Humans have a source code too

Yesterday I discovered that a food I had in my freezer could kill me. Not in a vague, long-term, “red meat might give you cancer” kind of way. More like: skin turns yellow, urine goes dark, dead in two days. I found this because I gave AI agents my genome and let them run on a GPU cluster for 48 hours.

Let me back up.

Sequencing

In 2024 I decided to sequence my DNA. The technology seemed viable and I figured the sooner I did it, the more benefits would compound (acting early against genetic weaknesses). It took over a year to get my results, after sending 4 vials to 3 different companies, all under aliases.

Once I confirmed it was actually my DNA (blood type match, etc.), I had some fun querying the source files with easy questions: do I carry this Parkinson’s mutation, how fast do I metabolize caffeine, and so on. GPT suggested many interesting queries but I was disappointed. Most insights were tips that apply to everyone: sleep properly, eat properly, exercise. What I wanted was something far more advanced: identifying synergies between alleles, cross-referencing my drug sensitivities to recommend precise habit changes, maybe even molecules that would work particularly well given my genetic design. Given the complexity of the human genome, a static, generalist set of queries was doomed to failure. I needed something that could refine itself based on its own results.

But recently I regained hope.

A Paradigm Shift?

With the release of Opus 4.5, my usage of AI changed dramatically. Instead of using AI as a copilot, agents became the main drivers. I would watch them work like an air traffic controller watching dots crawl across a screen. My computer quickly became a limiting factor, agents would spawn a debugging browser in the middle of my Netflix episode, so I built an orchestrator to run them remotely for days and leave my laptop in peace. If you’re curious, it’s open source: sandboxed.sh.

What surprised me was how long these agents could work autonomously. Well-designed harnesses, suitable tools, and smarter models allowed them to verify their own work and keep going, recalibrating automatically. The first real moment of clarity came when I realized I was the bottleneck.

I asked an agent to update a Minecraft plugin, my favorite example of a complex task, poorly documented, requiring many skills. It needed my feedback to verify its work: would the new items render correctly, would the interactions behave as expected. The fix was obvious in retrospect: build a tool to replace me. A modified version of the game the agent could control via command line, testing all its changes autonomously.

That got me thinking. Maybe I could do the same with DNA.

I was ready to hand over the tools and turn myself into a self-experimenting lab rat. But I had no desire to let OpenAI or anyone else train on my most personal data. Fortunately, my friend Ben and I recently invested in a little toy from Nvidia: the DGX Spark. Slow for inference, but very capable of loading large models. It could run for days without bothering me. I’d already been waiting two years.

Methodology

I needed a plan, so I did what any reasonable person would do in 2026: I asked AGI.

Opus managed the hardware constraints of the Spark and built a router to swap between LLM models depending on the request, in exchange for a 15-second loading time per transition. He created a repo for fellow Nvidia Blackwell-at-home users: dgx-spark-router.

The very slow but talented GPT 5.2 Pro came back with a plan: split the work in two.

1. Set up the environment, tooling, and strategy with top-tier models (Codex and Opus)
2. Upload my DNA and task the local LLMs to run the pipeline

I ended up with a bunch of Python scripts and dozens of gigabytes of scientific literature. Local models being considerably less capable, I used oh-my-opencode 3 for the final task, a harness that treats models as lazy, stupid liars and supervises them accordingly. After some trial and error, this configuration seemed to work.

Trading kWh for Life Points

Moment of truth. I pressed enter and heard the soft noise of the chip getting to work.

After hours of waiting, the initial results came back with a few problems:

• False positives from X chromosome artifacts.
• Generic advice I could have found on WebMD.
• A fixation on everything that could go wrong.

It took me two days of back and forth to refine the research (mostly me waiting) but it was worth it. I got a detailed report of findings and implications, plus a concise version accounting for the compound effect of characteristics and comparing my statistics to the average population.

And damn, it was a gold mine.

For the equivalent of what a Tesla Model 3 consumes in 20 kilometers, I had replaced a team of researchers.

And among other things, it saved my life.

What I Found

I won’t share everything, but here are three things.

1. I am cursed with alcohol. My metabolism metabolizes it twice as fast as average, I am literally designed to love it… but life is cruel and I have between x10 and x50 the risk of developing pancreatitis if I drink so I should avoid it as much as possible. This happened twice to someone from my close family, this is probably why.

2. I have “G6PD Deficiency”. It’s benign unless I take a few rare medications I have never heard of or …eat fava beans. If I ate a bit too much it would cause a rapid destruction of my red blood cells which my liver couldn’t process. In two days my skin would turn yellow, my urine dark and I could die. Guess what I liked to put in my soups but I will never eat again?

3. On the good side I have a variant of the NOS3 gene (chr7:150992991) with huge cardiovascular, athletic and metabolic benefits which I can maximize by increasing my nitrate intake, avoiding certain kinds of mouthwashes (wtf?) and supplementing with Methylfolate + Methylcobalamin.

I’ve scheduled tests to confirm the results. In the meantime, I’ve stopped eating fava beans.

If you want to reproduce this, be careful with the data. And most importantly, be sure you want to know the answers to the questions you’re about to ask.

Take care,

Thomas