Without a doubt, the most exciting trend in the past few weeks has been the realization that AIs are finally starting to discover new science in areas such as materials science and medicine, and more recently, physics and mathematics, including offering the first real progress on a Millennium Prize problem in more than a century.

If true, it would make good on one of our main predictions in our 2025 list at the beginning of the year, in which I predicted AI would have its first real discovery. And what do we know, that prediction might come true next week.

But is this what we think it is? What has changed?

Today, you will learn how AI is changing the world in magical ways, and more importantly, what evidence and research back the realization that, this time, it’s for real, including a brand-new research paper that has proven one of my long-lasting beliefs to be wrong.

Let’s dive in.

Whenever we think about AI, we usually think first of productivity.

That is, AI’s most significant benefit to civilization is the capacity to make our world more productive, allowing tasks, businesses, and processes that yield economic value to be enhanced, doing more with less, or even automating them altogether.

But is productivity the best AI can give us?

I believe not, as more than a year ago, I was championing the idea that AI’s greatest superpower would be its ability to answer humanity’s most burdening questions.

And we might finally start to see this at scale. But first, let’s do a quick rundown of AI-led discoveries in the last few years.

We can begin with the clearest case of discovery: materials. Almost two years ago, DeepMind’s GNoME used AI to predict the stability of inorganic crystals at an unprecedented scale.

The headline result is 2.2 million new crystal candidates, with roughly 380,000 predicted to be stable; hundreds have already been independently realized, and an autonomous “A-Lab” synthesized dozens guided by those predictions.

The claim here isn’t that AI mixed chemicals on its own, but that it drastically expanded the viable search space and provided experimentally useful “recipes,” turning months of trial-and-error into a ranking problem that labs can actually prosecute.

These days, companies like recently-unveiled Periodic Labs want to take AI’s pattern-matching superpowers to professionalize what DeepMind showed us with a fully-fledged AI scientist. The key is that while AIs can propose new materials and other stuff we’ll see below, they lack the capacity to test those potential discoveries.

Thus, after raising $300 million, the idea of this startup is to create autonomous laboratories where AI can not only propose new ideas, but also test them.

The goal? Create new knowledge.

Besides material engineering, another area that holds great promise is none other than healthcare, particularly in the field of antibiotics.

MIT’s team trained deep models on ~39,000 compounds against MRSA (bacteria), then screened ~12 million molecules to attribute antimicrobial activity to specific chemical substructures.

Crucially, they didn’t stop at “ok, these work.” They also employed a method to determine which parts of each molecule the AI considered most important.

That clue helped them figure out a likely way the drugs work: they scramble the bacterium’s “battery”. In other words, the AI didn’t just find new shapes of antibiotics, but it pointed to why those shapes work, giving chemists a head start on tweaking and designing even better ones.

Another way that has proven capable of ‘connecting the dots’ came from Professor Buehler from MIT. He built large, multimodal knowledge graphs, then “hunted for” isomorphisms across domains.

For instance, he found out that Beethoven’s Ninth Symphony and certain biological materials share structural motifs, or that Kandinsky’s Composition VII suggests a hierarchical mycelium composite with concrete, testable properties.

All in all, AI’s capacity to connect seemingly unrelated domains or capture key patterns in materials or antibiotics has been known for quite some time.

So why all the fuss now? What’s new?

And the answer is the potential emergence of physics and mathematics’s ‘ChatGPT moments’.

The Clay Mathematics Institute defined a set of mathematical problems that, if solved, would earn the solver a million dollars. These are known as the Millennium Prize Problems, and to this day, only one of them has been solved.

Among them is the challenge of proving whether or not the Navier–Stokes equations (which describe how fluids like air and water move) can produce singularities. A singularity is an event where the equations predict something physically impossible, such as a fluid’s velocity becoming infinite.

Understanding whether such singularities can arise would tell us where our current equations for fluid motion stop working, and might even point toward new laws of physics.

And a few days ago, Google DeepMind, working with mathematicians from several top universities, used an advanced AI system to discover new examples of singularity-like behaviors directly from the equations themselves. These are the most precise and complex examples ever found, revealing how fluids might collapse into infinite motion under the right conditions.

Also in the last week, such claims of AI’s capacity to discover have focused much more on the mathematical domain, based on two recent comments by renowned mathematicians Scott Auronson and Terence Tao (the latter considered by many—and officially, I believe—as the smartest human alive, sometimes referred to as the ‘Mozart of Math’).

Scott Aaronson recounts that, during work on an oracle-separation paper (“Limits to black-box amplification in QMA”), a key technical move (the right function to analyze) emerged from a back-and-forth with GPT-5 Thinking, after earlier attempts with weaker models had failed.

Aaronson is careful about the claim, however: the system didn’t write the paper, but it provided the crucial “should-been-obvious” insight that unlocked the proof. That’s a qualitatively different claim of discovery: not enumerating candidates, but injecting a mathematically relevant idea at the right time in a genuine research argument. 

On the other hand, Terence Tao offers a parallel but complementary exciting story. He reports an extended conversation with an AI, starting from a plan to generate code to search for counterexamples, then pivoting into an interactive reasoning session that proved more productive than the brute-force route.

The point is not that the AI “proved a theorem” end-to-end, but that it acted like a capable and persevering collaborator, capable of following hints, adjusting its strategy, and keeping the discussion within the space of mathematically meaningful moves. For working mathematicians, that’s the part of discovery that actually matters day-to-day. 

We can’t disregard the importance of these claims.

These aren’t incumbents with money on the line anymore, these are famous mathematicians and physicists being brutally honest about AI disrupting their own field, one they would be tempted to protect from the risks of AI job displacement, being ‘invaded’ by AI models that many call “just stochastic parrots” yet are now actively participating in the discovery of new maths (always with the help of a human, of course).

But beyond maths, there are reasons why these last two “discoveries”, the maths ones, are different from the rest we have covered. AI is no longer just pattern matching; it’s doing something else; they are discovering things they couldn’t have possibly known before.

And to answer why that is the case using proven research and knowledge, we are diving deep into the roots of creating AI models to understand why, ending with a very recent research paper that has forced me to correct one of my strongest beliefs regarding AIs:

The fact that, after all, AIs might not be just the imitation of intelligence, an illusion of intelligence, but the latest models might be the beginning of the creation of real intelligence.

We’ll even have time to describe what comes next based on the comments by legendary AI researcher Sutton. So we aren’t just answering what’s emerging now, but also what’s coming next.