Picture a world where cold tumors become visible targets for your immune system, and fusion reactors generate limitless clean energy. Sounds like a future that's "perpetually thirty years away", right? Well, buckle up... because AI is making major breakthroughs on both fronts, and Google and OpenAI are both doubling down on AI for the physical sciences in an effort to prove the validity of the industry's current spending trajectory and model progress.
But let's be clear about what "breakthrough" actually means here. What we are talking about is AI moving from "helpful research assistant" to "active scientific collaborator that generates genuinely novel hypotheses"... and then proving those hypotheses work in the lab. That's a fundamentally different capability than anything we've seen before, and it's happening right now across multiple scientific domains simultaneously.
Google's Gemma Model Discovers Novel Cancer Therapy Pathway
Google DeepMind released a quite amazing piece of news this week: their Cell2Sentence-Scale 27B model (built on the Gemma family) not only analyzed cancer data, but predicted an entirely new approach to fighting it, then proved itself right in the lab.
Here's the setup: Many tumors are "cold," meaning they're invisible to your body's immune system. To make them "hot" (and therefore killable), you need to force them to display immune-triggering signals through antigen presentation. The challenge? Finding a drug that acts as a conditional amplifier—one that only boosts the immune signal when low levels of interferon (a key immune protein) are already present.
Google's model screened over 4,000 drugs across two contexts: one with intact tumor-immune interactions, one without. It identified silmitasertib (CX-4945), a kinase inhibitor, as a conditional amplifier. The kicker? Inhibiting this particular kinase has never been reported in scientific literature to enhance antigen presentation. This was a genuinely novel hypothesis.
So they tested it. In human neuroendocrine cell models—a cell type the model had never seen during training—the combination of silmitasertib plus low-dose interferon produced roughly a 50% increase in antigen presentation. That's a massive jump in making tumors visible to the immune system, opening a promising new pathway for immunotherapy.
The model's prediction worked because it could reason about context-dependent effects—an emergent capability that smaller models simply couldn't handle. Between 10-30% of the drug hits were already known in prior literature, but the rest? Completely new discoveries.
This didn't happen overnight. Google has been building toward this moment for a decade, quietly assembling the tools needed to crack genomics at scale. Since 2015, they've developed AI tools like DeepVariant (which helped complete the final 8% of the human genome), AlphaMissense (which predicts disease-causing genetic variants), and most recently, DeepSomatic—an AI model that more accurately identifies cancer-related genetic mutations, catching key variants that prior state-of-the-art tools missed.
The cancer therapy breakthrough isn't a lucky shot—it's the culmination of years of work creating what Google calls an "innovation flywheel": real-world problems drive fundamental research, which leads to solutions, which uncover even more interesting problems to solve. After 10 years of driving discoveries from research to reality, Google now has a suite of scientific tools being used by researchers worldwide to improve healthcare and preserve biodiversity.
OpenAI Recruits a Physicist—And Enters the Science Game
While Google was hunting cancer therapies, OpenAI made its own power move: hiring Alex Lupsasca, a theoretical physicist known for black hole research and recipient of the 2024 New Horizons Breakthrough Prize in Physics. He's OpenAI's first research scientist on their new OpenAI for Science team.
The timing isn't coincidental. OpenAI is betting big that frontier models can tackle complex, multi-step scientific reasoning—and they're putting their money (and compute) where their mouth is. Kevin Weil, OpenAI's VP of Science, announced they're building "the next great scientific instrument: an AI-powered platform that helps researchers move faster, test more ideas, and accelerate the pace of scientific discovery."
They've already got receipts. In the past few months alone:
- GPT-5 Pro improved a convex optimization bound by 50% in just 17 minutes of thinking (work by Sebastien Bubeck; here's the paper).
- GPT-5 outlined proofs and suggested related extensions in a quantum field theory paper
- Terence Tao used GPT-5 in an extended conversation to solve a MathOverflow question about LCM sequences—a task that would have taken him "multiple hours" to do manually. The AI spotted mathematical mistakes in his requests and fixed them before generating code.
- Scott Aaronson published a paper where GPT-5-Thinking provided a key technical step—suggesting a rational function (the trace of the resolvent) that encoded information about eigenvalues in a way that cracked their quantum complexity problem in about 30 minutes of back-and-forth
Aaronson's reaction? "I guess I should be grateful that I have tenure." He noted the solution was "obvious with hindsight," but that's exactly the point—AI that can "merely" fill in insights that "should've been" obvious is a huge deal because it speeds up actual discovery, not just LaTeX formatting.
The National Labs Get Frontier Models
OpenAI's science push isn't just academic posturing. They signed an agreement with the U.S. National Laboratories to deploy o1 (or another o-series model) on Venado—an NVIDIA supercomputer at Los Alamos that serves as a shared resource for Los Alamos, Lawrence Livermore, and Sandia National Labs.
Translation: roughly 15,000 scientists working across energy, physics, biology, materials science, renewable energy, astrophysics, and cybersecurity now have access to frontier reasoning models. The Labs will use these models for everything from identifying new disease treatments to enhancing cybersecurity, improving detection of biological and cyber threats, and supporting nuclear security programs (with "careful and selective review" and AI safety consultations from OpenAI researchers with security clearances).
This builds on OpenAI's earlier collaboration with Los Alamos to assess risks of advanced models creating bioweapons—the kind of unglamorous but critical work that determines whether AI benefits humanity or endangers it.
OpenAI + Retro Biosciences: Engineering the Fountain of Youth
But perhaps the most stunning result came from OpenAI's collaboration with Retro Biosciences, a longevity startup. They created GPT-4b micro—a specialized mini-GPT-4o trained for protein engineering—and used it to redesign the Yamanaka factors, a set of Nobel Prize-winning proteins that reprogram adult cells into stem cells.
The problem? Natural Yamanaka factors are painfully inefficient. Typically less than 0.1% of cells convert during treatment, and it takes three weeks or more. Efficiency drops even further in aged or diseased cells.
GPT-4b micro proposed redesigned variants (RetroSOX and RetroKLF) that differed by over 100 amino acids on average from the wild-type proteins. Over 30% of the model's suggestions in the screen outperformed the original—a hit rate 3x better than typical directed-evolution screens, which usually see less than 10%.
The results in the lab? The engineered proteins achieved greater than 50x higher expression of stem cell reprogramming markers compared to baseline. They also demonstrated enhanced DNA damage repair—one of the canonical hallmarks of aging—suggesting improved rejuvenation potential.
When tested with mRNA delivery in mesenchymal stromal cells from middle-aged donors (50+ years old), more than 30% of cells began expressing key pluripotency markers within just 7 days. By day 12, numerous colonies appeared with typical stem cell morphology, and over 85% of cells activated endogenous expression of critical stem cell markers. The derived stem cell lines confirmed healthy karyotypes and genomic stability suitable for cell therapies.
The model could handle this because it was trained not just on protein sequences, but also on biological text and tokenized 3D structure data—a mix most protein language models skip. This allowed it to work with intrinsically disordered proteins (like the Yamanaka factors) and handle prompts as large as 64,000 tokens, which is unprecedented for protein models.
The Money's Flowing Toward Physical AI Science
The market has noticed. Lila Sciences—a startup pioneering "scientific superintelligence", just closed a $115M second tranche of their Series A, bringing the total to $350M. That's just one month after announcing the first $235M tranche of money.
Lila integrates AI with physical labs to interrogate every part of the scientific process—from discovering new possibilities to testing hypotheses. Their first-mover advantage means they've already yielded breakthroughs in protein binders and DNA constructs, with plans to tackle challenges across life, chemical, and materials science.
And Lila Sciences isn't alone in this race to merge AI with actual lab work. The field is exploding with companies betting that the next breakthrough won't come from bigger language models—it'll come from robots that can test hypotheses 24/7.
Lila and Retro are just two nodes in a larger network of companies betting that the next AI breakthrough won't come from scraping more internet data, but from robots running actual experiments, 24/7, generating data that's never existed before.
Another major player in this field, Periodic Labs, just emerged from stealth with a $300M seed round (led by a16z, backed by Bezos, Schmidt, and NVIDIA) to build "AI scientists" with autonomous powder synthesis labs. Founded by former OpenAI and DeepMind researchers—including one of ChatGPT's original creators—they're building labs where robots mix powders, heat materials, measure properties, and feed the results back to AI models. Their first target? New superconductors that could transform power grids. They're already helping semiconductor manufacturers solve chip heat dissipation problems.
Then there's Emerald Cloud Lab, which pioneered the concept of remote-controlled science. Scientists design experiments online, and over 200 different robotic instruments execute them around the clock in Austin. Carnegie Mellon built the first university cloud lab using Emerald's platform—meaning students can now run real molecular biology experiments from their laptops. Arctoris is another player focusing on automated, AI-driven drug discovery through remote labs.
Other AI scientist efforts include tiny startups like Tetsuwan Scientific, nonprofits like Future House, and the University of Toronto's Acceleration Consortium, all racing to automate the scientific method itself.
On the drug discovery front specifically, Google's throwing massive weight behind AI-powered therapeutics:
- Isomorphic Labs (DeepMind's spinout) raised $600M and expects its first AI-designed drugs in clinical trials by year-end. Based on AlphaFold technology that won the Nobel Prize, they've inked deals worth billions with Eli Lilly and Novartis.
- TxGemma is Google's open AI model collection for drug discovery—it understands small molecules, chemicals, and proteins, helping researchers predict drug properties.
- AlphaFold3 predicts protein structures and molecular interactions, while AlphaProteo designs novel protein binders for drug targets.
Beyond Google's efforts, the AI drug discovery field is crowded with well-funded players:
- Recursion Pharmaceuticals (NASDAQ: RXRX) combines AI with high-throughput robotics, generating massive biological datasets from cellular imaging. They've partnered with Roche, Bayer, and others, with multiple candidates in clinical trials.
- Insilico Medicine claims the first "true AI drug" advancing to human trials with its Pharma.AI platform handling everything from target discovery to molecule design. Their lead candidate for lung fibrosis showed encouraging Phase 2 results.
- Enveda Biosciences takes a different approach—using AI, metabolomics, and robotics to explore natural chemical diversity in plants. They're translating molecules from medicinal plants into new drugs, with their first IND applications filed for atopic dermatitis.
The pattern is unmistakable: AI has exhausted the internet's finite data. The new frontier is nature itself, and that requires actually running experiments, not just predicting outcomes. As Periodic Labs put it: "Unless you have an experiment in the loop, we're just thinking. Until you try it, you're no further along."
Google DeepMind Goes All-In on Fusion Energy
One of the most ambitious physical AI lab flywheels could actually be a nuclear fusion reactor. Google DeepMind just announced a partnership with Commonwealth Fusion Systems (CFS), a global leader building SPARC—a compact, powerful tokamak machine that aims to be the first magnetic fusion device in history to achieve net fusion energy (producing more power than it takes to sustain).
Fusion—the process that powers the sun—promises clean, abundant energy without long-lived radioactive waste. The problem? Keeping ionized plasma stable at temperatures over 100 million degrees Celsius is brutally complex physics.
DeepMind already showed that deep reinforcement learning can control the magnets of a tokamak to stabilize plasma shapes (work done with the Swiss Plasma Center at EPFL). Now they've developed TORAX, a fast, differentiable plasma simulator written in JAX that CFS is using to test and refine operating plans by running millions of virtual experiments before SPARC even turns on.
Devon Battaglia, Senior Manager of Physics Operations at CFS, said TORAX "saved us countless hours in setting up and running our simulation environments."
DeepMind and CFS are collaborating on three fronts: producing fast, accurate plasma simulations; finding the most efficient path to maximizing fusion energy; and using reinforcement learning to discover novel real-time control strategies. When SPARC runs at full power, it will release immense heat concentrated onto a small area—heat that must be carefully managed to protect solid materials. DeepMind's RL agents are learning to dynamically control the plasma to distribute this heat effectively.
The stakes? SPARC crossing "breakeven" would be a landmark achievement on the path to viable fusion energy, potentially solving humanity's energy crisis.
Where This Is All Going: Trivializing Hard Problems
So what's happening here? We're watching AI move from "helpful assistant that writes code" to "active collaborator that generates novel scientific hypotheses and validates them."
This is OpenAI's stage 4 of AI development—where AI contributes to novel scientific discovery—and it's aligned with Google DeepMind's stated mission to "solve intelligence and use it to accelerate scientific discovery." Both camps are proving the same thesis: AI can now generate new physics data, chemistry data, and biological insights, then use that data to crack problems that have stumped humanity for decades.
Cancer immunotherapy? Fusion energy? These are two of the hardest problems we face, and they're perfect targets for AI because they require crunching numbers, exploring vast solution spaces, and reasoning about complex, multi-step processes—all things models are increasingly good at.
Jerry Tworek, OpenAI's VP of Research, went on the MAD Podcast recently and made it clear where all this is headed. "Every week or every other week on Twitter, I do see what I think are credible reports of actual scientists using some of our reasoning models to help them perform calculations, solve hard technical problems," he said. "Solving competitions is cool, but people solve competitions to prove they can go to actual frontier level jobs and solve new technical problems. And this is what we want from our models as well."
If this works—if the positive feedback loop between AI-generated hypotheses and lab validation holds—then the only limits are physics itself and available compute. The AI bubble doesn't pop; it compounds. You crank the compute, the models get smarter, the breakthroughs accelerate, and suddenly we're living in a world where cold tumors become treatable and fusion reactors power our cities.
Terence Tao saved a few hours on a math problem. Scott Aaronson got unstuck on a quantum complexity proof. Google found a novel cancer therapy pathway. OpenAI redesigned Nobel Prize-winning proteins. DeepMind is chasing fusion energy. If the only limit to solving these hard problems is time spent in the lab, better data, and more compute... we'll need more compute to solve all three, so we'd better keep investing in it.
