yourself in the foot is easier than ever, and talking about code smell for most application (especially frontend) does it even matter?
for an administrative tool the criticality is not determined by the network boundary but by which decisions within the plant are based on these figures
I doubt it would be better served as a customized off the shelf software. They certainly could have done it with our existing systems with SmartSheet, Microsoft Fabric and/or Power Apps with SharePoint for document storage.
From an enterprise perspective this becomes complicated for various reasons. RBAC is one area. In the perfect world you have a system to handle roles and rights to every system, something that you can give managers access to so they can maintain the access available to their employees, something linked with HR. In reality you have EntraID with a hieracy which is sort of automated by HR data, but not really, because sometimes HR puts everyone on the CEO level by mistake, and, if you trusted HR as authoritative that would've just broken all the EU laws. So you have all those Entra groups and you need IT Operations to maintain them and since you want to build it on job roles and not people you'll typically not be able to maintain them in the off the shelf system. Which means that you would have had to build a web portal for the plant managers manager where they could maintain a couple of Entra groups in a web interface. That or you'll have to setup an IT support flow where you add yet another system that IT has to maintain access for.
Then we get to the actual customization. Maybe you buy a custom API on top of your BC365 platform. Maybe your C-levels deciced that paying €50k a year to avoid outages on major updates isn't worth the cost. Then when things predictably and completely avoidable fail you're going to hav to deal with the literal shitstorm. You'd think that all the people being locked out of their jobs and the €150k cost of getting an immediate and prioritised update to the system would mean you'd start paying for that $50k service after this. You'd be wrong. Ok, to be fair, in this particular example it would be a different scenario. For a small system like this you'd find a cheap consultant house in your area and get them to build the customization for you. Only they would outsource it to some solo developer who will build it in a way that basically requires that specific person to alter it. Then when it breaks or needs to be customized futher a year down the line, that person is no longer a solo developer. So you reach out to another cheap consultant house and do it all over again, from scratch.
This doesn't even mention how poorly all those 300 off the shelf systems work together. I mean, I don't maintain a SDK delivering a way to use Apache-Arrow to write and read parquet files from our datalake in the same manner for fun. I do it because those 600 container apps which basically simply translate data from one system to another need it to be as slim as possible.
“I wanted to see if I can get in on this 10x magic. I decided to put my money where my mouth was and I bought $10 worth of DeepSeek credits to use with a project I was working on.“
I'm sure DeepSeek isn't the point here. You can change the name to whatever you prefer and the article still holds.
Actually, I think the author put DeepSeek on purpose to avoid the obvious ChatGPT/Claude comparison — because whatever he chose, there would be a question of why model A and not B, while the point of the article isn't about models comparison at all.
It is the point. Also open source model enthusiast tell you otherwise, there is a coding quality gap between these models. If I use DeepSeek, I do so knowing that I have to limit to simpler tasks on smaller, well specified prompts. What the author did, letting the model do the planning, is not something DeepSeek will excel at. I'm using GPT (Terra, Sol, Luna), Claude (Opus 5, Fable), Qwen 3.8 and GLM 5.3 Flash daily and have to vary which model I use where because there's a huge intelligence step function difference here. That's why this article is so useless:
Imagine someone trying to make the case that riding bicycles is a terrible experience and their whole argument is that they took a random cheapo bike with flat tires and rode it for 3min and that wasn't fun. Sure, but if you buy a 25k carbon bike you will have a different experience. I'd not trust that person. If someone told me they have 10 bikes they ride daily and can explain the differences, in detail, between their bikes, and what they excel at. I'd trust that person's opinion.
Excellent analogy. This paragraph invalidates the entire post and honestly just looks lazy. The author may be right anyway, but with that level of experience with these tools, he is really just guessing.
Right. Better throw money at 5 different ones and then people come and tell you that you just need MORE agents and throw MORE money at it or you're not doing it right.
> I'm also convinced the effect would not be there If I had a 10$ budget.
Sure it would be. I pay $10/month to OpenCode for a Go subscription, it's fine for day-to-day coding tasks. I wouldn't necessarily try and one-shot a production app on that budget, but with decent planning and test-driven-development, it gets the job done
The cheapest coding subscription is $20 and you can get a refund if you cancel within 7 days.
Actually, I lied, Codex is free and I developed my first AI written application using it and the free tier limits were generous enough to work on it for several months.
I got here and... I agreed with everything up to this point.
The author makes a good point. If you don't know what you're doing, AI accelerates that. No question.
But they put a whopping... ten bucks into using DeepSeek and weren't impressed with the initial results.
I know they try to cover this with "you just aren't prompting correctly!" but if, in 2026, you aren't able to have an LLM generate decent quality code... IDK what to tell you. Good luck I guess?
DeepSeek has been the most used model on open router by far, only temporarily overtaken by ox alpha when it was free, so it's the most representative experience
It is funny that only real research on productivity gains from AI shows at best very minimal gains, but AI bros will always tell you "no no no, you have used wrong model, try a different one, there are more of them, you have to try, trust me" and call that a "research".
Ah yes I'm an AI bro because I said the author must do thorough testing before coming to a conclusion about LLM's.
Looking forward to the authors next blog post on how after driving one car they find that all cars are slow and uncomfortable. Then the follow up after trying one phone and that all phones have bad cameras and battery life.
Bit of a humbling/jarring moment when I realized that people are doing real paid work using LLMs that they could not otherwise do. I mean, it's quite obvious I suppose. But up until now I just assumed it was only a (massive) catalyst for things people would already be able to do with enough time. But nope -- it seems people are right now employed in roles that they would not be able to fulfil the tasks within if AI wasn't there telling them what to write/say/produce. Nobody is really going to come out and say that ... it's not something the less-AI-literate superiors would take kindly to.
> people are doing real paid work using LLMs that they could not otherwise do
FWIW, this is not exactly new; those same people were just using other sources like Stack Overflow, blog posts, etc. before, cobbling together random code snippets, libraries, and so on without actually understanding any of that at a relevant detail level.
Sure, with LLMs, one can naturally tailor this much closer to the current need (or at least the need one thinks they have) and iterate ("spew") faster, but it's not a new phenomenon in general.
Yeh but I also mean outside of software engineering. Within the gamut of 'being a programmer' it seems fair to bleed into adjacent areas without too much cheek. We've all done it; it's part of the learning curve. But I was talking more about people in other knowledge work who have to come up with a lot of prose-like material about {insert thing}. Marketing, consultants, PMs, or even domain-specific analysts, .. ya know, the types of office jobs where people basically write emails, attend meetings, discuss reports, and produce mostly text-or-data artefacts all day long.
I think this is fine as long as the success criteria are strict enough that they are forced to learn something in the end.
Any significant testing will inevitably create that situation. The LLM won't have enough context to handle the more precise business requirements. The dev will have to read the code carefully and make their changes by hand. Additional rounds of testing may cause thrashing between regressed states until something clicks for the developer. That lightbulb going off is called "learning" and they are human after all!
I mentioned this then other day, my university class had maybe 80% of people who could not code. No joke. And a lot have gone onto to code professionally.
You can learn to code in a few months in your free time and get good in your job. That's different from refusing to learn it, because you outsource it.
Also coding is not that hard. You kinda only write the algorithms down you make up in your head. And you do have an algorithmic understanding when you graduate. You also invented some/several languages in during your studies. The learning an actual developed language IS learning, but it's just boring learning, nothing compared to what you did for your exams.
aka there are a lot more bullshit artists around these days
i know of several engineers who produce absolute slop and who probably would have produced nothing at all in pre AI times (which would have been preferable) and probably let go or never hired (even better).
they impose such an enormous drag on productivity that they more than wipe out any gains from people using the tools responsibly.
In general, the frontier models are not capable of reliably authoring non-trivial code without careful oversight yet. They are great at producing code that can pass tests, but not neccessarily a code review. This means if you care about code quality you still need a human in a loop understanding what has been done, and that becomes the bottleneck. And less disciplined folks will indeed become increasingly dependent.
However, over time the complexity of problems where you can get away with less/no oversight is increasing. And the models are already great at solving certain classes of problems where one doesn't really care that much about code quality, that wouldn't have even been attempted in a pre-LLM world. Over the weekend I was using Claude to add features to the compiled (no source available) firmware of one of my audio devices, adding workflow features by patching assembly and custom DSP code.
In coding, as with other areas, what's emerging is jagged intelligence.
Not at any decently sized company. Delivering features is all that matters. By the time you're cleaning up the buggy mess, they were promoted two times and work somewhere else now. If anything, AI makes it much easier to suck and still get things done to an acceptable degree.
Nope. That would require the people paying for the code (those requesting the projects) to see something wrong with the code. As a coder, you would look in an see a mess. As a user they'll probably see something relatively nice, with some maybe strange behavior and non ideal, but what software sin't non ideal? It does most of what we wanted, and you finished it how fast!?
'Or else admit this is a dopamine game that makes you feel like The Universe's Most Special Programmer™ when it's really just gamified mass-scale intellectual dependency.'
Alternatively, AI teaches you lessons about technical debt, code maintainability, and architecture faster too. Traditionally it took at least one or two years to really bump into those problems. Now any student can get into these problems within one or two weeks.
There are some good points, and I ask the question of where is the ground breaking stuff myself, but severely weakened by
* stretching the timeline: the actual real programming ability appeared in LLMs in the last 6-8 months, not 3-4 years,
* using the weakest possible tool: and I bought $10 worth of DeepSeek credits that is a far cry from Claude with Fable.
Also, I know nothing about marathons but for most uses putting the app, database, and background processes on the same server is very much the right starting point. With the next steps being employing Cloudflare or similar solutions long before managing a fleet of servers.
Even as someone using AI on the regular I'm starting to hate the "You didn't actually use this exact most expensive model so your point is invalid" argument.
This is fair to say if someones last experience with AI was copy-pasting code into GPT3 chat windows years ago, but Deepseek is a more than capabale model and enough for someone to get an informed opinion about the technology.
If people have actual counter argument, use those. And if some of those counter argument are "What you say isn't possible, the neweste model can do and here are examples of that", that is fine.
But a blanket "Nuh-uh, it wasn't Model X" is not only a poor argument but also automatically invalidates any criticism when a new, better model comes out - and that can't be the basis of a good argument.
In this case, the argument is valid though. Only recently have some of the models become powerful enough (for some) to actually be useful in day to day programming, without too much hand-holding. For most, this change occurred with the introduction of Claude Opus, and OpenAI and Google have caught up. Unfortunately, none of the "open" models is at this level yet.
The field is moving fast, and asking for scientific arguments is not realistic. It takes an extreme amount of effort to show what exactly is different.
We were in a similar position with static vs dynamic typing for decades. There is still no scientific proof that one is better than the other, but it is quite obvious to professionals which flavor works better in a given situation.
So, even though the argument might be sloppy, I subscribe to it. Using DeepSeek to dismiss better models is the bad argument here.
Edit: added "(for some)" as a disclaimer that you still need to be a fairly decent programmer to actually benefit.
> Even as someone using AI on the regular I'm starting to hate the "You didn't actually use this exact most expensive model so your point is invalid" argument.
What the author of the article is doing is dismissing a technology so disruptive that it's basically all everyone's talking about in the "tech space" at the moment (I mean look at HN frontpage for the past few months), by trying a relatively mediocre (but still quite good) model for about 10 seconds.
The reality is that frontier models suddenly got very good in the past 3-6 months. It has it's problems, and you need to learn how to use this new tool (as with any tool).
But models can and do generate good code. They also can and do generate absolute garbage (even Fable).
You need a good harness, tools to help the model check it's own output, good context, and a good idea of what you actually want. If you have those 4 things, the chances of generating absolute garbage are pretty slim (but yeah, still there).
The models got good starting with 2026 and that isn't some attempt at excusing it. Companies like OpenAI started building dedicated models around their coding harness called Codex, there was gpt-5.3-codex and it was both cheaper and better at using the harness than the regular models. Then they started merging the two model types into their main release models. All of this happened like 6 months ago.
You don't have to pay money to use Codex, there is a very generous free tier that costs you nothing, you just have to accept being told you're out of tokens every day. Because your token limits are low, you need to make sure that you accept or reject everything manually and when it tells you that it wants to run a command you have to paste in the command into your terminal and only paste the relevant output back otherwise it floods the context window.
I remember being told in early 2025 that only now has "real programming ability" appeared in LLMs.
And since then there's been 3 more "now real programming ability has been made available and previous stuff was just toy examples" cycles (summer 2025, winter 2025 and spring 2026)
Looking forward to the next "everything before this was trivial and bad, here's the good stuff" moment
If you haven't been using the latest NanoQwen27B-5.2.10-Turbo-KimiGLM-Pro-10.4 in the past 13 hours, your information on what LLMs can do is outdated and you're falling behind. Check your priors.
Yeah, it's interesting how the current thing is always the end of programming, but then when a new thing comes it turns into absolute trash that can't be used for anything.
> “after four years of open source LLMs, we should have three AirBnBs, two Stripes, and three Dropboxes thanks to the power of AI.”
Airbnb, Stripe and Dropbox were created in a different time when the market was much less competitive.
Saturation of software development velocity doesn’t increase large scale product opportunities in the market. It can also mean that opportunities get filled even more quickly by niche players, and nobody gets to grow to Airbnb scale.
IMO the latter is what’s currently happening. AI-powered companies are like little mammals scurrying around between the feet of the dinosaurs, and commentators like the OP look at the evolution of the brontosaurus as evidence that the mammals don’t seem to be growing as they should.
That's one of the few consumer-facing areas where there are still standards in place, namely PCI-DSS. As far as I know the audits require the name of a human who is responsible for payment security. Card companies can one-hit kill your startup if you're breaking those rules (maybe purely blockchain startups are exempt).
Yes, you can offload this to stripe, but then your app should never see the card number and certainly not the CVV. You end up storing these, even by accident, both stripe and the card companies will hate you.
I think my 10x comes elsewhere than writing code. For ex llms are much better at bash than me. So good I don't have to leave my shell. Jira? ServiceNow? Ha! Claude you go update that ticket. Go check my history and git commits and write me my standup. I'm not leaving the terminal.
I don't disagree with the title of this weblog post, but its premise is wrong: the existence of AI has an impact on the total number of tech companies, but not on the number of companies that make it real big (the AirBnBs, Stripes and Dropboxes the author is looking for).
AI does make me suck more, which is a good thing. It sped me up and has me iterating on things in 3 days that normally would have taken me more than a month to do. Surprise, surprise: I'm discovering that I have a lot of blind spots and false presumptions about what good programming is. My understanding of the craft is patchy, but that wasn't as noticable when going at the speed of human-programming.
It's like cars. If you only ever drive at 30 km/h, a 30 yo car, that had its last oil change in 2008, might be great. But, if, in the same car, you start driving at 200 km/h, you might notice that there's a lot of room for improvement. My programmer-self is the 30 yo car in this metaphor.
I tried really hard to get senior management to understand this where I work. Yes, AI is an accelerator, but that doesn't necessarily mean it's going to accelerate you in the "right" direction.
I have a coffee cup with the writing "Do stupid thing faster with cafe". That's how I feel about myself when I use AI carelessly... The speed with which I can make a mess is astronomical!
> Borrowing Disesdi's logic in their arithmetic (which is hard to argue against), after four years of open source LLMs, we should have three AirBnBs,1 two Stripes,2 and three Dropboxes3 thanks to the power of AI.
>
> So. Where the fuck are they?
I think this is a great question, and I have a completely unqualified theory.
I work in a fairly specialised field, I've notice when I use AI to try to generate code in fields where I'm very experienced, it produces a worse output at a slower rate than I can produce manually[0].
When I use it in fields I'm not an expert in by any stretch, like webdev, genAI massively improves my output by a huge factor.
I think this makes the productivity gains a little fuzzy. Claude code makes me 10x faster at webdev, but I'd be a very slow webdev. Ultimately, the next innovation in a field, will probably be from an expert in that field, and they therefore won't see the kind of "10x productivity boost" that non-experts do.
[0] I still find AI helpful for exploring code etc, just not so much producing output.
- when are all the "I am running agents 24x7 a day on max plan" people on HN gonna learn this?
- best case scenario bro: LLMs get infinitely better and nobody needs to code anymore
"software requirement prompting is the new skill" I can write those pretty well bro
- worst case scenario: AI market crash, cognitive debt spikes across every major organization filled with vibe coding juniors that have never spent a single day in their life debugging a production setup without AI.
- Now the whole world is filled with 90% programmers that cannot add 2 numbers in c++ without using GPT
- Guess what bro? I am now one of the most valuable programmers there is :)
How you use AI as a person who can actually code is up to you.
You can use AI in two ways, either…
1. set the plane on autopilot and arrive at the destination having no idea how you got there - like any old idiot who can type a prompt. And also pay a very expensive air fare for this super power auto pilot.
2. You can pilot the plane yourself. Plan the route in advance. Set waypoints. Make decisions along the way. And land the plane manually. And the reward is you know the route and you pay 1/10 the cost.
At least once a week I have to drop what I'm doing and race in to stop the non dev manager from merging a fully vibe coded patch which will blow up production. Claude Code has made this guy a negative 10x developer.
They don't have permissions but if I don't rush to block the PR with a bunch of review comments, one of the other AI brained devs will just hit approve on it without reading or understanding anything.
Yes it is hell here. But it's a new kind of problem that just showed up.
At worst, needs another manager to approve. Now they can both saddle up on the velocity/time to market wave... together. The racket begins in earnest.
How does a manager do this, they ask? I answer: the same way they do everything. Lunch. If the effort fails, at least they tried! Deck chairs, Titanic, etc.
If one truly believes it's not in a bubble, then borrow and leverage an unlimited amount and bet the house & your family on creating a business with infinite growth and value. Don't know how to do your own prompts? Borrow and then hire others to do it for you. If those who you hire don't know, they can hire others too.*
[* This is not financial advice. Please don't actually do this.]
You could make the converse argument, too: why don't you borrow and leverage and bet and short these AI companies, if you so vehemently think it's a bubble?
The bottom line is that markets are complicated, useful technologies are often accompanied by bubbles, investors are not always rational, and people generally try to make the best decision with the information they have available. The answer is likely somewhere in the middle, but that's a lot more boring and a lot less inflammatory.
> You could make the converse argument, too: why don't you borrow and leverage and bet and short these AI companies, if you so vehemently think it's a bubble?
Because going long and going short are very different. For successfully shorting something you need to have a pretty precise estimate of when the crash will happen. Being a comparatively short time off can cost you everything.
You mean for companies that make -$2.50 for every $1 of revenue, and bury their groups $2Tn debt in a footnote?
It is not a question of if the bubble will go, but when... but you are right that Bears or Bulls always get it wrong predicting the future (if they are a legal investor.)
Personally, Shrek movie release correlation with market corrections is funny, and a new film is due June 2027. Please hedge your bets with a diversified portfolio. =3
AI is like a tool used by mega-corporations to change the world according to what they want. We see this with the increase in RAM prices. I don't want to pay the overprice here - AI companies owe us money. People seem still mostly in the AI hype phase, but a lot more criticism has amplified in the last months. It is only a matter of time until the hype phase is over.
> The chatbot recommended some of the dumbest shit you could possibly do
The "I tried it and it sucked" is borderline conspiratorial at this point. There are enough talented, thoughtful developers saying there's something real here, and it is worth believing them and investing some time to understand it, even if you come out the other side and decide you don't want to use LLMs.
Use a very good model. Set up a good harness. Spend some time on your system prompts and skills. Develop your intuitions about how the model works, what it's good at, how to scope the work, and how to steer it. Recognize when it's alleviating menial work and recognize when it's making choices you really need to understand yourself. Be patient and accept that the failures are going to be very painful for a while.
Don't write it off until you've genuinely seen the upsides.
You describe my experience with Claude (except my gut feeling is that system prompts and especially skills are useless, and I've even done limited testing that agrees with that).
However, your discourse is not the usual borderline religious discourse that's used by LLM advocates. You're not extreme enough either pro or against LLMs :)
Ha, well that's good because I would hate to be labeled an "LLM advocate". I've been on the negative side of the conversation many times, and I think the downsides are very real and often ignored. But calling them useless is just not reality.
As far as system prompts and skills, my approach has been to start with nothing and add very specific instructions as I notice issues or feel like I need them. Mostly around trying to keep it terse, kill the obnoxious rhetoric ("It's not X. It's Y"), and give specific guidance on how to write code. Skills are things like "use edge-tts to generate spoken audio for this answer and send it to me" kind of thing.
> except my gut feeling is that system prompts and especially skills are useless
I tend to agree. Not with the extreme version of this statement: there are some genuinely useful instructions I give in my Claude.md and prompts. Comment style, how much to push back, which subagents to orchestrate for tests, reviews, etc.
But I've tried the single-sentence versions of those and versions with multiple files of long prompts for the orchestrator and its subagents, and I can't tell the difference in output quality. If anything, the shorter version is better
> Don't write it off until you've genuinely seen the upsides.
This is ridiculous. Basically you're saying "if you try it and find it sucks, you're wrong, keep at it until you change your mind". That isn't a tool at that point, it's a religion.
No, they're saying there are upsides, even if you think they are small, or think they don't outweigh the downsides, you at least have to be honest about where the utility is.
I largely agree with this piece but the author isn't really giving the AI development process a fair shot. It's not a matter of "prompt better" but of establishing processes that account for and correct agent shortcomings.
For me, it's taken thousands of dollars and months of experiments to get a process that produces working and maintainable code. You can't just go in with $10 and be like "Wow it doesn't work" after playing around with it for 30 seconds, which is how far $10 takes you on models that are doing work. All that does is expose the shortcomings that need correction.
You can get around these not by just "prompting better", but by establishing rules and pipelines that are enforced with standard software engineering tools like compilers, linters, type checkers, borrow checkers, profilers, benchmarking, smoke tests, proof assistants, and so forth. Those things we know work really well, so AI coding shouldn't be viewed as AI doing all the work, but AI orchestrating all of the tools that we know work so well in software development.
The magic of agentic engineering is only possible because our software tools are so good not because agents are.
Why does it have to lead to a 10x increase in revolutionary companies?
Next to maintaining and expanding my own network of websites I 10xd writing boring CRUD applications for companies that were otherwise unable to afford it, making all their employees more productive. There's true economic value in that.
This is fine, but software also needs maintenance, especially security updates. If we have an explosion of new software without the corresponding increase in maintenance work - very likely from clients that pay peanuts for software - the world will be taken over by botnets and malware.
I just allow access to their application only for their IP address(es) and my deployment server checks if I need to update anything. That's not expensive.
His AI experience is $10 of DeepSeek - but I think he is right anyway in the main points.
You can most definitely ship crap much faster than you used to. It's obvious to anyone when you're shipping crap. And it seems everywhere I look people are shipping crap. Both software and writing.
The hardest part of shipping quality software is not and has never been "writing the code". The hard parts are product taste, architecture and ensuring your product actually solves the problems it should in an efficient and secure manner.
If you don't have an intuition of those, you will likely still be shipping crap. Just faster.
Yep, using AI to remove your thinking process and judgement, and just let it generate stuff as fast as possible, will definitely make you ship crap faster! If you did the same thing with a junior developer you would get the same result.
Where AI shines for me is accelerating the learning and exploration process. I can get up to speed with new tech fast. It is good at spotting issues in designs and code. It can knock out quick tests or benchmarks to support me. The quality of what I can produce with AI support is much higher than I could without it.
So it really just depends on what you use it for. If the goal is "replace humans and ship fast" that's one thing. If the goal is "explore the problem space in greater depth", it's another.
> and ruining the quality of internet search engines.
Tbf, I'm not sure we can really lay this one at the feet of GenAI. The SEO bros had pretty thoroughly ruined search before LLMs took off - the process just accelerated a little at that point.
I also changed the way I do review because it has been more than a year and the juniors/new hire are still lost, wether on domain knowledge for the older new hire, or just capabilities for the juniors, and discussing with other departments, it's the same for like 95% of them. Now, rather than correcting the PR or adding a request for change, I add a whole unit/functional test to the PR and let that as an exercise to pass the test. They can use AI but I tell them to try to find what part of the code doesn't work before generating the fix, hopefully they'll take ownership of the code if I keep doing that.
this makes me kinda sad. they don't do this on their own? are they not even a little curious
For me and you, "fixing" something means finding which assumption was violated and redesigning the solution in light of that. Sometimes this makes the solution smaller.
For a lot of people, "fixing" means "adding code to make it work". This always makes the solution larger.
So, the models ARE dumb. They just are very good at finding patterns in their training data. I mean, when they code minecraft clones, it is not because they can cook up how to write minecraft clones, rather, their training data includes a lot of minecraft-like games code, and they just reuse that.
Don't get me wrong. It's not great. It would never pass any of our policies for things that actually operate stuff on the power grid, but as an administrative tool that can live in total isolation from the vital networks. It's perfectly fine. It's also not like we would have hired the best software companies to build it otherwise. We'd hire some low-level cheap consultant house who would then likely get cheap student labour to build it. With that in mind though, the AI is much better than what the realistic alternative would be.
Money wise it's also cheaper. It's been roughly €1000 + the time it's taken us both. If I had known they were doing it, I would have rolled out the developer cowork app/skills/whateveryoucallconfigurationsthesedays to them. This would have avoided their AI building it to be depoyed on a VM rather than in our managed k8s in our Azure. It would also have written the code a little different, used UV and maybe django rather than flask. But hey. For what it is, it's like a 90% cost saving compared to buying what would've been a less maintainable and lower quality system.
I think perhaps the greater issue will be finding people who want to extract the gold from the heap of shit and getting it to run in production. I don't personally mind, but it's not like any of my colleagues would've wanted the task.
The insecurity in a vibe-coded web portal isn't that someone hacks it with XSS, it's that after the next vibe-coded release, some X quietly becomes −Y somewhere no one expects.
From this perspective, having no software at all might be better, or as in your case, safer.
As you point out this portal isn't that, but what protects us is the processes around compliance. This can't grow from X to Y because not even the CEO has the authority to overwrite our compliance gates. The EU is a tremendous help in this area since personal liability changed things completely.
> We're in the European economic area and are completely NIS2 compliant and we use LLM's to aid in our software development for high risk systems.
Please pick one. Either way this is a nightmare level of threat to sovereignty.
1. https://news.ycombinator.com/item?id=47063153
I still doubt we will ever give an AI access to run code on our systems directly though. In isolation, sure, but other than that.
This is a side note, but my personal favorite part of Cowork is that I can roll out our compliance policy to every developer as a Microsoft Teams app (no, that makes no sense to me either). So when they try to install some package that isn't pre-approved their Cowork agent won't let them and will instead explain how they might get approval. If they then continue to reference it, Cowork will even alert us.
Well, I'm speechless.
I agree, the EU can't become sovereign or have privacy, if the citizens constantly work against it.
From my perspective it looks like were just allowing hostile developers within our environments now lol.
There's not a lot of software where users dont really care if it goes wrong.
From an enterprise perspective this becomes complicated for various reasons. RBAC is one area. In the perfect world you have a system to handle roles and rights to every system, something that you can give managers access to so they can maintain the access available to their employees, something linked with HR. In reality you have EntraID with a hieracy which is sort of automated by HR data, but not really, because sometimes HR puts everyone on the CEO level by mistake, and, if you trusted HR as authoritative that would've just broken all the EU laws. So you have all those Entra groups and you need IT Operations to maintain them and since you want to build it on job roles and not people you'll typically not be able to maintain them in the off the shelf system. Which means that you would have had to build a web portal for the plant managers manager where they could maintain a couple of Entra groups in a web interface. That or you'll have to setup an IT support flow where you add yet another system that IT has to maintain access for.
Then we get to the actual customization. Maybe you buy a custom API on top of your BC365 platform. Maybe your C-levels deciced that paying €50k a year to avoid outages on major updates isn't worth the cost. Then when things predictably and completely avoidable fail you're going to hav to deal with the literal shitstorm. You'd think that all the people being locked out of their jobs and the €150k cost of getting an immediate and prioritised update to the system would mean you'd start paying for that $50k service after this. You'd be wrong. Ok, to be fair, in this particular example it would be a different scenario. For a small system like this you'd find a cheap consultant house in your area and get them to build the customization for you. Only they would outsource it to some solo developer who will build it in a way that basically requires that specific person to alter it. Then when it breaks or needs to be customized futher a year down the line, that person is no longer a solo developer. So you reach out to another cheap consultant house and do it all over again, from scratch.
This doesn't even mention how poorly all those 300 off the shelf systems work together. I mean, I don't maintain a SDK delivering a way to use Apache-Arrow to write and read parquet files from our datalake in the same manner for fun. I do it because those 600 container apps which basically simply translate data from one system to another need it to be as slim as possible.
Am I jaded? Sure. But who isn't in enterprise IT?
Wow. I guess that’s the punchline!
Actually, I think the author put DeepSeek on purpose to avoid the obvious ChatGPT/Claude comparison — because whatever he chose, there would be a question of why model A and not B, while the point of the article isn't about models comparison at all.
Imagine someone trying to make the case that riding bicycles is a terrible experience and their whole argument is that they took a random cheapo bike with flat tires and rode it for 3min and that wasn't fun. Sure, but if you buy a 25k carbon bike you will have a different experience. I'd not trust that person. If someone told me they have 10 bikes they ride daily and can explain the differences, in detail, between their bikes, and what they excel at. I'd trust that person's opinion.
I'm also convinced the effect would not be there If I had a 10$ budget.
Sure it would be. I pay $10/month to OpenCode for a Go subscription, it's fine for day-to-day coding tasks. I wouldn't necessarily try and one-shot a production app on that budget, but with decent planning and test-driven-development, it gets the job done
Actually, I lied, Codex is free and I developed my first AI written application using it and the free tier limits were generous enough to work on it for several months.
The author makes a good point. If you don't know what you're doing, AI accelerates that. No question.
But they put a whopping... ten bucks into using DeepSeek and weren't impressed with the initial results.
I know they try to cover this with "you just aren't prompting correctly!" but if, in 2026, you aren't able to have an LLM generate decent quality code... IDK what to tell you. Good luck I guess?
Why not use actual frontier models, and you know do some real research, before writing a blog post?
My point is, there is a monumental difference between flying Ryanair and Emirates First Class for example.
It is funny that only real research on productivity gains from AI shows at best very minimal gains, but AI bros will always tell you "no no no, you have used wrong model, try a different one, there are more of them, you have to try, trust me" and call that a "research".
Looking forward to the authors next blog post on how after driving one car they find that all cars are slow and uncomfortable. Then the follow up after trying one phone and that all phones have bad cameras and battery life.
Sure, with LLMs, one can naturally tailor this much closer to the current need (or at least the need one thinks they have) and iterate ("spew") faster, but it's not a new phenomenon in general.
If the company want's hire someone who doesn't have a clue and only uses SO, that's of course fine, but I doubt, that this is the case.
Any significant testing will inevitably create that situation. The LLM won't have enough context to handle the more precise business requirements. The dev will have to read the code carefully and make their changes by hand. Additional rounds of testing may cause thrashing between regressed states until something clicks for the developer. That lightbulb going off is called "learning" and they are human after all!
AI is a godsend to this cohort of code monkies.
Edit: we graduated in 2006.
Also coding is not that hard. You kinda only write the algorithms down you make up in your head. And you do have an algorithmic understanding when you graduate. You also invented some/several languages in during your studies. The learning an actual developed language IS learning, but it's just boring learning, nothing compared to what you did for your exams.
i know of several engineers who produce absolute slop and who probably would have produced nothing at all in pre AI times (which would have been preferable) and probably let go or never hired (even better).
they impose such an enormous drag on productivity that they more than wipe out any gains from people using the tools responsibly.
However, over time the complexity of problems where you can get away with less/no oversight is increasing. And the models are already great at solving certain classes of problems where one doesn't really care that much about code quality, that wouldn't have even been attempted in a pre-LLM world. Over the weekend I was using Claude to add features to the compiled (no source available) firmware of one of my audio devices, adding workflow features by patching assembly and custom DSP code.
In coding, as with other areas, what's emerging is jagged intelligence.
Now with AI I can 10x my output and 10x my bad code!
Disesdi Shoshana Cox
* stretching the timeline: the actual real programming ability appeared in LLMs in the last 6-8 months, not 3-4 years,
* using the weakest possible tool: and I bought $10 worth of DeepSeek credits that is a far cry from Claude with Fable.
Also, I know nothing about marathons but for most uses putting the app, database, and background processes on the same server is very much the right starting point. With the next steps being employing Cloudflare or similar solutions long before managing a fleet of servers.
This is fair to say if someones last experience with AI was copy-pasting code into GPT3 chat windows years ago, but Deepseek is a more than capabale model and enough for someone to get an informed opinion about the technology.
If people have actual counter argument, use those. And if some of those counter argument are "What you say isn't possible, the neweste model can do and here are examples of that", that is fine.
But a blanket "Nuh-uh, it wasn't Model X" is not only a poor argument but also automatically invalidates any criticism when a new, better model comes out - and that can't be the basis of a good argument.
The field is moving fast, and asking for scientific arguments is not realistic. It takes an extreme amount of effort to show what exactly is different.
We were in a similar position with static vs dynamic typing for decades. There is still no scientific proof that one is better than the other, but it is quite obvious to professionals which flavor works better in a given situation.
So, even though the argument might be sloppy, I subscribe to it. Using DeepSeek to dismiss better models is the bad argument here.
Edit: added "(for some)" as a disclaimer that you still need to be a fairly decent programmer to actually benefit.
What the author of the article is doing is dismissing a technology so disruptive that it's basically all everyone's talking about in the "tech space" at the moment (I mean look at HN frontpage for the past few months), by trying a relatively mediocre (but still quite good) model for about 10 seconds.
The reality is that frontier models suddenly got very good in the past 3-6 months. It has it's problems, and you need to learn how to use this new tool (as with any tool).
But models can and do generate good code. They also can and do generate absolute garbage (even Fable).
You need a good harness, tools to help the model check it's own output, good context, and a good idea of what you actually want. If you have those 4 things, the chances of generating absolute garbage are pretty slim (but yeah, still there).
You don't have to pay money to use Codex, there is a very generous free tier that costs you nothing, you just have to accept being told you're out of tokens every day. Because your token limits are low, you need to make sure that you accept or reject everything manually and when it tells you that it wants to run a command you have to paste in the command into your terminal and only paste the relevant output back otherwise it floods the context window.
And since then there's been 3 more "now real programming ability has been made available and previous stuff was just toy examples" cycles (summer 2025, winter 2025 and spring 2026)
Looking forward to the next "everything before this was trivial and bad, here's the good stuff" moment
Or B that the ideas are there but are not get released as the code is ai slop?
Airbnb, Stripe and Dropbox were created in a different time when the market was much less competitive.
Saturation of software development velocity doesn’t increase large scale product opportunities in the market. It can also mean that opportunities get filled even more quickly by niche players, and nobody gets to grow to Airbnb scale.
IMO the latter is what’s currently happening. AI-powered companies are like little mammals scurrying around between the feet of the dinosaurs, and commentators like the OP look at the evolution of the brontosaurus as evidence that the mammals don’t seem to be growing as they should.
You won’t get capital let alone VC if you’re not AI.
It’s infected everything much like crypto did just 3-4 years ago.
That's one of the few consumer-facing areas where there are still standards in place, namely PCI-DSS. As far as I know the audits require the name of a human who is responsible for payment security. Card companies can one-hit kill your startup if you're breaking those rules (maybe purely blockchain startups are exempt).
Yes, you can offload this to stripe, but then your app should never see the card number and certainly not the CVV. You end up storing these, even by accident, both stripe and the card companies will hate you.
Sshh, don't say that part to loud, especially on HN.
Neither was ever about some bottleneck on pumping out code and everything about marketing and network effects
It's like cars. If you only ever drive at 30 km/h, a 30 yo car, that had its last oil change in 2008, might be great. But, if, in the same car, you start driving at 200 km/h, you might notice that there's a lot of room for improvement. My programmer-self is the 30 yo car in this metaphor.
I think this is a great question, and I have a completely unqualified theory.
I work in a fairly specialised field, I've notice when I use AI to try to generate code in fields where I'm very experienced, it produces a worse output at a slower rate than I can produce manually[0].
When I use it in fields I'm not an expert in by any stretch, like webdev, genAI massively improves my output by a huge factor.
I think this makes the productivity gains a little fuzzy. Claude code makes me 10x faster at webdev, but I'd be a very slow webdev. Ultimately, the next innovation in a field, will probably be from an expert in that field, and they therefore won't see the kind of "10x productivity boost" that non-experts do.
[0] I still find AI helpful for exploring code etc, just not so much producing output.
- when are all the "I am running agents 24x7 a day on max plan" people on HN gonna learn this?
- best case scenario bro: LLMs get infinitely better and nobody needs to code anymore
"software requirement prompting is the new skill" I can write those pretty well bro
- worst case scenario: AI market crash, cognitive debt spikes across every major organization filled with vibe coding juniors that have never spent a single day in their life debugging a production setup without AI.
- Now the whole world is filled with 90% programmers that cannot add 2 numbers in c++ without using GPT
- Guess what bro? I am now one of the most valuable programmers there is :)
You can use AI in two ways, either…
1. set the plane on autopilot and arrive at the destination having no idea how you got there - like any old idiot who can type a prompt. And also pay a very expensive air fare for this super power auto pilot.
2. You can pilot the plane yourself. Plan the route in advance. Set waypoints. Make decisions along the way. And land the plane manually. And the reward is you know the route and you pay 1/10 the cost.
Seriously, that was the lowest level of skill I saw in 30 years of development.
Now we have people who can't even do that proudly showing off PRs to widely used products.
"B-b-but I do the systems design and hard thinking".
Sure, buddy.
Yes it is hell here. But it's a new kind of problem that just showed up.
How does a manager do this, they ask? I answer: the same way they do everything. Lunch. If the effort fails, at least they tried! Deck chairs, Titanic, etc.
[* This is not financial advice. Please don't actually do this.]
The bottom line is that markets are complicated, useful technologies are often accompanied by bubbles, investors are not always rational, and people generally try to make the best decision with the information they have available. The answer is likely somewhere in the middle, but that's a lot more boring and a lot less inflammatory.
Because going long and going short are very different. For successfully shorting something you need to have a pretty precise estimate of when the crash will happen. Being a comparatively short time off can cost you everything.
Most look forward to picking up the assets at a heavy discount. =3
https://www.youtube.com/watch?v=rE75WvOtcu8
It is not a question of if the bubble will go, but when... but you are right that Bears or Bulls always get it wrong predicting the future (if they are a legal investor.)
https://www.youtube.com/watch?v=wTiYaWFP59Q
Personally, Shrek movie release correlation with market corrections is funny, and a new film is due June 2027. Please hedge your bets with a diversified portfolio. =3
The "I tried it and it sucked" is borderline conspiratorial at this point. There are enough talented, thoughtful developers saying there's something real here, and it is worth believing them and investing some time to understand it, even if you come out the other side and decide you don't want to use LLMs.
Use a very good model. Set up a good harness. Spend some time on your system prompts and skills. Develop your intuitions about how the model works, what it's good at, how to scope the work, and how to steer it. Recognize when it's alleviating menial work and recognize when it's making choices you really need to understand yourself. Be patient and accept that the failures are going to be very painful for a while.
Don't write it off until you've genuinely seen the upsides.
However, your discourse is not the usual borderline religious discourse that's used by LLM advocates. You're not extreme enough either pro or against LLMs :)
As far as system prompts and skills, my approach has been to start with nothing and add very specific instructions as I notice issues or feel like I need them. Mostly around trying to keep it terse, kill the obnoxious rhetoric ("It's not X. It's Y"), and give specific guidance on how to write code. Skills are things like "use edge-tts to generate spoken audio for this answer and send it to me" kind of thing.
I tend to agree. Not with the extreme version of this statement: there are some genuinely useful instructions I give in my Claude.md and prompts. Comment style, how much to push back, which subagents to orchestrate for tests, reviews, etc.
But I've tried the single-sentence versions of those and versions with multiple files of long prompts for the orchestrator and its subagents, and I can't tell the difference in output quality. If anything, the shorter version is better
Are those "skills" or just instructions?
My definition for "skills" is the "you are the greatest software architect ever born" type bullshit.
This is ridiculous. Basically you're saying "if you try it and find it sucks, you're wrong, keep at it until you change your mind". That isn't a tool at that point, it's a religion.
For me, it's taken thousands of dollars and months of experiments to get a process that produces working and maintainable code. You can't just go in with $10 and be like "Wow it doesn't work" after playing around with it for 30 seconds, which is how far $10 takes you on models that are doing work. All that does is expose the shortcomings that need correction.
You can get around these not by just "prompting better", but by establishing rules and pipelines that are enforced with standard software engineering tools like compilers, linters, type checkers, borrow checkers, profilers, benchmarking, smoke tests, proof assistants, and so forth. Those things we know work really well, so AI coding shouldn't be viewed as AI doing all the work, but AI orchestrating all of the tools that we know work so well in software development.
The magic of agentic engineering is only possible because our software tools are so good not because agents are.
Next to maintaining and expanding my own network of websites I 10xd writing boring CRUD applications for companies that were otherwise unable to afford it, making all their employees more productive. There's true economic value in that.
It's going to be a worldwide software experiment.
You can most definitely ship crap much faster than you used to. It's obvious to anyone when you're shipping crap. And it seems everywhere I look people are shipping crap. Both software and writing.
The hardest part of shipping quality software is not and has never been "writing the code". The hard parts are product taste, architecture and ensuring your product actually solves the problems it should in an efficient and secure manner.
If you don't have an intuition of those, you will likely still be shipping crap. Just faster.
Where AI shines for me is accelerating the learning and exploration process. I can get up to speed with new tech fast. It is good at spotting issues in designs and code. It can knock out quick tests or benchmarks to support me. The quality of what I can produce with AI support is much higher than I could without it.
So it really just depends on what you use it for. If the goal is "replace humans and ship fast" that's one thing. If the goal is "explore the problem space in greater depth", it's another.
you want more SaaS?
I don’t want to solve a new problem with an LLM programming assistant anyways - that’s the fun part of my work, why would I get rid of it?
Tbf, I'm not sure we can really lay this one at the feet of GenAI. The SEO bros had pretty thoroughly ruined search before LLMs took off - the process just accelerated a little at that point.