How GitLab's Founder Used AI to Fight His Cancer

At four in the morning in late 2022, Sid Sijbrandij sat in an emergency room because a pain near his heart had refused to go away for two weeks. The doctors reassured him: there is no such thing as a two-week heart attack. They took an X-ray and sent him home.

A few hours later his GP called and opened with a strange question: do you know how to meditate? Sid's blood pressure was dangerously high and the pain might be his aorta starting to tear. Back in the emergency room, the scans showed his aorta was fine. They also showed a six-centimeter tumor growing from a vertebra in his upper spine. The diagnosis, in November 2022: osteosarcoma, a rare and aggressive bone cancer. Sid was the co-founder and CEO of GitLab, and he had taken the company public the year before.

What followed was the standard playbook, and it was brutal. Surgeons removed the tumor in December 2022 and fused his spine with a metal frame. Then came radiation and chemotherapy so aggressive that he needed four blood transfusions. Sid calls the standard treatment medieval, and he says it as someone grateful that it exists.

What happens when the standard of care runs out?

Osteosarcoma is rare. In the United States it is diagnosed roughly 1,000 times a year, about half of them in children and adolescents. When it comes back after first-line treatment, the prognosis is often serious and the remaining options are few. That is exactly what happened to Sid: in June 2024 the cancer came back, at the Th4 vertebra in his spine, and in January 2025 it progressed again.

His oncologist had no standard medicines left to recommend and told him to look for clinical trials. For a disease this rare, there were none. Most patients hear that sentence at the weakest moment of their lives, and the system offers them little beyond it. Sid decided to go at his cancer the way he had built GitLab: full founder mode, as he calls it. Understand every detail, own every decision, hand off nothing that decides your life. He even quit his day job to stay fully focused.

What does founder mode against cancer look like?

Three decisions define it. First, maximal diagnostics. Instead of testing only what would change the next treatment decision, Sid's team ran every technology they could get access to: whole-genome and whole-exome sequencing, bulk and single-cell RNA sequencing, long-read sequencing, spatial transcriptomics, multiplex tissue imaging, nine whole-body PET/CT scans plus experimental tracers. The result is about 25 terabytes of data, and all of it is public at osteosarc.com so that researchers anywhere can work with it.

Second, treatments in parallel. Most cancer patients try one therapy, wait, and try the next when it fails. With an aggressive tumor, that waiting is what kills. Sid's team prepared several options at the same time, so that a working plan B already existed before plan A failed.

Third, making medicine where none exists. Through a friend's biotech startup, Sid learned that the US drug authority FDA allows treatments manufactured for one single patient, a pathway called single patient IND, and that it approves nearly all such requests. In Germany and the EU, related routes exist, such as compassionate use programs and the individual treatment attempt. It turned an assumption upside down: you do not have to wait for a big trial when you are out of options.

How did AI actually help?

The person running what Sid calls the enterprise of his care is not a doctor. Jacob Stern is a geneticist who spent six years at the sequencing company 10x Genomics. In the talk the two gave at the OpenAI Forum in March 2026, Stern walked through his real ChatGPT history to show what the AI contributed.

The first example is almost banal. In the summer of 2025 he uploaded the raw CSV file of a bulk RNA sequencing run, gene names and counts, and asked a paid ChatGPT model what it saw. The answer flagged a surface protein called B7-H3 as conspicuously overexpressed. That protein later became the target of one of Sid's most important planned therapies.

Today the setup goes much further. The team built a system in which a question in plain language spins up agents that search the literature, form a hypothesis, write analysis code and run it against roughly 600,000 single cells sequenced from Sid's blood. When a lab result hinted at a dangerous late effect of his old chemotherapy, Stern asked the system, waited half an hour and got back an analysis with plots, conclusions and the Python code it had written. Specialists then checked the finding properly and gave the all-clear.

Stern is explicit about the limits. He does not trust the output blindly, and it has not made him a specialist. It has made him a competent counterpart: someone who can ask experts reasonable questions, understand their answers and push the programs forward while owning a single objective, keeping Sid alive.

Which treatments came out of it?

The clearest success so far began with single-cell sequencing. It showed that Sid's cancer cells carry large amounts of FAP, a protein typical of fibrous tissue. A doctor in Germany offered an experimental therapy that couples a FAP-binding molecule to a radioactive payload. Sid flew in twice, and each time the treatment ended with a stay in an isolation ward. The result: 60 percent of the tumor tissue died, the tumor shrank by a fifth and detached from the membrane around his spinal cord, and in April 2025 surgeons at Memorial Sloan Kettering in New York could remove it. Since then there has been no evidence of disease, a status Sid still reported in mid-2026.

The rest of the arsenal is insurance, built in parallel for the day the cancer might return. A personalized mRNA vaccine, built like a COVID shot but encoding mutations specific to Sid's tumor, went from project start to injection in six months. A TCR T-cell therapy is in preparation: immune cells fitted with receptors fished out of Sid's own sequencing data. And a CAR-T therapy (immune cells re-engineered in the lab to hunt cells that carry one specific mark) targets B7-H3, the protein the AI had flagged, as what Sid calls his nuclear option.

The CAR-T story shows why the diagnostics matter. To check where B7-H3 appears in his body, Sid traveled to Beijing in October 2025 for an experimental scan. The scan found no cancer anywhere. It also showed his liver glowing three and a half times brighter than in the twenty people scanned before him. Translated: his healthy liver also carries plenty of B7-H3, so a CAR-T aimed at it might have destroyed the liver along with the cancer. The fix came from the data as well: the therapy is now built like a double lock (engineers would say a logical AND gate). It only kills cells that carry both marks at once, B7-H3 and FAP, and since FAP is nearly absent in the liver, the liver stays safe.

Two more threads run in the background. Sid's tumor expresses extreme levels of MDM2, a protein for which drugs were developed years ago and then abandoned because the market seemed too small. Sid now pays to keep the manufacturer's freezers running and is looking for a way to bring the compound to market after all. The team also found a barely studied protein that appears around ten thousand times more often in his tumor than in his healthy tissue. Because it repels water, standard water-based assays keep missing it, which may explain why nobody had published on it. Sid's point: with 25 terabytes of data, AI has more patience than any human, which is how details like this stop slipping through. The team is now trying to engineer a molecule that docks onto this protein and turns it into a drug target.

What can you take from this without a founder's bank account?

The obvious objection: Sid is the absolute exception. He is wealthy, extremely well connected, and much of the above is expensive. He addresses it himself. Bulk RNA sequencing of a tumor sample costs about 50 dollars today, a whole genome starts around 500 dollars, and the AI tools he considers essential cost 20 dollars a month. Many drugs worth discussing are cheap generics. The expensive part of his playbook is the custom-made biologics, and even there his team is starting companies to push the costs down and pave the road for the next patients.

The deeper lesson is about advocacy. At his lowest point, a radiology report said his lungs were full of metastases, inoperable. Six of seven doctors accepted the finding. The seventh noticed that the pattern did not match how osteosarcoma spreads, and he was right: it was residue from a COVID infection. Sid does not read the episode as doctors failing. He reads it as proof that a patient who gathers data, asks for second opinions and prepares a differential diagnosis question with AI can be the difference between giving up and continuing.

He is also blunt about incentives. A doctor, he argues, is structurally pushed to minimize liability, while a patient with a lethal disease wants to maximize survival, and those two goals do not always point at the same treatment. He advises nobody to treat themselves. What he does advise is coming to your oncologist well informed, asking why not about combinations and understanding that reasoning about side-effect profiles is possible even where no hundred-million-dollar trial exists. If this tension interests you, we have written about why medicine still reacts too late.

The playbook is already traveling. Scott McKinney, an OpenAI researcher who has fought the same disease since 2020 and who introduced the talk, says the approach Sid and Jacob shared with him changed how he handles his own case. Sid has since hired a full-time person whose only job is helping other patients who reach out.

None of this is medical advice, and Sid's case is a single patient whose outcome cannot be attributed to any one intervention with certainty. He says so himself: maybe he got lucky. What makes his case worth telling is the approach: a technologist using the tools that exist today on a disease the system had given up on, in the hope that other patients benefit later. Several things his team built by hand, personalized vaccines designed with model support, scans before drugs, AI that reads your tumor data, are what he believes will be standard care decades from now. He is trying to pull them into the present.

Where can you follow the story?

Everything is open. The site osteosarc.com hosts the 25 terabytes, the interactive treatment timeline and links to talks and interviews. The full OpenAI Forum conversation with Sid and Jacob Stern is on YouTube. On his Substack, Sid writes about the fight and about the changes he thinks the system needs, from cheaper clinical trials to the freedom to choose an independent ethics review board. His options went from zero treatments to thirty that he hopes never to need. As he puts it: it is great to have options.