While others sell us fireworks, she is rewriting science

OpenAI does not seek to create an AGI. Make sure we don’t stop talking about her. It is the maximum exponent of the “productization” of AI. It and other rivals focus on offering flashy options that improve our productivity but don’t change the world. Which is precisely what some companies are trying to do, among which one stands out in particular: DeepMind.

The AI ​​that helped science. For the past two years the tech industry has resembled a fireworks contest. Every few days or weeks a new model promises to write better emails, generate more realistic videos or have more and more human conversations. The cycle of novelty is often ephemeral —Studio Ghibli style images were a good example—but far from those “wow effects” there is a silent AI that does not seek to impress on social networks, but rather to help solve scientific problems that have been blocking new advances for decades.

The Thinking Game. The recent documentary about DeepMind titled ‘The Thinking Game’ and available for free on YouTube precisely shows us that other side of AI. Although the tone is not exempt from that epic that we already experienced with the documentary ‘AlphaGo’, what it tells us serves as a reminder of this dichotomy that the industry experiences. While the AI ​​bubble inflates in search of immediate profitability, DeepMind seems to have maintained its original spirit. One that wants to use AI not to imitate the human being, but to – in this case – decipher the code of biology.

From Pong to AlphaFold. This 84-minute documentary tells the story of DeepMind through the career of its co-founder, Demis Hassabis. This journey is fascinating and shows us how the startup began to develop AI models that taught themselves to play retro video games like Pong or Breakout (Arkanoid) to, little by little, evolve towards much more ambitious challenges. Specifically, being able to predict the structure of proteins through deep learning.

AI can change science. The challenge DeepMind engineers faced seemed impossible. Predicting the structure of these proteins was often misleading and required enormous computing power, but with AlphaFold 1 (2018) and especially with AlphaFold 2 (2020) DeepMind achieved spectacular results. In 2021 the company published both the source code of the project and a database with the structure of more than 200 million proteins available for any laboratory or researcher. It was an absolute gift for the scientific world. Then AlphaFold 3 would arrivemore oriented towards drug development and with a somewhat more commercial point.

A Nobel Prize-winning AI. Two of the 2024 Nobel Prize winners in chemistry work at DeepMind. These are Demis Hassabis and John M. Jumper, who received the award for their contributions to the prediction of protein structure. That work with AlphaFold demonstrated that AI could indeed contribute to scientific advancement, and put DeepMind on the throne of that segment more than ever.

A radically different approach. It is important to do pedagogy here. While LLMs (large language models like GPT-5) work by predicting the most likely next word in a sentence, “AI for science” predicts physical and chemical behaviors: while LLMs can hallucinate and lie like it’s nothing, scientific AI submits to the laws of physics.

From observation to simulation. Traditionally, science advanced through observation, hypothesis and experiment, which was often slow and expensive. With AI, an intermediate phase is introduced, massive simulation, which acts as a catalyst for this process. Thanks to AI, it is possible to rule out millions of dead ends before the scientist sets foot in the laboratory. DeepMind has seen this so clearly that created Isomorphic Labsa business spin-off dedicated exclusively to using this technology to discover new drugs.

DeepMind is not alone. Although the company co-founded by Demis Hassabis is the clear reference in this area, there are other examples that follow the same path:

• Microsoft– Achieved a striking milestone in collaboration with PNNL (Pacific Northwest National Laboratory) by AI-screening 32 million potential inorganic materials and finding a new one capable of reducing the use of lithium in batteries by 70%.
• M.I.T.: The prestigious technical institute used deep learning models to discover halicinean antibiotic capable of eliminating bacteria resistant to all known treatments.
• NVIDIA: The firm not only imperially dominates the AI ​​chip market, but has built a “digital twin” of the Earth called Earth-2. Its AI models (FourCastNet) predict extreme weather events thousands of times faster and consuming much less than traditional supercomputers.

The promise (a little) fulfilled. Almost since ChatGPT appeared, we were promised that AI would change the world. At the moment it has not done much, but what has been achieved by DeepMind and other companies in the field of science does seem to pose real revolutions. I recommend not missing the documentary: it is fantastic.

In Xataka | What the AI ​​pioneers awarded today with the Nobel Prize say now about AI and its risks