Disclamer
First of all I wanted to make a small disclaimer.
Iâm not a fan of the theft that is happening. Iâm well aware of how LLM training data is obtained. It does not matter if itâs OpenModel or proprietary one. The feed which is used to train is build on: stolen books, blog posts content, news websites and evil content from Reddit.
Itâs bad. I know itâs bad and still I have to use it. I feel like if I donât jump into this train in 2 years I wonât be able to find a job. A job that sustained me for most of my adulthood.
Iâm scared, frustrated and angry but I also know that this is literally the future technology we engineers will have to embed into our workflows.
So yes I disagree with how current LLMs are trained. Still I need to learn about them and use them to sustain my life.
My journey though LLMs
I rejected LLMs for many years. Itâs not like I didnât try to learn how they work or attempt to use them. They were just never part of my daily work.
This changed though after company I work for decided to give Copilot licenses to all employees. This was a breaking point.
My colleagues at work did not work faster or delivered better results. The managers though expected this to be happen. AI KPIs were created. Work of employees started to be semi monitored and the fear of being left out spread though the air.
It was around 1.5 year ago when I tried to use LLMs to deliver good quality software and at that time it was just not possible. No matter what I did LLMs were more unhelpful than simply not using them.
This changed though and it changed not that long ago.
Anthropic released Claude 4.5 [1] and it started to actually help me in daily tasks. Writing tests was the first place where I noticed that LLMs can really shine. I started with simple "Create me tests for this function" and evolved into Take this service and cover all corner cases with tests. Use [n] as example as much as possible. Keep it all in one file..
This approach allowed me to write a lot of code. Crazy amount of code and as Iâm true engineer all the generated slop went through my review. At first tests could be improved but quickly I noticed that I donât have anything to change. Tests were perfectly handling all scenarios which I found necessary.
Things I use LLS for
After that I started to use LLMs for way more.
- Writing unit tests which I didnât even specify constrains for - Yes. I donât specify constrains to most of tests I write now and itâs because LLMs is really good with finding it without my help.
- Writing integration tests between services - because I know how to architecture for various pararigms itâs possible for me to create prompts which actually generate integration tests that make sense. Sadly there my input is still needed and I really need to specify constraints to achieve good results.
- Writing E2E tests with Playwright [2] + Playwright-CLI [3]. This was actually mind blowing moment for me. Iâm a Full Stack Engineer and E2E tests were always the tedious part of my work. Finding html selectors, going through flowsâŚâ I just hated it. Thanks to agents and Playwright-CLI itâs no longer the case. I instruct agent to go through certain flow and generate test for me. SurprisinglyâŚâ THIS FUCKING WORKS. And I donât mean it works 30% of the time. It works 90% of the time. Itâs such a time saver I think developing some tool which could make it even easier would be a real game changer for developers.
- Writing logic based on my well defined input usually in form of PlantUML. Before I even write a tiny piece of REST service I always start with defining it in PlantUML which later is a key input for my coding agent. I literally ask LLM to scaffold me OpenAPI generated clients and cover flows written in PlantML. At first I noticed agent was generating trash but this changed half a year ago where it really started to implement flows without errors.
- updating libraries and configuration files - Everyone who writes software knows that the biggest struggle is usually setup and testing. Implementation part is the fun part that usually does not take that much of mental capacity. LLMs are just great with setting up projects. From config files to environment variables. Starting work was never easier.
- I use agents to interact with Azure DevOps and literally Iâm quite sure itâs possible to replace your PO with it.
Things at which LLMS are not that good at
First of all they are not good at reviewing code. Every time I use agents to review PRs I feel like Iâm not only wasting tokens but also time. For small PRs itâs okish but for big ones I always get cosmetic remarks or irrelevant comments.
Second of all I still canât use them with this whole "vibe codding" approach. Making markdown files and generating big chunks of code just does not work long term. It usually works till second big requirement comes in. After that project falls with domino effect. Thatâs why also Agents are so good at implementing REST services. Small scope to move or modify data is just the sweet spot for such tools.
I actually noticed that for big projects LLMs which refactor big parts of the code is still just a dream of tech bros. You can generate one semi working project and then to make it to useable state you need to pump so much money and time.
Also with big changes which are introduced by AI there is always a question of ownership. Who owns the product? Who can fix it and make sure itâs reliably running in production environment?
I feel like over 15 years IT evolved from professionals doing good work slowly to gardeners and baristas scripting websites for 20k euros a month. Now with AI such people are worse than ever and produce garbage at incredibly fast rate.
OpenSource developers and corporate
The current model where if you want to be productive you need to buy access to AI tools is horrible butâŚâ Looks like thatâs the future. At least in the corporate world of commercial software. This AI box wonât be closed and we will only see new advancements.
What about OpenSource though? Do you need AI to code free software?
I always looked at Open Source as I look at paintings in gallery or artistic expression. This view of mine did not really changed but software engineers evolved. They moved from professionals who code to make world a better place to machines driven by capital.
10 years ago seeing an open source project which core features were only enabled if you paid some money did not happen often. Today itâs happening for all major projects.
With this in mind I think there is a huge niche for true OpenSource developers. Ones who wonât be bought by governments or corporations. Ones that will develop code which AI guide rails would never allow to generate.
And to anyone who works this way I want to say Thank You. You are making our miserable, IT shithole a better place.
Recent drama with Torvalds
Recently Torvalds [4] said that people who reject LLMs in their work express stupid arguments and innovative tools are almost always good.
Sadly I agree with him in the most of his recent posts and comments but I also understand people who categorically reject to use LLMs. Itâs clearly a theft and even if itâs improving our life theft is a theftâŚâ
On the other hand AI is here to stay and if you want to make money you canât afford to miss out on it. If you donât want to make money and just have fun thatâs a different story. Develop code however you like but be aware that you will probably loose your job at some point. Thatâs sad but itâs true. Market in capitalism does not wait for people who reject change.
And sadly we live under capitalism that will milk us till we die.
But yeah Iâm going to wrap it up here. Take good care of yourself no matter if you use AI or not. We all need some peace as times are rough.
Future
For quite some time I felt safe about the future but now with AI advancements Iâm no longer have a peace of mind.
Some things are certain for me. Large language models are here to stay and they will change the world. Will they be proprietary models with trillions of parameters or small, optimized models? Time will tell.
Personally I think about them as technology in its early stage. Huge potential is there but for the money which is put into investments Iâm not really sure if itâs possible to get good return. Mostly because there is nothing innovative about way LLMs are trained. New "open" models pop up every week and clearly they start to be as good as proprietary ones. Itâs hard to deploy them though. For this you need either super expensive computer or couple thousand Euros a month.
Still I think current LLM optimizations are far from optimal. Most of weights are not fully used when network operates and for sure with nought math it will be possible to run even the smartest model on our smartphones.
Also, if LLMs are step towards AGI any model which will replicate itself with incremental advancement I think it would be also able to develop optimizations or techniques to run itself in cheaper environment.
How much Iâm willing to pay
Currently at home I use Mistral AI and i still donât pay for it. Maybe in couple months I will actually invest in some premium plan.
For sure I wonât pay more than 50 euros a month. Thatâs my hard limit.
Also for senior engineer jumping into working with those models was one of the easiest tasks I ever done. There are plenty of courses and "specialists" in AI field but itâs actually not that hard to use.
I feel like thatâs why I also donât find pleasure in grinding my LLMs experience.
Well see you all in one year. xoxo
PS:
This whole blogpost was written on this Logitech keyboard and âŚâ And a fucking Boox GO Color Android ebook reader.
Sources
[1] âIntroducing Claude Opus 4.5.â [Online]. Available: https://www.anthropic.com/news/claude-opus-4-5.
[2] âPlaywright.â [Online]. Available: https://playwright.dev/.
[3] âPlaywrightCLI.â [Online]. Available: https://playwright.dev/agent-cli/introduction.
[4] âLinus Torvalds shares his thoughts on Anti-AI coding rules.â [Online]. Available: https://www.reddit.com/r/linux/comments/1x6z4fq/linus_torvalds_shares_his_thoughts_on_antiai/.