I'm building a biodefense company

It's been nine months since I last wrote here. It has been hard, but more fulfilling than anything I've done before. In January I left my job at a venture capital firm to run a startup building what I believe is one of the most consequential technologies of our time, and this post is a preview of what we're building and why.

The short version is that we're finally able to read the chemistry floating around us. Almost everything gives off volatile molecules, and those molecules carry information we mostly let blow away: whether there's a gun in the room, whether a person's body chemistry has shifted directionally towards disease, whether one tree in an orchard is patient zero for an outbreak that could take down a state's entire citrus industry.

I think of it as dark matter. It's all around us, it's full of signal, and almost nothing measures it.

No technology today can give you those answers instantly and at scale. The best we have are what I call "fancy smoke detectors." Each is built for a single job, and even then most aren't sensitive enough or fast enough to be operationally useful: about a minute to an answer at best, and usually days once you count the sample preparation.

We want to build the first technology that can answer almost any of those questions in under a second. It will serve several markets eventually, and the biggest vision is a world where changes in a person's chemistry get noticed early enough to send them for a proper test.

Our first market, though, is national security and biodefense, which is what this post is about.

Suddenly everyone is paying attention to bioterrorism

Biodefense is having its day in the sun. In March the OpenAI Foundation pledged at least $1 billion over the next year, aimed in part at novel biological threats. Dario Amodei and other leaders of frontier AI labs now put biological risk on the same list as cyberattacks and misaligned AI. Anthropic's September 2026 threat report describes several attempts to use its models for pathogen engineering, including one by a state-associated infectious disease lab trying to quietly modify a smallpox-like virus. Those are only the attempts that got caught.

The press has noticed too. On September 23 the Wall Street Journal ran "Anthropic Leaders Back Startup Developing System to Detect Biological Threats", and five days later it profiled Red Queen Bio in "This Startup Is Using AI to Fight Off a Future AI Pandemic". A lot of people sent me the first one. This post is partly my answer to all of them.

Fear travels faster than the pathogen

This summer I read Annie Jacobsen's Biological War: A Scenario, and my takeaway was that the fear outruns the disease, and the fear does the lasting damage. It reminded me of an xkcd comic from 2010 about how tweets about an earthquake can reach people faster than the seismic waves do. A bioweapon works the same way. The news, the rumors and the panic arrive well before the virus does, and they do their own harm. Someone who used to work for the government's biological security program said to me "bio is the bomb that doesn't go boom."

Trust in each other and in our institutions would erode quickly. Resource panic would make the toilet paper runs of 2020 look quaint. Class resentments that already exist would harden into open conflict. Even if the biological event were short, the damage to the way we live would be chronic.

We've already seen a mild version of this with COVID. It wore down trust in our scientific institutions, and then in government, and then the very idea that we could believe in the same sources of information. We argued about which people, which outlets and which institutions to believe. I've come to think the real "long COVID" was civic.

The doubters have a point

Plenty of serious people think this is overblown. Some argue that worry about AI-enabled bioweapons is doomerism, and that the real bottleneck to making a pathogen is physical: lab space, equipment, materials and skill, not a model that can design sequences. I think they're right about the bottleneck. I don't think it changes what we should do.

The first reason is that prevention leaks. Open models get jailbroken, and covering every step of the path to a bioweapon would take a degree of global coordination we're nowhere near. The United States isn't the only place you can order DNA or set up a lab, and a motivated group can work out the rest. I assume that something will get out eventually.

Just a week ago, a 28 year old scientist died from a suspected plague at the Russian "anti-plague laboratory" in Irkutsk, and while we don't have to worry about this one, this is just the beginning.

The second reason is the asymmetric risk. A bioweapon is an existential threat with very little margin for error, and physical resilience is much harder to build than digital resilience. You can roll back a bad software update, but you can't roll back an outbreak. When the cost of being wrong is that lopsided, waiting for certainty is its own bet.

So the question I keep coming back to is this: if something does get through, how do we widen the margin for error? How do we buy time for the countermeasures to arrive?

The missing layer

It helps to think of biosecurity as a stack of four layers:

1. Prevention - to stop bioweapons from being made at all
2. Alarm - detect a threat at the border, air or wastewater
3. Detection - rapid diagnostics so that some parts of society can stay open
4. Neutralization - vaccines, treatments, and other medical countermeasures

All four matter, and a real defense needs every one of them. The layer we don't have is the third: slowing the spread while keeping society open. That's what we work on.

People often assume the second layer can cover for it. Alarms that sample the air or the wastewater are useful, but they work at the level of a population. They can tell you that something is in the air of a terminal or the sewage of a city, not the person who is carrying it. Sequencing takes time, so by the time the result comes back, the person who set it off is already in an Uber. And if all you know is that a sample came back positive, the only response available is to quarantine everyone it represents.

COVID made this concrete. We had the genome of the new coronavirus by January 11, 2020. It still took months to turn that into a test that could scale, and the early swabs went in so deep that people called them brain swabs. Meanwhile the school principal, the office manager and the person running the nursing home had no way to answer the question they actually cared about, which was whether it was safe to be in this room right now. They were left with the bluntest instrument there is: close everything, just in case. We also learned that Americans don't stay locked in their homes for long.

The company I want to build

The company I want to build keeps detection capacity deployed for non-emergency work, but ready to be redeployed in a national emergency. I think this is the holy grail of a dual-use company. Not making a consumer version plus a ruggedized camouflage "milspec" version. True dual-use means that the capex is used by consumers and military (like satellites).

Biodefense money follows fear. It surges during an outbreak and drains away between them, because nobody pays for prevention. So I hold the company to four rules.

1. The revenue has to survive the gaps between outbreaks. It should be wavy, not spiky.
2. It has to slow the slide toward collapse in ways that are plausible and actionable. What we do with the information matters more than the information itself.
3. It has to work within the next few years (preferably now), not the next few decades.
4. It has to build defense without building offense. Nothing we make should help anyone make a weapon.

The fourth rule comes from the asymmetry at the heart of all this. An attacker has to succeed once, and a defender has to catch everything. The best lever a defender has is speed: the faster you can tell who is carrying something, the less the attacker's head start is worth. Better detection pulls that lever without handing anything useful to the other side.

The third rule is why we didn't start from scratch. The fastest way to have something that works in a few years is to take the best sensor we already have, one that works better than anything we've ever engineered, and then supercharge it and scale it up.

I'm sorry I'm so vague about what we're building right now. I promise I'll write more about what we're building next time. I just want the time and space to write a manifesto around it instead of burying it in this blog.

Buying time

If I'm right about all this, the most useful thing biodefense can give us is time: enough of it to keep schools, offices and hospitals running while scientists do the slow work of finding a countermeasure. Prevention buys some of that time, and alarms buy a little more. I think instantaneous detection buys the most, and I think that there is a way to build this without having to spend billions of dollars on a very specific biodefense use case.

That's what we're building.

If you work on any layer of this stack, or you think I'm wrong about which layer matters most, I'd like to hear from you.