Vijay Pande spent years placing big bets inside a $4 billion practice at a16z. Now he runs a much smaller, AI-native fund called VZVC, and he is deliberately not trying to play the same game. His point is not that large funds are wrong. It is that the old playbook of spreading capital across dozens of bets assumes a certain kind of predictability, and biology no longer fits that mold. When he says, "We're not doing 30 bets a year," he is making a claim about focus, not just size. For anyone building with AI, that distinction matters more than the headline number.
The deeper shift Pande is pointing to is one we are seeing play out across technical fields: the move from discovery to engineering. Biology, he argues, has long been a science of finding things by accident or by exhausting search. But with AI, it becomes a discipline of designing solutions on purpose. That is not a subtle change in workflow. It changes what you optimize for, how you hire, and which problems are worth solving. We made a similar observation in our own work on Exploring Real-World Computer Vision, where the hard part is rarely the model architecture and almost always the messy, real-world constraints of deployment. The same logic applies to biology: the bottleneck is not the hypothesis, it is the pipeline that tests it. And that is why Pande's emphasis on open, shared datasets is so important. Walled-off data may protect a single company's lead for a quarter, but it slows down the collective learning that makes AI genuinely useful. We have seen the cost of that fragmentation in other domains, and the pattern is consistent: closed systems accelerate short-term wins and stall long-term progress.
That is also why his skepticism about clinical trials is worth taking seriously. He is not saying trials are unnecessary. He is saying they are brutally expensive, and that cost distorts every decision upstream of them. If you know the trial will eat your entire budget, you stop taking risks. You optimize for incremental improvements instead of transformative ones. That is a structural constraint, not a personal failing. We touched on a similar theme in our piece on Verify Your AI's Understanding, where we argued that verification is not a bureaucratic step but a design principle. The same holds here: if you cannot afford to test your assumptions cheaply, you will never test the bold ones. Pande's bet is that AI can change that equation, not by making trials cheaper at the margins, but by making the entire discovery process more efficient from the start. That is a different kind of ambition, and it is one that smaller, more focused funds are better positioned to pursue.
What should a reader take from this? The practical takeaway is simple: the size of your bet should match the shape of the problem, not the size of your ego or your fund. Pande ran billions and walked away. That is not a criticism of capital; it is a recognition that capital without alignment is just overhead. For founders and builders in AI, the question is not whether you can raise a large round. It is whether you have designed your process to learn faster than your competitors, and whether you are willing to share enough to make the whole field move forward. The specific thing to watch is whether open datasets actually gain traction in biology the way they did in early AI research. If they do, the cost of entry drops, and the advantage shifts from whoever hoards the most data to whoever asks the best questions. If they do not, we will keep paying for trials that tell us what we already suspected, one expensive step at a time.