DEV Community

Cover image for Is this Software Development now?
Alex Hyett
Alex Hyett

Posted on Originally published at alexhyett.com on

Is this Software Development now?

I must admit I used to be an AI sceptic.

I used AI a lot when Claude Code first came out, and it struggled with even simple bash scripts introducing a lot of mistakes that I would have to check and correct. There was so much back and forth just to get something working that it was questionable how anyone was saving time with this. Now with the latest models, it is truly impressive what can be achieved.

YouTube seems to be full at the moment with people vibe coding whole games. I have seen copies of Splatoon, Mario Kart, Minecraft, GTA, Spiderman and even Godzilla.

Someone even made a lawn mowing game.

As you probably saw on the news a few weeks ago someone used AI to clone the whole Adobe and Microsoft Office suite and put it up for free on GitHub.

Yes undoubtedly this code is riddled with bugs and is nowhere near production ready, but it is usable and completely free.

If before AI, someone came to you and said they made any of these projects by themselves, over the course of a weekend, you would definitely be impressed.

Anyone with an idea and a Claude subscription can now build whatever they need in a short amount of time.

Nearly all the tech YouTubers are saying the same thing. These models are so good now that you hardly even need to look at the code.

Tech YouTubers that used to go on about how rubbish AI was, have now done a complete 360, and have started talking about their AI pipelines and sharing their skill files.

Even the OG developers like Linus Torvolds or Robert C. Martin, aka Uncle Bob, are AI promoters now.

I watched a few interviews recently with Uncle Bob, and he is using static analysis tools such as cyclomatic complexity to wrangle his agents into writing maintainable code.

DHH went on stage at Ruby World and announced to a whole room full of developers that, “Writing code by hand is no longer an economically productive enterprise”.

And it's easy to see why.

For $200 / month, Claude Code will generate more code than a whole team of developers could write even a year ago, and most of the time it will work fine.

I have seen so many developers on social media say they haven't written a line of code for months or even over a year at this point.

I don't think there is any doubt now that these models are very capable at writing working code.

If your job involves producing software to make money, then you have to ask yourself, what is the point of spending hours writing code by hand, when you can achieve working code in minutes, by using frontier models?

How does all this work? #

Training Data #

These large language models can clearly produce some impressive stuff, which would have taken months before. If you are using AI every day and plan to completely pivot your career around it, it is important to understand how it works.

Is this all magic, are the machines sentient and have developed some sort of “super intelligence”? How have these companies managed to achieve all this?

Well, the first thing these companies did was to illegally download millions of pirated digital books, as well scrapping the whole internet including all the publicly accessible code, while completely ignoring any of the licence agreements.

Some code used is GPL licensed, that requires all distributed derivatives to also be open source and use the same licence, as well as attributing the original source code.

Other codebases may have the Apache licence that also requires attribution. Of course no one is crediting the original source code because they can't tell where the AI got it from.

Images generated from OpenAI have started forging comic artists signatures though, so that's the same thing right?!

It is not just code and text that has been illegally used but also images from artists, songs, videos, literally all of human creation has been pumped into this thing.

In a leaked document, one of the Microsoft execs said it's the “largest theft of labor in human history”.

The problem is, not all of this human labour such as books, have digital versions available, especially if they are a bit older. So AI companies have started buying secondhand books en-mass, and had them shipped to their warehouses to be scanned in.

Of course, it is entirely possible to non-destructively scan in a book, but then that would mean it could end up in the hands of those pesky humans, who wouldn't need to pay for AI subscriptions any more.

So naturally, they cut off the spine of the book so they can scan in all the pages quickly and then destroy the remains. There weren't many copies of these old books left so that solves the pesky human problem.

But what about all the code that isn't publicly accessible. The code that makes up the software that people are actually willing to pay for. How on Earth were the AI companies going to get their hands on that?

Well, step one was to convince all the CEOs and upper management that AI is inevitable and if your company doesn't start using it, you are going to be left behind.

This then caused AI usage to be mandated on to developers. The subscriptions are so heavily subsidised as well that it just doesn't make economic sense not to use them. Especially when you consider your competitors are likely using them as well, and will be able to overtake you.

With all the developers now trying their hardest to keep their jobs, they spend their days feeding the AI companies all the employers proprietary code and helping to train the AI when it makes mistakes.

Of course the enterprise plans come with a pinky promise not to use your data to train the AI models. You have to trust these companies, after all, when have they ever trained their models on stolen data.

Running the models #

So what does it take to run these AI models. Can everyone just download them and run them on their computers?

Well yes, there are “open weight” models but most of them still require a pretty beefy setup. The smaller models that people can run on the computers at home, are nowhere near as capable as these frontiers models.

To cope with all the demand for the highly subsidised plagiarism machine, sorry AI, they have had to build data centres, a lot of data centres. These data centres are kitted out with the worlds supply of graphics cards and RAM for the next few years, which has caused the price of consumer hardware to skyrocket.

Which not only angers computer enthusiasts and gamers like myself, but also has the added benefit of stopping most people from running these free open weight models themselves.

There isn't an awful lot of concrete data around how much water and electricity these data centres use. Mainly because the companies have decided that these details are “trade secrets” and don't need to be shared with the public. How convenient!

Water usage is definitely up for debate, all we know is it is somewhere between that of a leisure centre and a small town.

Electricity again depends on the size of the data centre and ranges between 20 megawatts and 17 gigawatts. Either way, electricity usage is incredibly high and most of this energy is likely to come from non-renewable sources like gas power stations.

From the perspective of people in the tech industry, it seems like everyone is using AI. In reality, less than 1% of the worlds' population is using AI extensively for things like coding and mathematics, with maybe 20% using it like a chatbot for search and other small tasks.

The fact that in the UK alone, they are looking to build over 100 data centres over the next few years, I can only dread to think what will be needed if everyone starts using it.

The chances of looking out of your window and seeing a data centre is getting more and more likely, and is a reality for many already.

This is an incredibly inefficient technology and is often being used for simple tasks that could be done more efficiently by a human.

Why is it being used so much #

Despite all the negatives, AI use just keeps growing. This is partly because, like everything that comes out of Big Tech, it's addictive.

I am not sure if it is by design, or if it is just the probabilistic nature of the technology, but the dopamine hit from using AI, is similar to using a slot machine or scrolling on TikTok. You don't know whether your next prompt is going to be a one shot wonder or require another round of prompting.

This combined with the timed usage reset windows, creates a sense of FOMO, where you don't want to feel like you are wasting the credits that you are paying for each month.

It should be no surprise then, that despite the claims that AI is going to save us time, we all seem to be working longer and longer hours.

There are of course some positives to AI seen in other fields.

OpenAI recently managed to solve hundreds of problems in mathematics which mathematicians haven't been able to solve before. The proofs are technically correct, but many are nearly incomprehensible.

It is the equivalent of a developer sending you an 100,000 line PR request to review which they vibe coded and didn't bother reading themselves.

At least overall, AI should be a net positive for humanity, assuming it is in the right hands.

To quote Sam Altman:

We believe that the world should accept some bad things happening for the benefits of this technology

Or how about this one:

I think AI will probably, like, most likely, sort of lead to the end of the world, but in the meantime, uh, there will be great companies created with serious machine learning.

All the AI CEOs are convinced that their product will lead to the end of humanity, but before that happens they want to replace all of your jobs too.

I can't deny that these large language models can generate good enough code in most cases. AI lets people create software and other things that they wouldn't otherwise have the skills or the time to produce themselves.

Even non-developers can put something together that works just by using natural language. I think this is a positive thing. It makes computers more accessible for more people.

With everyone creating their own tools, greedy tech companies like Adobe can't get away with charging a monthly subscription, especially those shocking cancellation fees that even I got hit by as well.

I also think the advancements in speech to text models are fantastic too.

There is no doubt that this technology is useful, but like asbestos, which was incredibly useful and cheap, it doesn't mean it is not without severe downsides.

The lesson to learn from AI, is that it is amazing what you can do with billions of dollars in funding and no consideration for the law, ethics, humanity, or the environment.

Effect on developers #

Take software development for example. It has always felt like a hamster wheel, where we are constantly shipping features, but all AI has done has accelerated that.

It is like the hamster wheel now has a motor attached to it, making everything spin faster and making it impossible to get off.

No wonder developers are burning out quicker than before.

For many it is also sucked out the joy of the job. Part of the satisfaction came from working on a hard problem, using your brain and being proud of yourself once you solved it. A lot of that joy has now gone.

AI has also destroyed a lot of the developer communities. There are less people writing blog posts about programming and new frameworks. A lot of the posts on places like dev.to are now all AI generated slop. Obviously, no one bothers asking for help on Stack Overflow any more and the tech scene has been reduced to people bragging about their markdown files that they got Claude to write.

The part that bothers me the most though, is the fact that to work in tech now, means giving up on your morals.

If you look at any job listing for software engineers they all expect you to be AI native and to use AI to accelerate development.

Obviously, there are a lot of unethical ways to use AI, such as setting agents off and letting them hack into various companies, without even bothering to use the most basic of security measures or monitor what its doing.

Most people however, are not using AI unethically, but you can't say that the way these models have been produced is ethical and that the companies running these models have the best intentions.

By paying into these companies, you are essentially signalling that you are OK with companies plagiarising other peoples work, including your own. That ruining whole industries after stealing from them is absolutely fine, and that a small increase in productivity for yourself is more important than supporting human created work.

I shouldn't be surprised by this because it seems to be Silicon Valley's playbook. Spotify for example started with distributing pirated music. At least artists eventually got some money from this arrangement, but I don't see the AI companies paying all the websites or developers for all the data they used for training.

Given all this, I am not sure if you can be a morally good person while using AI on a daily basis, especially these frontier models.

I understand for a lot of people it's use AI or be laid off, and I completely get you need to put yourself and family first. Most people don't have the luxury of saying no unfortunately and are just trying to make the most of a bad situation.

However, being pro-AI and burning through millions of tokens a day with your software factories and your multi-agent systems seems a bit hypocritical if you are otherwise environmentally conscious by driving an electric car and recycling.

I don't think AI is going away, but it does feel morally wrong to use these frontier models knowing what the AI companies agendas are. All we are really doing is providing training data for these companies so it can replace even more jobs, while paying for the privilege.

Productivity #

The main argument for using these frontier models is that it's the future, you need to use it or be left behind and that it makes us more productive.

The jury is still out on how much a boost AI gives us though.

Some surveys say it makes us 10% more productive, others say that it's 6x, 10x or even 1000x more productive.

Why are people seeing such a big discrepancy?

I think it is generally agreed even among pro-AI people, that having a decent knowledge of software engineering practices is critical in creating software with AI that isn't just slop.

Anyone can prompt and get something that works but to get decent software that is maintainable takes skill, and you need to understand what you are doing.

That is why every one is hiring senior developers and not juniors because we already know how to produce good software, and can guide the AI when it makes mistakes and stop it from producing spaghetti code.

Those that are only seeing maybe a 10 - 30% increase in productivity are developers that are still reading the code, reviewing everything that the AI is doing and making sure that the code is maintainable.

Obviously the less you care about making good software, the quicker you are going to be. When you stop reading the code and let AI do all the reviews, obviously it is going to be quicker.

What a lot of people don't seem to be talking about is the fact that coding is a skill. It is a skill that many of us have spent a decade or more mastering, and that is the reason we are able to get the most out of AI.

It is also true that if you don't actively develop those skills you will slowly lose them over time.

The more you use AI to do the coding, and increasingly you're thinking too, you're not getting any better at software development. You aren't learning new skills, and your abilities that make you good at guiding AI will slowly disappear. You eventually end up only marginally better than someone who is vibe coding everything.

If you have spent the last 6 months to a year only using AI, then try spending a day this weekend coding by hand. I suspect you will find it an incredibly painful and frustrating experience.

Things you knew how to do before will seem impossible now. It might be that you have just forgotten how to do them but more likely is that you no longer have the patience or discipline to do it the hard way any more.

At the end of the day you need to take responsibility for the code you produce whether it is with AI or not.

AI is not going to be the one that gets called up at 2am to fix a production issue, you are. If you don't understand your code base, and the code is all a mess, you just better cross your fingers and hope Claude hasn't gone down again when you need to fix things.

Code Is Cheap Now #

The common argument for those that are saying you “no longer need to look at the code”, "code quality doesn't matter" or "code is cheap now", are assuming that we are always going to have cheap access to these frontier models.

Yes a $200/month Claude subscription is incredibly cheap for what you get. You are effectively getting between $2,000 and $10,000 worth of tokens for that price. These companies are heavily subsidised by VC money at the moment.

You might say that is true for all Silicon Valley startups but not to these insane levels. Big Tech has around $3 trillion dollars of debt tied up in AI. This is why everyone is pushing it so hard, the amount of money they are set to lose if things don't work out is crazy.

OpenAI only expects to become profitable by 2030 yet at its current rate it is likely to run out of money by next year. Some estimates say they need to earn an additional $200 - $300 billion dollars a year in revenue just to stay afloat.

Anthropic isn't doing much better, as we saw from their S-1 prospective. With some creative accounting they can say they have had their first profitable quarter, but that is only on inference revenue and compute costs. It doesn't account for all the other expenses which would put them deeply in the red.

Anthropic has committed to over $400 billion in spending over the next 10 years that it can't back out of. How they are going to pay for that when it is earning less than $5 billion in revenue a year, I have no idea.

The way I see it, is that there are a few options on how this is going to play out.

The US might decide that AI is critical for national security, and that they are in an arms race with China.

In which case, they are likely to bail out these AI companies so they can keep going. This will likely mean high taxes at some point for Americans too.

They can't keep burning money, so the subsidized subscriptions will disappear. All the developers that have gone down the full agentic AI rabbit hole, will likely need to pay out huge sums of money to keep doing what they have been doing.

The other option is the frontier models get completely pulled or handicapped and only government agencies and a handful of companies will be able to keep using them.

The only option then is to switch to the open weight models, assuming the AI companies haven't succeeded in banning them. These models are mostly trained by using the frontier models for training data, so without subsidised access they might not advance as quickly as they have been up until now.

The comparable open weight models we have at the moment like GLM 5.3, are around 756 GB and even GLM 5.3 Flash is over 300 GB. No one is running these on their home machines. Even the heavily quantized models are huge and not as capable.

Some smaller models like Qwen3.8-27B are useful for software development, but they aren't in the same league as the frontier models, and still require a pretty beefy machine to get acceptable speeds from them.

Yes, if you are willing to spend $10,000 to get the latest machine, then you might be able to get frontier level speeds out of them, but not frontier level quality. You are still going to be back to checking all the code for mistakes.

There are plenty of people offering open weight models in the cloud via OpenRouter but supply and demand dictates that the price for these models is probably going to go up if frontier models become too expensive. They are also the easiest to ban if AI companies get their way.

At least with open weight models you aren't supporting companies like OpenAI and Anthropic, but they still are a product of the same system, and come with many of the moral concerns like IP theft and environmental damage.

What to do now? #

So what are we to do? I definitely don't have all the answers.

In one corner, you have people that have fully embraced AI and are now incapable of doing anything without it. If AI subscriptions disappear, or companies have to increase the token costs to be profitable, then they are going to be left with a huge bill. Maybe that is just the cost of doing business now?

In the other corner, you have people that are refusing to use AI because they don't want to lose their skills that they have spent years honing. They don't want to lose their critical thinking ability or compromise their own ethics and morals and become a worse version of themselves.

In the middle, we have developers that feel AI has sucked all the joy out of the job and are just trying to ride out the storm for as long as possible.

Personally, I don't think these frontier models are going to be heavily subsidised for much longer. I also don't think AI is going anywhere and developers will need to find ways to use it effectively.

The open weight models are catching up with the frontier models and I suspect many companies will move to hosting their own models in the future.

I am not anti-AI, I think there are genuine uses for it, but I hope in the future we have some AI models that are trained on ethically sourced data, and they become efficient enough to run on modest hardware. Maybe we will get some super efficient, AI on a chip, hardware models that will replace the need for data centres.

For now, if you want to be a software engineer and remain competitive in today's market then you will need to use AI and give up your morals at the same time. Either that or only work in regulated industries that haven't been impacted by AI yet.

Whichever way it goes, companies and developers need to be aware of all the downsides of using AI and use it responsibly to make software better not just produce it faster.

Understanding fundamentals, system design and what good code looks like is going to be more important than ever, whether AI is involved or not.

Those that are passionate about software will still want to take the time to learn, even if it is just to try to understand the code that an AI agent has generated for you.

I also don't think software development is dead because anyone can clone someone else's software.

If you take the ArtCraft example, where Brandon Thomas cloned the Adobe suite. This is someone with 15 years of experience. I don't think he could have done what he did without that. Also, it is estimated that just one of those applications cost around $2000 worth of tokens to produce.

So yes, if you are an experienced developer you too can spend $2000 to create a buggy version of any application. You will probably then need to spend another $5000 in tokens to make it production ready.

The lesson here is that the Adobe suite should never have been a subscription.

I bought Adobe Photoshop CS4 as a student for £150 in 2010. For non students it was around £650 which admittedly is a lot, but you would pay about that anyway with 2 years of a subscription.

I still think people are going to pay for software, at least for now. I maybe wouldn't bother marketing to developers, but that has always been a hard sell.

As long as you price your software reasonably, people are not going to spend hundreds or thousands in tokens to clone your app and then still have to maintain it, when they could just buy your working version for less.

I don't see people trying to clone DaVinci Resolve Studio, as they would spend a lot more than the $295 to create something that works. Especially given there is already quite a capable free version.

Hopefully we will just see fairer pricing in software and not the death of the industry.

However, I can't help but mourn what we have lost already. To get good software it used to require skilled developers that had spent years building knowledge and taking time to craft something carefully.

It now seems we are moving into the mass market, flat pack furniture, fast fashion era of software development. I hope I am wrong, but the way we are going, it is not going to be those with the deepest knowledge that are going to succeed but those with the deepest pockets.

That's just sad for all the indie developers, the underdogs and the upcoming startups that don't have the funds to compete.

We are now heading towards end game capitalism, unfortunately.

Top comments (0)