What Six Episodes of Beyond Tech Taught Me About AI

Besides the fact that I am not built for cameras…

If you looked at my calendar over the last few months, you would probably see a few blocks called Beyond Tech Filming. If you looked at my LinkedIn, you would also probably notice that I posted about each episode two or three weeks after the comms team reminded me to. I do not have a great excuse for this. I just forget LinkedIn exists.

João Almeida

Site Reliability Engineer|MB.ioneer since October 2021

Fun fact:

Unironically, he loves pinapple on his pizza.

Regardless, we’ve just wrapped the season: six episodes, all about Generative AI (GenAI), all filmed with me sitting under studio lights, trying to look like a normal person having a normal conversation. This was, let’s say, a tad outside my comfort zone. But I suffer from a serious condition called not being able to say no, and the comms team has clearly learned how to exploit that. Though to be fair to them, my own history didn’t help my case. When you spend the last few years fine-tuning GPT models on AWS, building some of our first RAG workflows for internal support, and submitting PRs to early open-source SDKs back when you still had to prompt tools using raw JSON and sheer hope, you kind of lose the right to claim you’re an innocent bystander.

So here it is, my post-mortem because, apparently, “engineering jokes” is the best thing I can do, when I’m nervous writing a recap of this project.

One thing that kept coming up in the leadership conversations, with Sílvia Bechmann, Dominik Zellner and Ricardo Jorge, was this uncomfortable truth: nobody really has a playbook for this. And that is particularly hard, especially in big companies. The instinct is to create all the rules first: define the guardrails, write the policies, lock things down, and only then let people experiment. But with Artificial Intelligence (AI), if you over-control too early, you can easily kill the thing you are trying to enable. But the better approach seems to be a bit more uncomfortable: accept that we do not have all the answers yet, let people experiment, keep teams curious, learn fast, and only tighten the system once we understand what needs tightening. Very easy to say, I know, and much harder to do when the ground keeps moving.

On the engineering side, with Tiago Santos and Luís Mendes, there was one idea that really stuck with me: the ratio has changed. Before AI coding assistants, you might spend three hours writing code and one hour debugging it. Now it can feel more like five minutes writing code and three hours debugging it. That sounds funny, but it is also a real risk. When the model writes the code for you, you can skip the part where you build the mental model. All because you did not struggle through the design, you did not make the trade-offs yourself, you did not slowly understand why the code looks the way it does. And then suddenly you are not really steering the repository anymore, you are just watching it happen to you.

However, there were also very real wins. One example I liked was shared by Ricardo Jorge where he spoke about turning a six-hour presentation slog into two hours by using AI to help with the visual layer. That is a great use case as it allows the tool to help with layout, formatting, structure, and the annoying blank-page part, so the human can spend more time thinking about the actual message. And although that feels like the right trade, the same tool that saves you four hours on a deck is not the tool you want anywhere near something like a 1:1 conversation or a performance review. Nobody is outsourcing empathy any time soon.

Design with Filomena Cardoso was probably the most skeptical voice in the room, and honestly, fair enough. If you leave these tools alone, they tend to drift toward the average of everything they have seen. And that is dangerous: it is how products start to feel the same, how taste gets flattened, how you end up with something that looks “fine” but has no point of view.

And then there is the confidence problem. One thing we’ve talked about that got stuck was about asking a model for book recommendations and getting a list of books that simply did not exist. Completely invented titles, presented with total confidence. Fortunately, LLMs are getting better at recommending books. Unfortunately, they are still very good at sounding right when they are not.

On the product side, with Tiago Pinto, the appeal is obvious. AI is genuinely useful for blank-page syndrome as it can take a pile of user feedback, synthesize patterns, and turn it into a first draft of possible features in minutes. But one warning from the episode felt important: AI features are expensive. Expensive to build, to run, to monitor and to keep useful. So, if you skip the part where you stay close to real users, understand the actual problem, and tighten the feedback loop, you are not building a moat. You’re just creating tech debt with a nicer demo.

Lastly, Luís Mendes shared his experience on how he spent a winter break building a side project in full vibe mode with paste error, get code, run it, past next error. He asked the model to continue and repeat until something works. And it did work. They shipped something real, and fast. Then someone asked them to write a simple “Hello World” in that same stack from scratch. They could not do it, and that is the trade-off we do not talk about enough. If you skip the struggle, you can also skip the learning: the annoying parts, the trial and error, the debugging, the dead ends, the “why the hell is this not working” moments, are often the parts where the knowledge sticks. So yes, AI can compress a learning curve. But if we are not careful, it can also stretch one. Something that used to take five years to truly understand might quietly take twelve if people keep shipping things they do not understand.

Personally, hosting this season pushed me way outside my comfort zone, which apparently is where useful things happen. It was a very annoying discovery. It did not convince me that AI is coming for all our jobs, and we can all move to the beach, at least not for now. But it convinced me that the boring fundamentals matter more than ever: Clean architecture, good code review, understanding your data, talking to users, and knowing when the tool is wrong or when the answer sounds good but does not make sense. AI turns more of us into orchestrators. And an orchestrator’s job is to notice when something is out of tune, not just to make things move faster.

To the crew who edited out my fumbled intros, made me look somewhat comfortable on camera, and waited patiently for me to remember LinkedIn exists: thank you.

Season wrapped. Back to the terminal.

Find the full episodes are on Mercedes-Benz.io YouTube channel, with no delays on their end. Rest assured.