That gen AI hype I've mixed feelings about...

The year is 2026, and AI has swamped up every single conversation. Technical and non-technical people alike : like it or not, it exists and the fact that it is creating so much (heated) debates shows that things won't be the same anymore.

Reluctant to use Claude Code, Github Copilot and co at first, I eventually got curious when we got encouraged to use it by top management at work. I started quite small using the Github Copilot integration in IntelliJ, with chat sessions only, and providing context from the project with selected lines or added files. I was quite deceived by my first tests to be honest. That feeling stuck for a few months, until we gained Claude Code access and agentic features in Github Copilot. Then I started to use CLI tools and it was different. I was amazed, afraid and doubtful at the same time.

For the following of this post, I must tell a bit about my setup. I use Claude CLI most of the time, falling back to Github Copilot CLI when I reach the end of my quota. I have some custom skills, main ones being:

• a scala-code skill: to enforce some rules when asking for scala code. This one leverages metals, coursier, cellar which are excellent tools, by the way.
• a kotlin-code skill: same as above, but for kotlin
• a skill for documentation : useful for a large share of our components with similar characteristics, shared in the team
• a set of skills for migrating (terraform defined) workloads from a k8s cluster to another

Also, I use some MCP servers or connectors as well:

• metals-mcp: for faster scala compilation, formatting, testing,...
• Atlassian: to comment on issues, link tickets, provide more already written context
• Slack: to provide context as well
• Coralogix: the heart of our logs and metrics
• Backstage: to access all our internal technical documentation

I have not a long experience with it and here are my current opinions (Summer 2026) on this hype, categorized as the good, the bad and the ugly on a work usage only (let's keep aside social and environmental impacts).

The good

Getting shit done

This is where Gen AI shines. We are moving workloads from one (old) k8s cluster to a new one. Given the number of deployments and stateful sets to migrate, with a repetitive process, we are able to automate and handle problems quite easily. Instead of having to run manually shell scripts, we have a workflow where the Human just validates and merge PRs. The agent being able to track networking issues and reporting them with meaningful data to feed the infrastructure team to debug and fix them. For this task, I find that Gen AI is speeding things up, turning a cumbersome, manual task into a fully automated workflow. Also, it is such a boring thing to do that doing it by hand would have led me into some "laziness".

Exploring

My team owns a lot of microservices, feeding Coralogix (our monitoring solution) with tons of metrics and logs. Sometimes I find it easier to ask Claude for insights on a given component without having too much to worry about the queries to write. Same apply for code questions, Claude is able to find out things faster than me. Also, generating API clients from an OpenAPI specification (attached with a techdoc) hosted in Backstage is a breeze.

Answering support questions

I mentioned above that my team owns a lot of components. But in a team of 4, it's impossible (at least for me !) to know them all. While I was the only member of the team (others being on vacation), I received some questions and requests from product or support teams. I used successfully Slack and Jira integrations to provide context in my prompts. And plugged all that with backstage, CLI utilities (to get logs, metrics, commits, ...) Combined with basic knowledge of some components, I managed to answer faster.

Cross-cutting concerns, refactoring

One thing I find Gen AI to be good at is adding logs and metrics (Otel) in various places with a consistent structure. It can be tedious (and boring) to do, so giving it up to AI is good. Same applies to refactoring, although it could be sometimes achieved by using regex and integrated tools in the IDE...

The bad

Big features

I did try to vibe code big features using plan mode upfront without much success, or with an outcome that is difficult to review/evaluate. I get that this may evolve as agents and LLMs get better every day, but still, it shifts ownership of the code to ??? Reviewing large PRs spanning from code to deployment is not fun. It is tempting to squash and merge in YOLO style... and it will bite you later. Instead I found out that small plan/code/PR loops are better. For higher level architecture chatting about the design worked well on some occasions.

Real costs

For feature implementation in a step by step driven process, yes Claude Code works well. And the same applies for the getting shit done, as mentioned early. But beyond the tech is reality and the answer to questions like "why do we build the component ?", "what is the cost of developing it vs its value ?" Also, I mentioned how I used a lot of tools to get my way out of questions from support and product teams, but that came with high cost. Yes, answers were sourced, well structured and genuinely useful but it burned a lot of tokens. Even if I had taken some time to investigate by myself, would have it been more expensive ? I mean, is speed always the good strategy, what if those questions could have been avoided in the first place ?

The ugly

Confirmation bias

It happened to me a few times, especially when exploring the design space, to give too much context, or to hint in some direction, but wanting it to be challenged without much success. Sometimes, only the Human has the key piece of information that changes everything and is not adding it to the prompt because it is so obvious in his context. Gen AI and agents are good enough to make us believe they are way more than just probability based technologies, but we should never forget that. I didn't use Fable or any new model which apparently goes beyond in terms of thinking though.

Juniors ?

Last thing I wanted to highlight is how Gen AI is bad for juniors. When a system wakes you up at 3am, AI can help you of course, but experience will be faster when you are on autopilot, still sleepy, and perhaps out of AI quota.

The thing is, we all can use AI, it is easy to do so. It does not involve any kind of learning, it does not require effort. In a sense it makes the developer lazy. What is a junior then, what value should he/she bring to the table ?

Conclusion

(Gen) AI is changing how we work for sure, but things at stake are not well understood yet. I know there are many researchers out there trying to figure out the impacts of AI on technology, industry, society,... Solving challenges on a technical level is nice. I mean, do I really want to write myself yet another crawler, job manager,... ? Or should I generate it with AI ?

In the end, one remaining question for me is about ownership. After years of not writing a single line of code, trusting AI, will we be efficient in thinking by ourselves ? Won't we be tempted to thrive for more from AI than just code ? Owning models and the hardware will become leverage over millions of people and yet, we can count the owners in one single hand.

As for now, AI gives me time while I watch agents doing the grunt job to read Spinoza, and to question myself about the outcomes of all of this....