Twenty-five years and half a mile
The morning before my first day I went for a run through Chicago, and a couple of miles in I recognised the route.
The morning before my first day at Tempus I went for a run through Chicago.
A couple of miles in, I recognised the route. Not the way you recognise a place in a photograph. The way your legs recognise a hill before you do. I had run this before.
I had spent the previous six weeks wrapping up thirteen years at AstraZeneca and being with my family, and in all that time I never once looked up where in Chicago the office actually was. So I ended up standing on a sidewalk at seven in the morning doing arithmetic I had not planned on. It came out at twenty-five years and half a mile.
That is the distance from the web applications company where I interned in college to the building I walked into a few hours later as a VP.
I got the goosebumps. Then I finished the run.
The under-construction sign
I was working at a web company as a biology major, which takes some explaining.
I had been building websites since junior high. This was the animated GIF era, the era of the under-construction sign and the hit counter you quietly started at a thousand. GeoCities was where I lived. Nobody taught me any of it, because there was nobody to do the teaching. The web was about four years old and the people building it were figuring it out in public, badly, at the same time as me.
That is the part worth holding onto, because it is the part that repeats. I was not precocious. I was early. Those are different things, and only one of them is available to you.
Four times
I have now done the same thing four times.
The web in the mid-nineties. Genomics in the early two-thousands. Clinical sequencing around 2010. Agentic AI from about 2023. Each time I arrived while the field was still arguing about what it was, and each time I got in without the credential the people around me had.
That is not a claim about talent. It is a claim about timing.
I was not precocious. I was early. Those are different things.
Genomics, before there was software
I joined The Institute for Genomic Research in 2002. It later became the J. Craig Venter Institute, and I was there eight years.
The human genome had a draft and no finish. The word metagenomics had not reached most biology departments. What that meant day to day is that the software to assemble and annotate a genome at that scale did not exist as a thing you could buy, so you wrote it, and the version you wrote was wrong until enough real data had run through it to show you where.
Nobody at TIGR ever asked what my degree was in. They asked whether the pipeline ran.
In 2010 the group built a synthetic Mycoplasma chromosome, transplanted it into a recipient cell, and the cell ran on it. They wrote four watermark sequences into the DNA, using a lexicon that maps DNA triplets to letters of the alphabet. One of those watermarks spells out the names of forty-six people who worked on it.
My initials are in there. Somewhere there is an organism carrying my name inside its chromosome, which is the closest thing to tenure I am ever going to get.
I helped computationally assemble the first human genome, the first large-scale metagenome, and the first synthetic genome ever made. I mention all three together because of what they have in common, which is that none of them had a method until somebody wrote one down afterwards.
The startup, and the boring half
I left for EdgeBio in 2010 to build the sequencing and bioinformatic interpretation platform for a CLIA lab. Clinical-grade genomics was maybe three years old as a commercial proposition. There was no reference architecture, so we built one, and most of what we built was wrong in ways we only discovered by running it against real samples.
GeneDx acquired the services business in 2013.
Around the same stretch I was running a small web and digital marketing company of my own, Inspiring Design, alongside the day job. It taught me the parts of building a company that nobody finds interesting until the moment they need them. How to form an entity. How to write a contract that survives contact with a disagreement. How to invoice, and what to do when the invoice does not get paid.
I have used that more often than I have used anything I learned in a lecture hall. If you are technical and you have never had to be the person who signs something, go and be that person once. It rewires how you read every proposal you see afterwards.
The fourth window was not a young field
Then thirteen years at AstraZeneca. I joined to work on genomics infrastructure and left running the AI Centre of Excellence for R&D.
The fourth window looked different from the first three. AI was not a young field by 2023. The papers existed, the tooling existed, and plenty of people understood it better than I did. What was young was the fit between that field and a hundred-and-fifty-year-old pharmaceutical company, and that gap turns out to open the same kind of opportunity as an empty field does.
So the work was FAIR data foundations, a prioritisation layer, agentic systems that did real research tasks, and a governance pathway that took model provisioning from weeks to hours. That last one mattered most and got the least attention, which is normal. I ended up repeating a line often enough that people started saying it back to me: safe is the fast way.
None of those four platforms was assigned to me. I wrote the first version of each one myself, proved it worked, recruited people better than me, and handed it over. That is the only method I have.
What the degree cost
I have a bachelor's degree in biology. That is the whole credential.
Almost everyone holding a job like the one I have now holds a doctorate. This is the part of the essay where I am supposed to tell you it did not matter, and that would be a lie, so here is the honest version.
It cost me the benefit of the doubt. In a scientific room, a person with no letters after their name starts a notch below the line and has to climb over it by being useful. Every time. In every new room.
It cost me speed early on, because there are things a good advisor hands you in a year that took me four to learn by walking into them. And there are rooms where the credential is the entry condition and no amount of shipped work substitutes for it. I have been outside a few of those.
What it did not cost me was the work. Not once in twenty-five years has anyone stopped me building something because of what my degree said. The gate is on the room, not on the workbench.
When the window is open
Here is what those four moments have in common.
A credential certifies that you have mastered an existing body of knowledge. That is a genuinely useful thing for it to do, and I would want a certified surgeon. But when the body of knowledge does not exist yet, there is nothing available to certify, and the only evidence anybody can offer is a thing that works.
That is the window. It opens when a field is young enough that nobody has written the textbook, and it closes when somebody does. While it is open, building is the credential. After it closes, the credential is the credential, and getting in the way you got in before stops working.
The practical version, for anyone deciding whether to wait for permission: the window is not announced and it does not stay open long. In the web it was maybe six years. In clinical genomics, four.
You find out it closed when the job postings start asking for five years of experience in a thing that has existed for five years.
The part where I am supposed to talk about hype
I have now watched three technologies get badly oversold from the inside.
The web was going to replace retail by 2001. Genomics was going to give us personalised medicine within a decade of the genome being published. AI is currently going to do most of what you have read this year that it is going to do.
All three were oversold on timing. All three were roughly right on direction. The mistake people make is treating the collapse of the hype as a verdict on the technology, when the hype collapsing is just the money leaving before the infrastructure is finished.
For AI right now, the honest answer is that some of it will leave residue and some will not, and I do not think anyone can reliably sort them from inside the cycle. What I do think is that the residue is where the products are, and that finding out which is which is a measurement problem rather than an opinion problem.
Which is a reasonable description of the job I just took.
The address changed
Tempus is the first place I have worked where the data is not the constraint. Molecular, imaging and clinical outcomes on the same patients, at a size a model can learn something from. I have spent twenty-five years running into the absence of that, and I would like to find out what is on the other side of it.
People keep telling me this is a full-circle story. I do not think it is. A circle means you came back to where you started, and I never left. I have had one job for twenty-five years, which is to turn up somewhere the rules have not been written yet and start writing them badly until they get better.
The address changed. And that's about it.
I am still building, and still writing about it.