If you follow the news about the effects of AI on software development, you’ll encounter wildly different claims. The supporters claim that AI assistance improves speed by 40%, and that in 2030 95% of all code will be written by AI. But then there are studies that claim that efficiency actually drops by 25%, as so much time is wasted on refactoring the AI results. I’ve talked to programmers who barely write a line of code themselves anymore, others who only use it for specific tasks (like writing tests), and some who refuse to use AI at all – for entirely legitimate reasons, like environmental impact, intellectual property issues, rising costs, and worries about digital sovereignty and being dependent upon US companies.
Then there are the horror stories of open source projects being swamped with AI-generated pull requests and companies spending tens of thousands of dollars on API calls by accident. So what’s the reality, out in the (proverbial) trenches? After some initial tests, with mixed results, I’ve been using Claude Code for a couple of months now – first with a Pro subscription, now Max, for a multitude of programming problems and languages – so I think I’m somewhat entitled to give my view on where we are right now.
The memory problem
I think that many of the problems that people experience when coding with AI assistance is a failure to understand that Large Language Models don’t have a long-term memory that stores past experiences. Developers are sometimes stumped by their AI assistant making the same mistake that it did last week – I’ve had Claude edit the wrong files as it had forgotten that it created a new version in a subdirectory.
LLMs don’t learn during use. They only learn during training. After that, the weights that determine the model’s knowledge are fixed. The only “memory” that an LLM has is the context. This is all the information, such as the chat history, that is fed to the model on every request. The only way an LLM can “remember” things and take them into account is by having them fed to it over and over again. That’s why context size is so important – only with a large context window can an LLM take lots of information, like a large code base, into account. This is also why coding assistants often don’t know the latest versions of evolving libraries and frameworks and tend to produce deprecation warnings – their knowledge is frozen at training time.
A partial solution are file like AGENTS.md, CLAUDE.md, and other markdown files that contain instructions and lessons learned – some generated by the human developer, some by the coding assistant itself. Feeding them to the model with every prompt provides a context to the request. The problem is that large context windows require lots of tokens, so only the best and most expensive models offer very large context windows. Even then, understanding a very large code base and all its dependencies are difficult. You’ll notice that, even with a smaller project, a new session will have to ingest the entire code base (that it might have written just a few minutes earlier in another session) to understand what’s going on.
Claude has functionality called skills – this sounds fancy, but essentially just means that a session gets initialized with specific instructions like “you are a security auditor” etc. – this helps with steering it in the right direction, but does not magically unleash new powers that otherwise wouldn’t be available.
The consequences are quite obvious: AI models can perform great work in small, self-contained projects, but may struggle in larger projects. The more specific the task and the smaller the amount of code that it needs to take into account, the better the results will be. Even then, I’ve had Claude spend hours trying to diagnose a small issue, where it proposed lots of increasingly far-fetched solutions – until I pointed it at the ten lines of code where the error occurred.
The need for speed
When the problem is well suited to AI-assisted coding, the results can appear close to magic at some times. I’m sometimes genuinely astounded by the knowledge it has even about niche domains like structural health analysis and surveying problems. Then it knows more programming languages than I’ll every be able to learn. Not only have I used Claude to write code in Rust, Python, and Swift – it is of course also familiar with shell scripting, Docker, and Ansible.
AI excels at writing boilerplate code, the stuff developers find really boring. I especially like how quickly it can turn out GUIs, Writing a custom GUI for labeling in a machine learning pipeline is now a question of minutes, not hours or days.
If you’re still coding by asking questions in a chat window and transferring results with copy & paste. you’re doing it wrong. To truly unleash the power of AI coding, you need to use a coding agent like Claude Code or Codex. This gives it direct access to code and tools – just don’t let it do too much, or you’ll be able to tell another horror story of deleted production databases and other mishaps. But to iterate quickly, using agents is essential.
One interesting and perhaps not fully known aspect is that modern AI models are multi-modal – they do not only understand language, they can also “see”, i.e. interpret images. One of the most annoying things about photogrammetry are the many different conventions for orientation angles that usually require a lot of trial and error to figure out. Well, since Claude can see, it can also plot information and interpret it, testing these conventions itself and finding the correct one within a few minutes (most of the time – Claude has also told me that things looked great when an image and the overlaid point cloud were clearly mirrored).
The problems
AI use is currently heavily subsidized. Even the paid subscriptions don’t cover costs, so AI companies lose money and live off venture capital. I don’t think that this will go on forever – I expect that the three main coding AI model providers (OpenAI, Google, Anthropic) will at some point be forced to raise prices. Even so, token-based API use can already lead to significant bills, so I recommend using a subscription instead for development purposes.
The other threat is dependence on US AI providers, a lack of digital sovereignty. Anthropic was forced to suspend access to Claude Fable after just a few days. What happens if access is turned off completely at some point, possible as part of a trade war? Running your models locally could be considered a solution – here the problem is that the available open-source or open-weight models aren’t as capable yet as the commercial models, and the hardware necessary to run large and powerful models has only become more expensive in recent months.
The long-term impacts aren’t clear yet. Memory prices are expected to stay high for the near future. An environmental impact is undeniable now that Gigawatt-scale datacenters are under construction. It remains to be seen what the impact on the work of software developers will be. AI clearly can do the job of junior developers, but today’s junior developers are needed as they are the senior developers of the future. And I generally disagree with the use of AI to “improve efficiency”, i.e. to fire employees. Our aim should be to employ as many people as possible, not as few as possible – even from an economic standpoint.
So my main takeaway is this: AI clearly has a massive impact on software development and will continue to do so – but I void becoming too reliant on it. Use it where it makes sense and has benefits, but stay flexible and capable without it.
