When AI Brings Hope: What a Cancer Story Shows About the Future of Medicine

If you follow AI news, you mostly get served two kinds of headlines right now: fear of job loss and fear of misuse. Deepfakes, security flaws, mass layoffs, surveillance, manipulation. Much of that is real and deserves attention. But because of that, something else almost disappears: the same technology can also compress knowledge, reduce friction, and in some cases give time back to people when they need it most.

I have already written here on the blog in another post about how quickly everything is shifting right now. This article is, in a way, the medical counterpoint to that: less market, less hype, more human reality.

One story from the OpenAI Forum is exactly that kind of case. It is not sentimental, not convenient, and definitely not a miracle promise. But it shows something we will likely hear much more about in the future, including from people who are not millionaires.

In short:

• In this case, AI did not replace a doctor. It massively accelerated analysis, research, and coordination.
• Today, this kind of path is still unfairly distributed, because it takes money, time, access, and the will to work through medical complexity.
• That is also how many major technologies begin before they eventually become normal for far more people.

If you want to see the original talk, you can watch it directly here:

A Story That Is Easy To Miss

At the center is Sid Sijbrandij, co-founder of GitLab. He was diagnosed with osteosarcoma, a rare and aggressive bone cancer. In the forum, he describes very openly how brutal the standard treatment was: surgery, spinal fusion, radiation, chemotherapy, blood transfusions, and physical collapse. For many people, that alone would already be more than enough.

An important early turning point for him was so-called click chemistry, a Nobel Prize-winning method that allows certain compounds to be linked in a very targeted way. For Sid, that was not only scientifically interesting, but mentally decisive, because it showed him that medicine does not always have to move only through large standard trials. Through paths such as a single-patient IND, there can also be legal and official routes for highly individualized treatment in extreme situations.

Then came the harder part: the cancer came back. There were barely any established standard options left. No clean pipeline. No obvious next step. No simple “we’ll just do treatment B now.” And that is exactly where something happened that feels very characteristic of this moment in time: he quit his day job and went into founder mode, not for a startup, but for his own survival.

That sounds absurd at first. A tech founder suddenly treats cancer like an engineering problem. And honestly, my first reaction would also be skepticism. Medicine is not a SaaS product. A chatbot is not an oncologist. An API call is not a therapy.

Still, it would be too easy to dismiss this story as a Silicon Valley special case. What becomes visible here is not a “ChatGPT miracle cure,” but a new way of working in medicine: more data, more pattern recognition, more parallelization, faster hypotheses, better questions, and shorter loops between finding and decision.

What AI Actually Did In This Case

The most important point first: AI did not cure the cancer. That would be a dangerous and false simplification. What helped here was a combination of diagnostics, specialists, experimental approaches, clinical experience, personal persistence, and AI as an accelerator.

In the talk, Sid Sijbrandij and Jacob Stern describe several concrete examples.

They pulled together enormous amounts of data, including, by their own account, roughly 25 terabytes of material. That included bulk RNA sequencing, single-cell analysis, pathology, scans, blood data, organoid models, and multiple experimental diagnostic methods. That depth of data is far out of reach for most patients today. But the interesting part is what you do with it once you have it.

Jacob Stern, a geneticist himself, says explicitly in the forum that he is not a doctor. What AI gave him, in his words, was not magical knowledge, but something like an Iron Man suit for highly specialized domains. In other words, the ability to get up to speed quickly, talk to experts intelligently, ask better questions, and prepare the right next steps.

One early example: they fed a CSV file from bulk RNA sequencing into ChatGPT and asked the model for an initial assessment. Even that was useful enough to surface certain markers and dynamics in the tumor environment more quickly. Later, they pushed much further: natural language as input, then agentic literature review, hypothesis building, bioinformatics steps, code, plots, and a combined report as output.

In one example from the talk, such an analysis ran for around 30 minutes and cost roughly 20 dollars in API usage. In the background, it was connected to about 600,000 single cells from different blood time points. That contrast alone is remarkable: huge biological complexity on one side, relatively low-cost analytical assistance on the other.

Things become even more interesting when the talk turns to therapy development. They discuss:

• a personalized mRNA cancer vaccine based on specific tumor mutations
• TCR T-cell therapies designed to target the tumor more precisely
• CAR-T approaches with added safety logic to reduce dangerous effects on healthy tissue
• the search for understudied targets, meaning proteins or structures that may be highly relevant in one individual case, but barely researched in the mainstream

Sid says at one point in the talk that his therapeutic ladder grew from zero to more than 30 options. Not because a miracle drug suddenly appeared, but because data, diagnostics, specialist knowledge, and AI together opened up a much larger search surface.

The best example of this needle-in-a-haystack effect is, to me, Penexin 3. According to the talk, this protein appeared roughly 10,000 times more strongly in his cancer than in healthy tissue. And yet there was almost nothing in the literature. The team’s suspicion was that because the target is hydrophobic, it simply falls through the cracks in many standard tests. That is exactly where AI shows what it can do in cases like this. Not because it is “smarter than every researcher,” but because it is patient enough to search huge data spaces again and again until unusual signals like this become visible.

That is where AI’s actual strength shows up. Not as an oracle, but as a tool for turning thousands of papers, data points, markers, side-effect profiles, and biological connections into something operational much faster.

That is the core of the story for me: AI compresses specialized knowledge. It does not turn a layperson into a doctor. But it can bring committed patients, families, and interdisciplinary teams much closer to current literature and real decision-making space.

This Future Is Still Unequally Distributed

We should not romanticize this. Right now, this story is still highly elitist.

To proceed like this, you need:

• money
• time
• very strong specialists
• access to rare diagnostic methods
• enormous mental energy
• the ability to tolerate medical complexity

They say this quite openly in the talk themselves. Some parts are extremely expensive. Others have become surprisingly cheap. And that mix is exactly what makes this moment so interesting. We are watching a transition: actual therapy development is often still expensive, slow, and bureaucratic. But some of the layers beneath it are already getting cheaper.

In the forum, they cite rough figures such as around 50 dollars for bulk RNA sequencing, around 500 dollars for whole-genome sequencing, and around 20 dollars for a more advanced AI-assisted analysis. Those are not the full costs of cancer medicine, but they are signals. The insight layer is starting to deflate.

And that matters enormously, because medicine today often does not fail first because of bad intentions, but because of friction: data is scattered, specialist knowledge is siloed, studies take forever, rare cases fall outside standard pathways, and too much time passes between finding and decision.

What I also find powerful about this story is that it is explicitly not meant to remain an elite playbook for a single case. In the talk and on the related project site, it becomes clear that data, learnings, and workflows are deliberately being prepared so other patients and teams can build on them faster later.

How Many Big Technologies Begin

That is exactly why this story made me think of an old pattern. New technology is almost always reserved at first for people who have more money, more access, and more room to experiment than everyone else.

In many periods of history, bathtubs were mostly for kings, nobility, or very wealthy households. Today, warm water is an ordinary part of daily life for millions of people.

And that is not an exception. Three more examples show the same pattern:

• Electric light in the home was once a luxury for wealthy buildings, hotels, and industry. Today, you only notice how standard it is during a blackout.
• Air travel used to be a symbol of wealth, status, and exception. Today, it is not equally affordable for everyone, but it is clearly mass mobility rather than an aristocratic privilege.
• Genome sequencing once took billions of dollars and many years. Today, in some cases, we are talking about hundreds of dollars and days rather than decades.

That is exactly how I think we should look at this new form of personalized medicine. What looks today like concierge medicine for highly connected founders could later become more broadly available in a simplified, standardized, and automated form.

Technology scales. First slowly, then suddenly.

Not automatically. Not by itself. Not fairly overnight. But historically, the pattern is clear: first luxury, then tool, then infrastructure.

Why This Still Makes Me Hopeful

What impresses me most about this story is that it does not rely on a miracle. It relies on compression.

AI compresses:

• literature
• diagnostics
• pattern recognition
• communication across disciplines
• the ability not to give up too quickly on rare cases

And that is enormously valuable in medicine. Rare diseases and unusual courses often do not fail because there is no information at all, but because nobody has the time to pull it together in a useful way.

If thousands of pages of medical history, lab values, scans, RNA data, and studies can be turned into a workable model, that is not only relevant for wealthy outlier cases. It is a preview of a medicine that could become more precise, more personal, and hopefully more fair.

What still looks like exceptional treatment today could become standard later in a reduced and more practical form:

• a system that automatically highlights rare markers
• an assistant that pre-sorts relevant studies for doctors and patients
• a tool that makes possible therapy paths and side-effect profiles more transparent
• diagnostics that take individual biology more seriously instead of only following standard paths

And yes, we will almost certainly hear much more about this, not only from founders with resources. The tools will improve. Costs will fall. Interfaces will get easier. That is where scaling begins.

My Conclusion

That is why I think this story matters. It helps straighten the view a little. AI is not only risk. Not only layoffs. Not only hype. Not only a security problem. It can also become a tool that helps people navigate a sick, slow, and often overloaded system more effectively.

Of course, we should not fall into technological romanticism. The most important health foundation remains surprisingly unglamorous and very human:

• enough sleep
• movement
• nutrition
• everyday discipline

We still control a large part of that ourselves. It remains the foundation.

If you are interested in exactly those topics, there are already practical and personal posts on the blog about health tracking and health data and also about Whoop, recovery, sleep, and strain .

But right after that, I am convinced, we will need AI. Not as a replacement for doctors, not as a miracle button, but as a second layer of intelligence above data, research, and complexity.

And that is also where Google matters to me. With AlphaFold, Google DeepMind made predicted structures for more than 200 million proteins openly accessible, and with AlphaFold 3 it pushed toward the next step, where not only proteins but also their interactions with other molecules can be modeled more effectively. That is not “the cure for all diseases.” But it is a massive foundational lever because it makes biology much faster to read for researchers.

So I remain hopeful. Not naive. Not blind. But hopeful. If we can think lifestyle, prevention, good medicine, and AI together in a clean way, the future could genuinely bring more health to far more people.

Until next time,
Joe