AI and the New Pace of Scientific Progress
For much of the last decade, AI has been discussed in the context of efficiency: faster workflows, better automation, smarter analysis. What’s becoming clearer now is that AI is starting to influence something much more fundamental which is how scientific discovery itself happens.
In fields like pharmaceuticals, biology, and materials science, progress has traditionally been slow and expensive by design. Experiments take time, lab work is capital-intensive, and iteration often happens one hypothesis at a time. AI is beginning to change that rhythm, not by replacing scientists, but by giving them new ways to explore, test, and refine ideas far earlier in the process.
What links companies like Project Prometheus, Xaira, Periodic Labs, and Thinking Machines Lab is not just technical ambition, but a shared belief that the next generation of breakthroughs will come from combining AI with deep scientific infrastructure, long-horizon thinking, and highly specialized teams. Together, they offer a glimpse of how AI is reshaping the mechanics of innovation.
From Optimization Tools to Discovery Engines
In pharma specifically, AI is increasingly being used upstream rather than at the end of the pipeline. Instead of simply analyzing trial results, models are helping researchers identify viable targets, predict how molecules will behave, and narrow down the most promising paths before clinical development even begins. This doesn’t remove risk, but it does reduce wasted effort and shortens the distance between insight and action.
A similar shift is happening across the physical sciences. AI-driven simulation and modeling are allowing teams to explore design spaces that would previously have taken years to map through manual experimentation. For executives, the implication is subtle but important: AI is no longer just a productivity layer sitting on top of R&D. It’s becoming part of the core mechanism through which discovery happens.
Why This Matters for Leaders, Not Just Scientists
This evolution is starting to reshape how science-led companies are built and scaled. When discovery cycles move faster, decision-making changes with them. Capital is deployed earlier, teams grow differently, and the line between research, operations, and commercial strategy becomes less distinct.
It’s also changing where serious talent and long-term investment are flowing. Rather than chasing short-term applications, investors and operators are increasingly backing platforms that aim to rewire the foundations of how science is done. The companies emerging in this space don’t look like traditional software businesses, but they don’t behave like legacy research institutions either.
Project Prometheus and the Shift Toward Physical-World AI
Project Prometheus sits within a growing wave of companies pushing AI beyond language and into the physical sciences. What sets Project Prometheus apart is the scale and ambition behind it — alongside the fact that it is spearheaded by Jeff Bezos. Backed by $6.2 billion in funding, the company has the resources to compete in an area where progress is expensive and slow by nature.
It has already recruited close to 100 researchers, many drawn from leading AI labs, and is reportedly exploring systems that learn not just from data on the internet, but from interaction with the physical world itself. This approach mirrors efforts at companies like Periodic Labs, which plans to run large-scale experiments using robots, allowing AI models to learn through physical trial and error rather than purely statistical pattern matching.
The significance of this shift is hard to overstate. While existing AI models excel at recognizing patterns in text, code, and images, physical-world intelligence requires systems that can grapple with uncertainty, experimentation, and cause and effect. That’s where breakthroughs in areas like drug discovery, materials science, and advanced engineering are most likely to emerge. Project Prometheus represents a belief that the next generation of scientific progress will come from AI systems that learn the way scientists do.
Xaira Therapeutics and the Reinvention of Drug Discovery
If Project Prometheus represents AI’s push into the physical sciences at scale, Xaira Therapeutics shows what that shift looks like inside modern drug discovery. Launched in 2024 with $1 billion in venture funding, Xaira is built around a simple but ambitious idea: that AI can fundamentally change how medicines are designed, not just how quickly they’re tested.
The company was co-founded by Nobel Prize–winning protein scientist David Baker and led on the technical side by Hetu Kamisetty, whose career has tracked the slow maturation of AI in biology. For years, the ambition of designing proteins and antibodies from scratch existed largely in theory. What has changed, according to Kamisetty, is that advances in generative AI and diffusion models have turned that ambition into a usable technology. Tools like RFdiffusion and RFantibody now allow researchers to design functional biological molecules rather than discovering them through trial and error alone.
Xaira’s focus goes beyond molecule creation. Its platform is designed to address three long-standing bottlenecks in drug development: identifying the right biological mechanism, building a molecule that can meaningfully intervene, and matching that therapy to the patients most likely to benefit. By combining AI-driven protein design with biological insight and patient stratification, Xaira is aiming to tackle diseases that have historically resisted traditional approaches. For executives watching the sector, the significance lies less in any single drug and more in the operating model — one where AI is embedded at the foundation of discovery rather than layered on later as an optimization tool.
Periodic Labs and the Rise of AI-Driven Scientific Discovery
Periodic Labs is pushing AI beyond analysis and into the act of scientific discovery itself. Rather than training models solely on the internet’s finite supply of text and code, the company is focused on giving AI systems the ability to generate new knowledge through real-world experimentation.
Its approach centers on autonomous, closed-loop laboratories where AI agents can propose hypotheses, run experiments, analyze results, and learn directly from physical outcomes — including failed experiments that are rarely published but often highly informative. By generating proprietary experimental data at scale, Periodic aims to train models that understand cause and effect in the physical world, not just statistical patterns in existing datasets.
Backed by a $300 million round led by Andreessen Horowitz and supported by investors including NVentures, Accel, and Jeff Bezos, Periodic is initially applying this model to physics and materials science. Early applications include work on high-temperature superconductors and collaborations with semiconductor manufacturers facing complex engineering constraints. For business leaders, the implication is significant: faster discovery of new materials could unlock advances across energy, computing, and manufacturing. Periodic’s bet is that the next leap in AI capability will come not from more data scraped online, but from systems that learn by testing ideas against reality.
Thinking Machines Lab and Making Frontier AI Usable
Thinking Machines Lab approaches scientific AI from a different angle, but plays a critical role in the same shift. Founded by former OpenAI CTO Mira Murati alongside a group of senior OpenAI researchers, the company emerged in 2025 with a $2 billion seed round and a clear thesis: the biggest bottleneck in applying AI to science isn’t model capability — it’s usability.
Instead of chasing ever-larger foundation models, Thinking Machines has focused on making powerful models easier to adapt to specific scientific and industrial problems. Its first product, Tinker, allows researchers and developers to fine-tune advanced open-source models for narrow domains without the cost and complexity of large-scale infrastructure. By using techniques like Low-Rank Adaptation (LoRA), Tinker makes it possible to customize models efficiently while retaining strong performance — a practical requirement for fields like biology, chemistry, and materials science where general models often fall short.
For business and research leaders, the significance of Thinking Machines lies in this focus on translation. Breakthroughs don’t happen just because better models exist; they happen when teams can actually use them inside real workflows. As the company moves toward releasing its own models and expanding multimodal capabilities in 2026, it represents a broader industry pivot toward AI that is adaptable, cost-effective, and embedded directly into scientific work.
What This Means for the Next Wave of Scientific AI
Project Prometheus, Xaira, Periodic Labs, and Thinking Machines Lab all point to the same shift: AI is moving beyond productivity tools into systems that can generate new scientific knowledge and accelerate real-world discovery. Whether through autonomous labs, protein design, or more efficient AI infrastructure, each company is tackling how to shorten the path from idea to breakthrough.
What stands out is not just the technology, but how these businesses are being built. Long-term capital, deep scientific talent, and strong operational leadership are becoming just as important as model performance. Turning AI-led discovery into lasting impact now requires teams that can manage complex R&D cycles, scale responsibly, and support sustained innovation.
At Harmonic, we work with high-growth AI and deeptech companies as they build the finance and operational leadership needed to support this kind of scale. Fin Glanvill, Co-Founder of Harmonic, partners with some of the most successful AI companies at key growth moments. If you’d like to discuss how teams are evolving in this space, you can reach Fin at [email protected].
