Last week, Sandeep Pawar, Manbir Sandhu, and I attended the Precision Medicine World Conference (PMWC) in Silicon Valley. Conferences like this are a bit of a forcing function. You step away from the daily grind, talk to a lot of smart people, and come back with your thinking either sharper or scrambled. This time it was mostly the former. Here are the three things that stuck with me.

1. It’s Still Early Innings

It's easy to watch Anthropic, OpenAI, and the frontier labs and feel like the race is already won. The models are remarkable. The rate of improvement is genuinely hard to comprehend. But the framing of "it's already over" is exactly wrong.

The agentic AI revolution is just beginning. Yes, the foundation models are here. But deploying them in ways that actually change how scientists work, how data gets used, how organizations collaborate across institutional boundaries… that work is almost entirely ahead of us. There are enormous spaces to build in and enormous problems to solve.

For us, this means bringing these capabilities home to life sciences:  Tailoring them, grounding them in the domain, and making them work for how researchers and clinical teams actually operate. The frontier labs are building the engines. The opportunity in life sciences is to build the car.

2. Clinical Foundation Models Are Actually Happening

Two highlights at the conference pointed at the same exciting development: clinical foundation models are no longer theoretical.

Nigam Shah's talk on chat EHR was a standout. The core idea is elegant: take the sequence of events in a patient's timeline (lab draws, diagnoses, procedures, surgeries) and treat it as a language modeling problem. Instead of next-token prediction, you're doing next-event prediction. Train on enough of these timelines and you get a model that can be decoded toward almost any downstream clinical question: predicting mortality, relapse, readmission, disease progression. A single encoder produces an embedding space that flexible decoding heads can target in almost any direction. It's a powerful architecture.

Conversations with the Standard Model Bio team pointed at the same thing from a different angle. And in pathology, the work coming out of MSK (the Virchow model and now the Prism model) shows what becomes possible when you can read slides at scale, extract rich features, and plug them into predictive workflows.

The throughline: the data is finally being amassed at the scale needed to train these models. The measurement infrastructure that makes this possible (e.g., NGS cost curves continuing to fall, 10x Genomics pushing hard to bring single-cell omics down to the cost of bulk) is advancing rapidly. The principle is simple: if you can't measure it, you can't improve it. And the tools to measure biology are getting dramatically better and cheaper. That's the foundation everything else is being built on.

This matters to us directly. The richer the measurement infrastructure, the richer the multimodal datasets that get generated. And the richer those datasets, the more important it becomes to have infrastructure that makes them accessible, analyzable, and shareable. That’s exactly the problem Manifold is built to solve. Our recent work with the Komen Tissue Bank is a good example: as multimodal data assets become richer and more complex, the need for purpose-built platforms to unlock them for research communities becomes more acute, not less.

3. With All the Advances in AI, Staying Close to Customers Still Matters Most

One of the most energizing parts of the conference wasn't a keynote. It was the conversations. Getting to think through problems alongside visitors to our booth cemented our thinking around what’s needed today, what it means to onboard data as a producer, and where the friction really lives in the workflows that we support.

Here's the thing: amid all the genuinely exciting advances in technology and AI, staying close to customers and end users is still the thing that matters most. It's easy to get caught up in the momentum of the field and start building toward the technology rather than toward the problem. The antidote is exactly what conferences like PMWC enable: face-to-face time with the people doing the work, getting a clear signal on where the real pain is.

This is something we think about a lot at Manifold. In our Series B announcement, we wrote about the Life Sciences Chasm:  the gap between expert intent and technical execution that slows the entire field down. That chasm is real, and it's specific. It's not an abstract technology problem. It lives in the workflows of the scientists and data teams we work with every day, in the friction between how questions get asked and how answers get generated. Closing it requires staying deeply connected to what those teams actually experience.

The conversations at PMWC sharpened our thinking on exactly where that friction is highest and where the next version of our platform needs to go. That's the kind of signal no amount of market research can substitute for.

The field is moving fast. Foundation models for medicine are being laid in real time. The measurement tools are getting better and cheaper at a rate that should excite everyone working in this space. And the agentic AI revolution is just beginning. There's enormous room to build things that matter.

And we’re going to build them.