Thoughts about AI
Disclaimer: I’m not fond of letting AI write my blog posts. Therefore, these are my unfiltered thoughts.
Introduction
It’s been a while since I’ve updated the blog. I’ve been busy with my life and my work, and honestly I haven’t had the time nor the energy to write anything.
However, with the latest advances in the AI field I felt the need to share my thoughts.
What changed
A year ago models such as GPT-5 and its variants as well as the Sonnet and Opus 4.5 series were already showing an impressive performance, with Gemini lagging behind Gemini 3 series.
But there’s been a shift in the past month. After some expected incremental updates in their models, the latest GPT Astra and (my personal favourites) Sonnet and Opus 5.5, have crossed -in my humble opinion- an inflection point.
Current context
The capabilities that these models are demonstrating are breath-taking. It’s been now almost 4 years since ChatGPT was officially launched (based on GPT-3.5) but there’s been a MASSIVE leap in capabilities. Three years ago it couldn’t even sum numbers reliabily. It has barely been two years and half since the first agentic workflows started appearing, with a not so good performance.
But in those two years ago the advancement of their capabilities have been rising exponentially. I am truly shocked. For the past 3 years and half I’ve been telling everyone that by the year ~2035 I’ll probably be planting potatoes. Well, or I’d would be if I could afford a farm.
What I do know is that I’ll have a hard time keeping my job. And it’s not because I’m not confident in my skills. In fact, I consider myself a senior Platform engineer and probably not an average one. But despite that, the skills that the AI agents are showing lately had me wondering if I was too optimist when I thought about the ~2035 deadline -~2 years up or down-.
Last year models were already topping the IMO medals, surpassing most mathematics.
12 months later, they are solving problems that didn’t have a solution for decades. They are solving millenium problems. They are decompiling complex software in the time that I write a ~200 lines Python to parse an API.
About the future of coding
I was critizising vibe-coding just 6 months ago. I don’t do it anymore. With the right harnesses, prompting the AI agents is the future. There’s almost no point in coding manually anymore.
Sure, I still review the code that the agents issue, but I don’t pay much attention anymore to the detail. The agents definitely code much better than me. It feels like having super powers. It’s awesome and scary at the same time.
The future of coding is prompting the AI. Lots of software engineers are scared because their identity was built around their capability to write code. Well, that era is gone. We need to evolve.
Vibe-coding with harnesses
In my opinion the shift for the next 36 months or so is to learn how to properly prompt the agents while keeping a good harness; this means to define in advance the architecture to be followed, project structure, technologies, providing enough context, and so on and so forth.
I usually use an ARCHITECTURE.md file next to AGENTS.md and I instruct the agent to keep track of every functionality and classes relationships inside that file. This Works great for future modifications when the code grows larger (tens of thousands of lines and beyond) and this way I can make the prompting more token-efficient.
Anyway, since I’m not a full-time developer myself and I’m focused in managing our internal platform I’m just developing internal tooling to make our operations easier and focus in other problems that we usually didn’t had the time for.
What do I expect beyond 24-36 months
Honestly, I’m not sure. The recent advancements in AI make me feel that even platform engineering is close to an end as we know it.
In Europe we are lagging a bit behind but things are going to shift fast.
I am just surfing the wave for now.
I’ll keep you posted.