ChatGPT didn't invent anything.
When the world woke up astonished in November 2022 to this “magical” chatbot, few realized that this magic was the result of decades of research. The history of artificial intelligence begins in 1943, when Warren McCulloch and Walter Pitts proposed the first mathematical model of an artificial neuron. In 1956, at the Dartmouth Conference, John McCarthy coined the term “Artificial Intelligence” and the discipline was officially born.
The '60s and '70s were characterized by excessive optimism: people thought strong AI was just around the corner. Two “AI winters” followed – periods when funding disappeared and research slowed – because promises weren't materializing. But some continued working in the shadows. Geoffrey Hinton, Yann LeCun, Yoshua Bengio – those we now call the “godfathers of deep learning” – continued their studies on neural networks when no one believed in them anymore. The real breakthrough came with three converging factors: computational power (GPUs), enormous amounts of data, and better algorithms. In 2012, AlexNet won the ImageNet Challenge by an overwhelming margin, demonstrating that deep learning really worked. From there, an unstoppable acceleration.
Once upon a time in the Carboniferous...
Before ChatGPT exploded, my only knowledge of AI came from science fiction books. Philip K. Dick and his reflections on what it means to be human. Cyberpunk in general, with its technological dystopias. Gibson's Sprawl trilogy, where AIs live in cyberspace like digital deities. Those pages were my only window to a future that seemed incredibly distant.
When I hosted the podcast Caccia al Fotone (a nice thing, but now belonging to the Carboniferous period...), I delved deeper into the subject. I read several papers published on arXiv and dedicated two episodes to AI development. In 2019, during the pandemic period, I devoured “Artificial Intelligence: A Guide for Thinking Humans” by Melanie Mitchell – a book that also helped me write a “thing” (those who know, know; those who don't, never mind...) on the evolution of computer systems and surveillance capitalism.
I thought I had a clear picture. I thought I was prepared.
Mea Culpa
Then ChatGPT arrived.
November 2022. First approach: total amazement. I couldn't believe my eyes. I kept asking questions, and despite all the initial hallucinations I encountered, I continued to have that “wow effect” typical of a child finding the most beautiful shell on the seashore (forgive me Newton for stealing that phrase, but it's always too beautiful).
And here's my mea culpa: I set aside all my protective filters that I generally have regarding privacy, open source, control over my data. I let myself go for hours of conversations on the most diverse topics. Until one night – one of many sleepless nights – I found myself discussing with that LLM about depression, various mental disorders, and how one or more abuses can influence a person's life.
When I realized what was happening, I stopped abruptly. I deleted the conversation, canceled my OpenAI subscription and didn't touch any LLM for more than a month. I was entrusting my most intimate thoughts to a proprietary system controlled by a corporation. I was betraying every principle I believed in.
But I work in IT. This is a huge revolution. I couldn't afford to fall behind, nor could I simply reject it on principle. I had to find an alternative. I began to study seriously.
Local, Always Local
I encountered the first models I could test locally. I discovered Hugging Face, and it was like finding an oasis in the desert. I began studying transformers, the datasets developed by the community. And I was astounded.
Transformers are the architecture that revolutionized AI. Presented in the 2017 paper “Attention Is All You Need,” they replaced old recurrent neural networks (RNNs) with a more elegant and efficient mechanism: the attention mechanism.
In simple words: instead of processing text word by word in sequence, a transformer looks at all words simultaneously and calculates which ones are most relevant to the context. When you read “The bank of the river was green,” the attention mechanism understands that “bank” refers to the river and not the financial institution, because it evaluates the weight of each word relative to the others.
This architecture made models like BERT, GPT, and all modern LLMs possible. It's scalable, parallelizable, and extremely powerful.
Hugging Face
Hugging Face is much more than a platform: it has become the Library of Alexandria of the artificial intelligence era. Founded in 2016, it now hosts over 500,000 pre-trained models, 250,000 datasets, and thousands of demo applications.
Their transformers library has democratized access to AI. With a few lines of Python you can download and use models that would cost millions of dollars to train from scratch. Hugging Face isn't the only platform doing this – there are also Ollama, LM Studio, GPT4All – but it's certainly the most extensive and collaborative.
Here, praise must be given to the developers: this community of people scattered around the world is doing extraordinary work. They release open source models, share knowledge, meticulously document everything. They're building a real alternative to Big Tech's monopoly on AI.
Watching this explosion of open models, global collaboration, shared code, I had a powerful déjà-vu. This is incredibly similar to the open source revolution that happened 30 years ago. In the '90s, Linux and the free software movement challenged Microsoft's dominance and proprietary systems. Many said it was impossible, that free software would never work. Today Linux powers 96% of the world's servers, all Android smartphones, and much of the Internet infrastructure.
Now the same thing is happening with AI. Llama, Mistral, Falcon, GPT-Oss, Mixtral – “open weight/open source” models that compete with (and often surpass) their proprietary counterparts. History repeats itself, and this time I know which side to be on.
Another Server in My HomeLab
I resumed studying Python, a study I had left on standby years ago. I began experimenting with training local LLM models. I added old scripts to provide my writing style (yes, it seems incredible but every coder has their own style, and it says a lot about their personality). I used Llama 3 to improve my Bash coding.
And when I was ready, I decided to make an important purchase: I bought a small server – to add to my homelab: Proxmox, pfSense, Nextcloud, WireGuard etc... – that I would transform into an OpenWebUI system.
OpenWebUI is a self-hosted web interface for local language models. Like ChatGPT, but running entirely on local hardware, without sending a single byte to someone else's servers.
For the nerds reading: the simplest way to install is obviously through Docker. Here's a basic example:
docker run -d -p 3000:8080 \
-v open-webui:/app/backend/data \
--name open-webui \
--restart always \
ghcr.io/open-webui/open-webui:main
Once installed, just connect OpenWebUI to Ollama (the runtime for local models), download your preferred models, and you're operational.
GPU usage is fundamental: a medium-sized LLM requires a lot of RAM and computing power. A dedicated GPU (like an NVIDIA GTX of various types) makes an enormous difference. For those using AMD, there's ROCm. With 16GB of RAM and an 8GB GPU, you can comfortably run 7B parameter models quantized to 4-bit.
My favorite combo?
AMD, Debian, Docker, OpenWebUI, Ollama and Mistral.
A Revolution. And a Choice to Make
We're facing a great revolution that we cannot avoid. There are two roads ahead of us.
The first: avoid it now, close our eyes, hope it passes or that someone else deals with it. And then, in twenty years, find ourselves chasing an evolved AI, probably impossible to understand, completely in the hands of those who controlled it from the beginning.
The second: study it, analyze it, use it and understand it today to be able to handle it better tomorrow. Actively participate in its evolution. Contribute to the open source community, ensure that this technology remains accessible, understandable, in the hands of many instead of a few.
The choice depends on us. And as I've learned on this (small) journey, choosing to understand – even when it's difficult, even when it means admitting you were wrong – is always better than passively submitting.
AI is not magic. It's mathematics, code, hardware, and above all: it's made by people. And if it's made by people, it can be understood, modified and shaped by people. For the better, not for the worse.