Major advances in AI in science for 2024 – 12/20/2023 – Science

The other day I was talking to a very intelligent and dear friend who reminded me of the phrase sometimes attributed to the physicist Niels Bohr, sometimes to the baseball player Yogi Berra: “It’s difficult to make predictions, especially about the future.” We laugh. Great truth.

What is proving increasingly viable is the possibility of using indicators such as investments, allocated intellectual capital, economic impact and scientific viability of innovation to model scenarios and thus reduce uncertainty somewhat.

In parallel, I am convinced that it is essential to incorporate the “spirit of the time” in these analyses. The mentality of an era makes certain things more favorable than others, despite what utilitarian calculations based on cold indicators indicate.

Applying this reasoning, I arrived at the four trends below, using the criterion of maximum coverage in the broadest domains of knowledge.

Protein engineering with AI generating new therapeutic strategies

Amino acids linked in a special way have properties that cannot be found in their components and, therefore, receive a specific name: protein. Understanding the connections that generate these properties is essential for the development of the most varied cell therapies, elimination of toxic substances and much more.

In 2022, the problem of protein modeling began to be considered “solved”, with the recording of the three-dimensional structure of all known ones. The main driver was DeepMind’s AlphaFold, which has been used to solve hundreds of biological challenges, from antibiotic resistance to the emergence of Parkinson’s disease.

It turns out that the broader goal is not to understand, but rather to prevent and cure. This is the trend that should gain strength in 2024, due to the unfolding of the above process in the creation of proteins with therapeutic functions on a scale. Molecular clinical treatments must proliferate.

Development of new materials using AI

At the end of this year, materials engineering was shaken by the discovery of 2.2 million new materials, of which at least 381 thousand have applied potential.

The feat came again from DeepMind, which developed the AI ​​GNoME, whose immediate output the authors argue is worth 800 years of human labor. Right now, there is an intense movement by companies and governments to convert these findings into prototypes and even products next year, mainly in two areas: semiconductors and batteries.

The idea is to extrapolate the principles that led to the discovery of materials to simulate their properties in realistic scenarios and also determine their ideal fits on silicon wafers and other AI platforms. The chip wars are sure to boil over.

If you are a good observer, you may have noticed that there is a deep relationship between this trend and the previous one. This is the effect of the spirit of the time. We are moving from the phase of discovering things that exist, but that take a lot of work to identify manually, to translating them into practice using deep learning.

Decoding extinct languages ​​and communication systems in other species

Understanding what whales are saying as well as crows, elephants and others, is the goal of initiatives that use modified generative algorithms to refine and scale research into animal behavior. Finally, the ladies of Jardins will be able to receive messages from their dog on their cell phone.

Tens of thousands of hours of effort are being condensed by AI, which should lead us to the first results in 2024. It seems like a lot of condensed human effort, but it is only a fraction of what we have already invested in another decoding process: that of extinct languages. Suffice it to note that translating the hieroglyphs from a single source, the Rosetta Stone, took 23 years.

For this process is also beginning to change through the use of AI for its mass decoding, which in turn should expose hidden relationships in the sociolinguistic puzzle of the past. Archaeology, anthropology, history and linguistics will experience an explosion of knowledge, starting next year, with the methodology that began to take shape through the algorithmic translation of a million pieces from the Akkadian language into English.

AIs equipping robots that learn through cultural transmission

There are many challenges that cannot be faced by isolated algorithms on servers, as they are irreducible to written and even audiovisual records. Examples include relational subtleties, contextualized decoding of aromatic molecules (eg, rotting cheese can be good, rotting meat is bad), and displacement in complex spaces. We learn these things through cultural dynamics that involve our actions and bodies.

Awareness of this should lead to a new phase of AI that I called embodied. The first study showing how robots equipped with newly developed algorithms can learn through cultural transmission has just been published, suggesting that this transformative approach to mechatronics and computer science is likely to gain traction next year in combination with the use of generative algorithms like ChatGPT so that robots can “tag the world” and thus understand what we are talking about when we refer to what exists in the real world.

Like the first pair of trends, this last one is also strongly correlated; in this case, by expanding the scope of generative algorithms. A sign of the times, without a doubt.