From GPT-2 to ChatGPT: My Journey into the AI Boom (and Why I Can’t Look Away)

God moves the player and the player moves the piece
What God behind God began the weaving
of dust and time and dream and the throes of death?
Jorge Luis Borges

The Beginning of the Interest?

I don’t remember the exact moment, but after reviewing my Twitter, I discovered that my interest in AI dates back more than 9 years.

source: x search

To give you some context, 2016—compared to the more recent period after 2023 with the launch of ChatGPT-3—was a pre-boom era.

source: Google Trends

However, 2016 wasn’t a dead year. Quite the opposite, it was the year when DeepMind’s AlphaGo beat Lee Sedol, a historic milestone in the development of AI. I remember trying to bring IBM Watson into the agency where I worked, but I failed and we didn’t manage to get anything out of it. The concept was still difficult to sell to brands, and the development costs were too high.

Lockdown, and Getting to Know GPT-2

In 2020, when the pandemic and lockdown hit, I signed up for a short course on AI applied to art (led by WIP Arte digital with Matthias Gatti as the professor). At that time, we were using versions of GPT-2, which wasn’t even a chat interface! Instead, it was something clunky that required searching, adjusting, and even training with our own data.

Although there wasn’t ONE cutting-edge model at the time, GPT-2—which was a text model with interesting results—started being used for experimentation with other mediums, such as images and sound. The key step was passing those files to a text format, and then training the models to return a text that, WITH LUCK, might result in coherent images.

For example, here a developer passed hundreds of Pokémon to text format to then train a GPT-2 model to generate text that, once composed, looked like a Pokémon.

https://www.reddit.com/r/MachineLearning/comments/jyh0h4/p_generating_pokemon_sprites_with_gpt2/

https://github.com/MatthewRayfield/pokemon-gpt-2

2022 Boom and Current Day 2025

2022 is a PRE and POST year. It has similar importance to the year the iPhone was launched for many industries (one of which personally affects me: digital marketing). Both technologies have in common that they created and expanded a market where there was very little before.

One difference worth highlighting is the speed with which OpenAI grew with ChatGPT. The innovation in ease of use for a massive public was immediately reflected in the time it took them to reach one million users, breaking all records.

This speed doesn’t seem to be decreasing; it seems to be accelerating. The number of days between the launch of versions of the different leading models is shortening, and on top of that, the growth of capabilities (and quality of results) is improving. https://artificialanalysis.ai/

This is why, in the last few months, I’ve been increasingly immersing myself in learning about and implementing AI. I’m going to be logging some progress and learnings on this blog from time to time. And now that it’s here… I can’t look away…

Attention…attention is all you need


Posted

in

by

Tags:

DK

Daniel's Portfolio

Online