Happy Accidents Still Matter
How Deep Tech Week Redefined Scientific Discovery in a Post-AI Era
In 1928, Sir Alexander Fleming returned from vacation to find that a petri dish of Staphylococcus cultures had been contaminated — an accident that led to the discovery of penicillin. Ninety years later, in 2018, Japanese scientists accidentally altered an enzyme and found it could break down plastic significantly faster.
What connects these seemingly unrelated moments isn’t meticulous adherence to the scientific method. These “happy accidents” weren’t born from controlled, procedural experimentation. Instead, they reflect how scientific breakthroughs often emerge from veering off course, driven not by protocol, but by the curiosity, intuition, and creativity that define human discovery.
At San Francisco’s fourth annual Deep Tech Week, events like AI Accelerated Science and Ligandal’s Biotech Summit emphasized the important role of agent-aided scientific experimentation in developing next-generation technologies to advance humanity. Yet, with the recent launches of superintelligent scientific agents from companies like FutureHouse and Lila Science, some curious attendees speculated that AI might soon replace scientists altogether.
But the founders, researchers, and operators at Deep Tech Week pushed back on this notion, emphasizing that such a perspective overlooks the criticality of human spirit to experimentation. Agents aren’t replacements — instead, they’re simply tools for accelerating it. As Kristin Schmidt, Strategy Assistant for Accelerated Discovery at IBM, puts it: The emergence of AI agents is analogous to the arrival of computers. Just as computers enhanced what mathematicians could do, AI agents will amplify scientists’ ability to explore, imagine, and stumble upon the next “happy accident.”
But, how exactly will agents do this?
While the preliminary questions and creativity that drive scientific experimentation will likely remain human, hypothesis building and testing has long been one of the most manual and time-consuming aspects of the scientific method.
Constructing a viable hypothesis requires extensive interdisciplinary literature review, careful consideration of extraneous variables, and, most importantly, repeated cycles of trial and error. This iterative process often results in thousands of failures before yielding a single success. Take drug discovery, for example: Before the advent of generative AI, out of every 10,000 compounds initially screened, only about 250 advanced to preclinical testing, and just 1 in 5,000 ultimately received market approval.
AI agents will help researchers generate hypotheses more efficiently in response to their scientific questions. As Theofanis Karaletsos — Head of AI at the Chan Zuckerberg Initiative and Co-Founder of Achira — explains: Embracing agents in experimentation doesn’t mean replacing researchers, it means shifting from knowledge-based hypothesis creation to quantitative hypothesis prompting. AI can’t replicate the foundational curiosity that drives human experimental design, but it can creatively predict hypotheses that inspire new lines of questioning, like novel protein structures or mathematical theories.
Consider the idea of “vibe coding” as a parallel. In half a year, everyone might be writing code via vibe coding. Yet programmers won’t disappear — Instead, as coding copilot platforms become more precise and autonomous, programmers will need to become better at asking questions and writing prompts. The same is true for scientists. AI will not replace them just yet, but empower them to think differently, more creatively, and more quantitatively.
To understand this rather abstract concept, let’s take a look at three examples showcased at Deep Tech Week of how scientific discovery has been accelerated via hypothesis building agents.
1. LLMs as Copilots for Mathematical Theorem Proving
Mathematical theorem proving is the process of showing that a mathematical statement (or theorem) is definitively true by following a rigorous series of logical steps based on known facts. This task is especially difficult due to the abstract nature of mathematics, its dependence on earlier results, and the need for rigorous reasoning. These challenges create an enormous search space for possible proofs — much like the range of hypotheses in scientific discovery. And like science, theorem proving relies not only on logic, but also on human intuition and creativity to guide where to search.
To manage this complexity, many mathematicians use Lean, an interactive theorem prover and programming language. Lean lets users formally state theorems and helps them construct proofs step by step. Still, navigating the vast space of potential proofs can be slow, repetitive, and filled with trial and error.
That’s where LeanCopilot comes in. Developed by Dr. Animashree Anandkumar’s lab at Caltech, LeanCopilot is an AI assistant designed to support — not replace — mathematicians. It speeds up the search for promising proof strategies, handling much of the tedious trial and error while leaving the deeper creative insight to humans. In practice, LeanCopilot has helped complete nearly two-thirds of proofs with minimal human input.
LeanCopilot is a compelling example of how large language models (LLMs) can assist in the search and construction of mathematical proofs. Rather than replacing human reasoning, it frees mathematicians to focus on formulating more ambitious theorems — reminding us that while AI can extend our capabilities, intuition and insight remain distinctly human strengths.
2. Predicting the Structure and Interactions of All of Life’s Molecules
Proteins are the engines of life, performing countless functions in living organisms. Their function depends on their precise 3D shape — but figuring out that shape is no easy task. A single amino acid sequence can fold into an astronomical number of possible conformations. In the 1960s, molecular biologist Cyrus Levinthal famously noted that if a protein tried every possible shape, it would take longer than the age of the universe — a thought experiment now known as Levinthal’s paradox.
Enter AlphaFold, the AI system from Google DeepMind that has revolutionized computational biology. It predicts a protein’s 3D structure from its amino acid sequence in minutes. The latest version, AlphaFold 3, goes even further — accurately modeling not just proteins, but DNA, RNA, ligands, and their interactions. This marks a major leap for drug discovery and our broader understanding of biology.
Still, each prediction is just that — a hypothesis. While AlphaFold’s insights are deeply data-driven and vastly more efficient than human computation, their real power lies in how scientists use them. It takes expertise and curiosity to ask the right questions and interpret the results in ways that drive meaningful breakthroughs.
3. Collaboration with Virtual Lab Teams
One of the clearest examples of human-led, agentic scientific experimentation comes from Dr. James Zou’s Virtual Lab at Stanford. This platform brings together interdisciplinary AI agents that collaborate with human scientists to tackle complex, open-ended R&D challenges.
The Virtual Lab operates like a real research environment — just faster and at scale. A principal investigator (PI) agent kicks off each project by posing a scientific objective. Specialized agents — such as immunologists, machine learning experts, computational biologists, and scientific critics — then respond, propose ideas, and debate next steps. Multiple agent teams can run in parallel, generating a rich pool of hypotheses. A “professor agent” reviews their transcripts and flags promising directions for deeper exploration. Throughout the process, human scientists serve as high-level advisors, steering and validating the agents’ work.
To ensure continuous improvement, Dr. Zou has also built “virtual lab schools” where agents train on a range of scientific problems. The results have been striking. One agent, CellVoyager, helped uncover that CD8 T cells in COVID-19 patients were primed for inflammatory cell death. It also surfaced signs of transcriptional noise linked to aging in the brain’s subventricular zone — offering new clues into cellular aging.
Zou’s work highlights a critical point: agentic AI doesn’t replace researchers — it augments them. The Virtual Lab accelerates the cycle of hypothesis generation and validation, but it’s human curiosity and scientific judgment that guide the questions — and ultimately, the discoveries.
So, what does this mean for the future of scientists?
Each of these examples underscores a central truth: while AI agents are getting better at generating and testing hypotheses, there’s still something uniquely valuable about human-led science — rooted in our curiosity, creativity, and the beautifully imperfect way we explore the unknown. It’s something agents haven’t yet — and may never — replicate.
Just as we once embraced the telephone, computer, and smartphone, scientists must now learn to work alongside agents — not just to speed up their workflows, but to ask more ambitious questions. By offloading routine tasks and cutting down time constraints, agents give researchers space to think more creatively, take bolder risks, and stumble into those serendipitous “happy accidents” that often lead to breakthroughs.
The question isn’t whether AI will replace scientists, it’s whether scientists are using agents to ask the right questions.
Reach out if you’re redefining scientific discovery.
At Alumni Ventures, we view the agentic revolution in scientific discovery as a chance to redefine the pace and potential of research. The three examples we’ve highlighted represent just a few of the many entry points into this expanding ecosystem.
We’re especially excited to meet founders building agents that accelerate discovery, whether through hypothesis-generating models, scientific foundation models, or virtual lab environments. If you’re a founder or operator in this space, or a scientist applying these technologies in research, we’d love to connect.
About the Authors
Meera Oak
Partner, Seed FundMeera’s background includes strategic, financial, and operational experience from her time at Yale University, where she managed a $1B budget (of a $4B organization), led M&A transactions, and secured business development relationships with corporate partners. Most recently, she worked with early-stage venture funds and incubators like Create Venture Studio and Polymath Capital Partners and was responsible for launching business ventures and sourcing investments in enterprise SaaS, infrastructure, and ecommerce. Meera has a BA in Economics from Swarthmore College and an MBA from the Tuck School of Business at Dartmouth.
Anika Chebrolu
Summer AssociateAnika Chebrolu is a Summer Associate at Alumni Ventures, where she focuses on pre-seed and seed-stage investing. She is also the Managing Director at Moso Capital and a Venture Fellow at Arkitekt Ventures, with a deep interest in health, bio, and climate tech. With experience spanning private equity, scientific research, and grassroots fundraising, Anika brings a multidisciplinary lens to venture investing, grounded in both analytical rigor and a passion for impact. She is currently pursuing her undergraduate studies at the University of Pennsylvania, where she combines her interests in business, science, and social impact through hands-on roles in early-stage consulting and impact investing.
Resources:
Deep Tech Week 2025, San Francisco
LeanCopilot Paper
AlphaFold 3
Dr. Zou’s “Virtual Lab”
Invest In Our Funds
Let our expert team handle it.
Our award-winning investment professionals will build a diversified portfolio of venture-backed companies.
Invest Deal-By-Deal
You make your own investment decisions. You’ll see 1-2 deals per month, including all AV diligence materials, and select which companies you want to invest in.
This communication is from Alumni Ventures, a for-profit venture capital company that is not affiliated with or endorsed by any school. It is not personalized advice, and AV only provides advice to its client funds. This communication is neither an offer to sell, nor a solicitation of an offer to purchase, any security. Such offers are made only pursuant to the formal offering documents for the fund(s) concerned, and describe significant risks and other material information that should be carefully considered before investing. For additional information, please see here. Example portfolio companies are provided for illustrative purposes only and are not necessarily indicative of any AV fund or the outcomes experienced by any investor. Example portfolio companies shown are not available to future investors, except potentially in the case of follow-on investments. Venture capital investing involves substantial risk, including risk of loss of all capital invested. This communication includes forward-looking statements, generally consisting of any statement pertaining to any issue other than historical fact, including without limitation predictions, financial projections, the anticipated results of the execution of any plan or strategy, the expectation or belief of the speaker, or other events or circumstances to exist in the future. Forward-looking statements are not representations of actual fact, depend on certain assumptions that may not be realized, and are not guaranteed to occur. Any forward-looking statements included in this communication speak only as of the date of the communication. AV and its affiliates disclaim any obligation to update, amend, or alter such forward-looking statements, whether due to subsequent events, new information, or otherwise.