A Story of Fall and Rise#
When people talk about great comebacks, they think of PSG vs. Barcelona in 2017, Robert Downey Jr., Kasparov vs. Topalov, or Apollo 13. I’ll add another one to my personal list: Google vs. OpenAI (Microsoft). Because it’s truly astonishing how unfavorable Google’s position in AI was just a few years ago, compared to the leadership it holds today.
Between 2019 and 2023, Google (the inventor of the Transformer) couldn’t afford to fall behind a startup that had, until recently, been a non-profit. So it responded to every flashy OpenAI launch with an internal prototype.
They presented Meena as a response to GPT-2, LaMDA as a response to GPT-3, and PaLM as a response to GPT-3.5.
They showcased Imagen and Parti against DALL·E 2, and even MusicLM and AudioLM against Jukebox.
But when ChatGPT (or GPT-3.5-Turbo) arrived, its popularity was so overwhelming that an internal prototype was no longer enough. Google was forced to release a model to the general public: Bard, in February 2023. And Bard failed. It was unusable compared to ChatGPT.
That’s when we realized that Google, despite having the talent, compute, and data, did not have models ready to compete.
An important detail is that all of the launches listed above came from Google Research (also known as Google AI), an organization made up of several internal groups, the most important being Google Brain.
And what about DeepMind? Well, after being acquired in 2014, the most prestigious AI lab operated with a great deal of independence. They had their own research and their own models: the entire Alpha series, Gopher, Chinchilla, Sparrow, Flamingo…
Sundar Pichai’s response, however, was swift. Just two months after Bard’s failure, Google announced the merger of DeepMind and Google Brain, forming Google DeepMind.
And in July, Sergey Brin (one of Google’s founders) returned to actively work at the company, collaborating with the AI teams. From that point on, everything began to turn around.
In December 2023, the Gemini series was released, with a fake demo—but a real model that, at the very least, didn’t suck.
In February 2024 came Gemini 1.5 Pro, with a one-million-token context window—very promising. That same year, they launched Gemma, their open source model line.
At Google I/O 2024, they presented Gemini 1.5 Flash, with decent performance and a super competitive price; they launched Veo 1 (still far from Sora) and Imagen 3 (a very strong model).
They brought back Noam Shazeer in September and, in December 2024, released the Gemini 2 series—now with little to envy from the competition: native multimodality, almost-real demos (this time, actually real), free access on aistudio.com, and a Veo 2 finally competitive with the state of the art.
And in 2025… that’s when the real comeback began.
Gemini 2.5 arrived, leading many benchmarks for months and becoming the best image editor of the moment. Veo 3 became the best video generator, and Genie 3 stunned everyone as a true world simulator—capable of rendering video in real time, controlled by a joystick, like a video game (no one else is even close to that). It was, personally, my favorite announcement of the year.
As of December 2025, following the disappointment of GPT-5, Google stands at the top with Gemini 3 Pro.
Without a doubt, an epic comeback.