What happens when machines stop solving problems⦠and start discovering the universe?
Key Takeaway
AI systems are now doing more than solving predefined physics problems β they are formulating novel hypotheses and discovering new physical laws from data, marking a shift from AI as a computational tool to AI as an autonomous scientific collaborator.
I’ve been watching AI eat into one discipline after another for years now. Coding, legal research, drug discovery. Each time, the story is roughly the same: AI gets very good at the task, people get nervous, things move on.
But something happened recently that stopped me in my tracks.
A collaboration between Google, Harvard, and Carnegie Mellon produced results that don’t quite fit that pattern. Their AI system didn’t just perform well at a pre-defined physics problem. It found new mathematics that human researchers hadn’t thought to look for.
That’s different. Genuinely different.
Key Takeaways
- π The Problem Physicists Couldn’t Crack
- π§ When AI Stopped Calculating and Started Inventing
- π¬ Why This Matters for Physics
- π This Wasn’t a One-Off
π The Problem Physicists Couldn’t Crack
To understand why this matters, you need a bit of context about cosmic strings.
Theoretical physicists believe these structures formed during phase transitions in the very early universe β microseconds after the Big Bang, if they exist at all. They’re hypothetical: incredibly dense, unimaginably thin defects in spacetime. If they’re real, they’d emit gravitational waves that observatories like LIGO might one day detect.
The problem is that modelling those signals is fiendishly difficult. There’s a particular integral, written as I(N, Ξ±), that describes radiation emitted by cosmic string loops. For years it resisted every attempt at an analytical solution. Physicists could approximate it numerically, but that gave them:
- No clean formula to work from
- No general solution that held across cases
- No deeper insight into what the physics actually meant
It was one of those problems that sits on the shelf. Acknowledged, partially worked around, never truly solved.
π§ When AI Stopped Calculating and Started Inventing
Google’s system used a neuro-symbolic framework β tree-search algorithms combined with automated mathematical reasoning. It explored more than 600 candidate derivation paths.
Most led nowhere. But eventually, the system landed on something that made the researchers sit up.
It turned out the integral could be tackled using Gegenbauer polynomial expansions β a technique that let the problematic singularities in the integrand be absorbed and rewritten as finite closed-form expressions. Six independent analytical approaches were found. The most elegant of them revealed connections between discrete mathematical structures and continuous Feynman parameterisation in quantum field theory.
These were connections human researchers hadn’t previously identified.
That’s the part that gets me. Not that AI solved it. But that the solution required noticing something humans had missed, across a space of mathematical possibilities too large to manually explore.
π¬ Why This Matters for Physics
Solving this integral isn’t just ticking a box. It gives physicists a much sharper tool for modelling gravitational wave signatures from cosmic strings β which feeds directly into our ability to read signals from next-generation observatories.
And those signals could, in principle, tell us things about:
- What the universe looked like fractions of a second after the Big Bang
- How symmetry broke down in the early universe
- Whether dark matter behaves as currently modelled
- Whether inflation happened the way we think it did
- Whether there’s anything to the multiverse speculation at all
These aren’t small questions. This particular bit of maths is a thread that, if you pull it, connects to some of the biggest open problems in physics.
π This Wasn’t a One-Off
The same system was reportedly run against 18 open problems across mathematics, physics, economics, and computer science. A few results stand out:
- A conjecture in submodular optimisation that had stood for a decade was disproved
- Previously unidentified cryptographic vulnerabilities were flagged
- New mathematical derivations were produced and formally verified
What strikes me about that list is the range. These aren’t variations on a theme. They’re genuinely different problems, in different fields, tackled by the same underlying system.
That breadth is arguably as significant as any individual result.
β οΈ The Paradigm Shift Few People Are Talking About
The standard history of science runs like this: humans ask the questions, humans build the tools, machines help with the computation. The creative act β noticing what’s interesting, framing the problem, interpreting the answer β stays firmly with us.
What’s changing now is the middle part. Not just the computation, but the exploration. AI systems can generate hypotheses faster than humans can test them. They can traverse mathematical search spaces that no individual researcher could navigate in a lifetime. They can iterate through thousands of conceptual paths in the time it takes a PhD student to read three papers.
None of that means they understand what they’re doing. But it might not matter.
π§ The Ronnie Huss POV
I’ve spent two decades building systems that sit at various edges β SaaS, AI, decentralised infrastructure. One pattern keeps showing up, and it’s not subtle.
Technology doesn’t just speed up existing work. It reassigns it. The people who were essential to doing X find themselves instead being essential to deciding which X gets done.
That’s what I think we’re watching happen inside science now.
“AI is not replacing scientists. It is becoming a new type of scientist.”
Human researchers still do the things that require meaning. They ask the questions worth asking. They interpret what the answers actually tell us. They catch when the machine has found a spurious pattern rather than a real one. But the exploration phase β the laborious sifting through possibility space β that’s where the shift is happening fastest.
π€ The Real Future: Symbiosis, Not Replacement
There’s an easy framing available here: machines are taking over science, human creativity is being automated out of existence, etc. I don’t think that’s right, and I don’t find it particularly interesting.
AI doesn’t have curiosity. It doesn’t know why the cosmic string problem matters. It has no stake in whether the answer turns out to be beautiful or ugly. All of that still comes from us.
But once a question is asked, an AI can search the space of possible answers with a thoroughness that simply wasn’t available before. That doesn’t diminish what humans bring. It changes what humans are for.
π The Real Question Ahead
We’re entering a world where AI proposes new theorems, derives new physics, identifies new materials, and generates testable scientific hypotheses at a rate no research institution could match with human labour alone. The humans then verify, contextualise, and decide what to do with it.
That division of labour is new. And it’s going to produce a lot of discoveries that wouldn’t have happened otherwise.
The universe hasn’t changed.
But how we go about understanding it has β more than at any point since we started writing things down.
Final Thought
For most of recorded history, the discovery of natural laws has been a human project. Slow, brilliant, deeply personal.
We may have just built something that can help carry that weight.
Not because it wants to. Not because it understands why it matters. But because it can, and we pointed it in the right direction.
The question isn’t whether AI will help us understand the universe. It clearly will. The question is whether we’re ready to keep up.
References
Brenner, M. P., Cohen-Addad, V., & Woodruff, D. (2026). Solving an Open Problem in Theoretical Physics Using AI-Assisted Discovery. arXiv.
Google DeepMind. Accelerating Mathematical and Scientific Discovery with Gemini Deep Think.
π¬ Let’s Stay Connected β Signal Over Noise
If this sparked something for you β a new insight, a deeper question, or a clearer signal β I’d love to keep the conversation going.
- π§ Blog: ronniehuss.co.uk
- βοΈ Medium: medium.com/@ronnie_huss
- πΌ LinkedIn: linkedin.com/in/ronniehuss
- 𧡠Twitter/X: twitter.com/ronniehuss
- π§ HackerNoon: hackernoon.com/@ronnie_huss
- π’ Vocal: vocal.media/authors/ronnie-huss
- π§βπ» Hashnode: hashnode.com/@ronniehuss
Frequently Asked Questions
How is AI changing scientific discovery in physics?
AI systems are moving beyond computation assistance to active hypothesis generation. They can analyse vast experimental datasets to identify patterns invisible to human researchers, formulate candidate physical laws, and propose experiments β roles previously exclusive to human scientists.
What are the biggest risks of AI-driven scientific discovery?
Key risks include: AI systems identifying spurious correlations that appear to be laws but are statistical artefacts, lack of interpretability making it hard to validate AI-generated hypotheses, and over-reliance on training data that may not represent the full range of physical phenomena.
Can AI replace human physicists?
Current AI systems are powerful collaborators but not replacements. They excel at pattern recognition in large datasets and hypothesis generation within known frameworks. Human physicists still provide the conceptual creativity, experimental intuition, and physical intuition required to formulate genuinely novel theoretical frameworks.
