Can a Biohacker With AI Tools Do Pharma's Job?
The hype around a tech entrepreneur, his dying dog, and a cancer vaccine he designed using AI tools gets almost everything wrong. The thing it gets right will one day reshape medicine.
You have likely read about Rosie the dog by now. If you have not, a Sydney engineer used ChatGPT and AlphaFold to design a personalized mRNA cancer vaccine for his dying dog, and the tumor shrank. The story went hyper popular, and it generally divided commentators into two camps. One says a guy with AI just did what pharma spends billions on. The other says this is one dog, no controls, no peer review, meaningless noise that does not deserve the attention of this magnitude.
I want to suggest that the real matter is different.
The question is not whether Rosie’s vaccine works — clinically, an uncontrolled N-of-1 in a dog tells us almost nothing.
And it is not whether biohackers with AI can “replace” pharma researcb — they obviously can not.
The question is what happens when the design of a personalized therapy, the part that used to require institutional teams and years and millions of dollars, becomes something a technically skilled and well-connected outsider can do with publicly available AI tools, while everything after the design remains locked behind the same institutional gates as before.
That gap between what one can now figure out and what one is allowed to do about it is the story. And it is bigger than one dog’s fate…
What Actually Happened
Let me lay out the facts, because the viral versions are overstating things.
In 2024, Paul Conyngham’s rescue dog Rosie was diagnosed with aggressive mast cell cancer. Surgery and chemotherapy failed, and vets basically gave her months. Conyngham has no background in biology, but the viral framing of a random entrepreneur and pet owner stumbling into science is misleading. He is a computing engineer with 17 years in machine learning, a co-founder of a data consultancy, and a former director of Australia’s Data Science and AI Association.
He paid about $3,000 for genomic sequencing of Rosie’s tumor at the University of New South Wales (UNSW), arguably used ChatGPT to navigate bioinformatics pipelines he had never worked with, ran the data through AlphaFold to model mutated proteins, and produced a half-page mRNA vaccine formula targeting Rosie’s specific mutations.
Martin Smith, the computational biologist who directs UNSW’s Ramaciotti Centre for Genomics, described himself as “gobsmacked” when he saw what Conyngham had done with the data. This kind of analysis pipeline would normally require an institutional team with a specialized biological background.
Then the wall.
Conyngham identified an existing immunotherapy matching Rosie’s cancer. The manufacturer refused to supply it for a dog. He pivoted: Pall Thordarson, director of UNSW’s RNA Institute, synthesized a bespoke mRNA vaccine from Conyngham’s formula in under two months. But regulatory approval to inject it took three more months, which was longer than designing the vaccine in the first place. Rosie got her first shot in December 2025.
The largest tumor shrank dramatically, by half to three-quarters, depending on the measurement. Importantly, and unlike what many media stories tell, she is NOT cured, the tumour did not go away completely. Moreover, another tumor has not responded.
That is the story. Accurate, interesting, and from a clinical standpoint, almost meaningless. Here is why.
mRNA cancer vaccines have been in development for over a decade, and the clinical record is sobering.
CureVac’s prostate cancer vaccine missed its survival endpoint in 2017. Argos Therapeutics’ Phase III trial in renal cell carcinoma was stopped for futility. BioNTech and Genentech’s personalized cancer vaccine missed efficacy signals in advanced solid tumors as recently as March 2025. Gritstone’s GRANITE missed its endpoint in colorectal cancer in 2024... and so on.
The consistent pattern: mRNA vaccines can provoke immune responses, but translating that into actual survival benefit in advanced cancer has been brutally hard. BioNTech’s chief medical officer, Özlem Türeci, acknowledged the core problem — the six to eight weeks needed to build a T-cell response is too slow to control rapidly growing tumors.
One dog’s tumor shrinking, with no controls and no peer review, does not change this record.
Professionals in oncology know that partial responses happen, spontaneous regressions happen, and the graveyard of promising N-of-1 results is vast.
The clinical skeptics are right in this case of a highly hyped anecdotal case.
Now, I should say that mRNA as a platform, the ability to write biological instructions as code and deliver them in lipid nanoparticles, is arguably the most important therapeutic technology to emerge in a generation.
COVID proved it could be manufactured at a population scale in an unprecedented time. The cancer vaccine applications are still early and struggling, but the platform itself is reshaping how we think about programmable medicine, from rare genetic diseases to personalized oncology to pandemic preparedness.
That story about who controls mRNA manufacturing, how the platform is being regulated, and where the next generation of applications will come from is its own essay, and I am already drafting it. I will publish in the coming weeks.
For now, the point is narrower: the specific application that went viral this week sits on the weakest branch of a very large tree.
What makes the story structurally important is something different: Conyngham compressed a bioinformatics pipeline that previously constituted one of pharma’s core competitive advantages into a solo project using tools available to anyone with an internet connection, his connections, and drive. Basically, a personal biohacking project.
It points to a possibility that the design layer of drug development is starting to leak out of the institution. But when he tried to act on what he’d designed, he hit a wall that had nothing to do with the quality of his science.
The Wall Is Older Than the Modern AI Trend in Bio
More than a decade before Conyngham opened ChatGPT, a medical student named David Fajgenbaum was dying at the University of Pennsylvania.
In 2010, during his third year, he was diagnosed with a rare subtype of Castleman disease — a disorder that triggered a catastrophic immune reaction, causing his liver, kidneys, and bone marrow to fail simultaneously. Over the next several years, he nearly died five times, cycling through rounds of chemotherapy that kept him alive but couldn’t stop the relapses. His doctors, some of the best in the country, told him they had nothing left.
What Fajgenbaum did next is worth sitting with, because it contains the structural seed of everything happening now. Between hospitalizations, he started running experiments on his own blood samples. He combed PubMed, reading hundreds of papers by hand. No AI, no AlphaFold, no generative models at the time were available to him. Just a desperately sick researcher with access to a medical library and enough training to understand what he was reading.
He found sirolimus, an immunosuppressant routinely prescribed to prevent kidney transplant rejection. It had been on the market for years, and it was cheap. It also had extensive safety data. And no doctor treating Castleman disease would have thought to prescribe it, because it was categorized as a transplant drug, not an oncology drug.
Long story short, Fajgenbaum took it, his disease went into remission, and it has now been in remission for over a decade.
This is the fact that reframes the entire Rosie the dog conversation: the wall between what could save a patient and what the system actually delivers to them existed long before AI touched drug development. Fajgenbaum’s drug already existed, it was already manufactured, and it was already proven safe. The only thing missing was someone willing to connect it to his disease, and the system, organized around disease-specific treatment algorithms, couldn’t do it. He had to do it himself while facing a sobering diagnosis.
He later co-founded a nonprofit called Every Cure, using machine learning to systematically screen thousands of approved drugs against thousands of diseases, industrializing the act of desperation that saved his life.
But the wall that made it necessary has not moved much. Drugs still get stuck inside regulatory categories, and patients still fall between the gaps in treatment algorithms. The system still optimizes for population-scale evidence, and individuals with rare or refractory diseases still get told: we’re out of options, good luck.
AI makes the design layer faster, and there is some evidence of that, but it does not make the wall meaningfully thinner for an average individual.
What It Actually Costs to Break Through the Wall
If you want to understand what it looks like when someone throws everything they have at that wall, look at Sid Sijbrandij’s truly remarkable story of biohacking his way out of cancer.
Sijbrandij co-founded GitLab, a publicly traded software company with over 2,000 employees that pioneered fully remote work at enterprise scale. In November 2022, he was diagnosed with osteosarcoma — a six-centimeter tumor growing from his spine. He went through surgery, radiation, and chemotherapy so intense that it required four blood transfusions. The cancer went into remission. Then it came back in 2024. His medical team told him, essentially, that the standard of care was exhausted.
What happened next is the most resource-intensive case of a patient taking over their own treatment that I’m aware of.
Sijbrandij deployed every frontier diagnostic available, including single-cell sequencing, bulk genomic analysis, multiple types of minimal residual disease testing, organoid models grown from his own cancer cells. He hired a full-time director of his care and assembled clinical and scientific advisory boards. He filed five separate expanded access applications with the FDA to obtain experimental drugs with no available clinical trials — each was approved, arguably, within 48 hours.
Then he flew to Germany for an experimental radioligand therapy targeting a protein called FAP that his sequencing had identified as overexpressed in his cancer cells. The treatment used Lutetium-177, the same radioactive payload in Novartis’s Pluvicto, delivered by a targeting ligand that had first been validated by diagnostic imaging. The tumor shrank enough to become operable again. Post-surgical analysis showed T-cell infiltration had surged from 19 to 89 percent. His cancer is now undetectable.
Here is the uncomfortable part, at least, it is what Sijbrandij describes based on his experience.
At every stage, Sijbrandij encountered institutional resistance that had nothing to do with the science. Hospitals wouldn’t release his own tissue samples without prolonged bureaucratic negotiation. Some of the most promising experimental drugs he found had been shelved by companies, not always because the science was wrong, but because the market was too small to justify development costs. Institutional review boards could block treatment on the basis of a single member’s concern.
Sijbrandij frames the core tension this way: getting a drug approved for a population costs around a billion dollars. Dosing a single person with a personalized therapy costs roughly a million. And that gap keeps widening because the cost of designing new medicines keeps falling while the cost of proving them at a population scale keeps rising.
He spent what appears to be millions of dollars on his treatment. He had a team of people working full-time on his care. He traveled internationally. He had the social capital to persuade companies and researchers to collaborate. He is alive, in remission, and building a new software company.
The question is what his story means for everyone who doesn’t have what he has…
The Real Shift — and Its Limits
The hype says AI is disrupting pharma. The reality is more specific and more consequential.
What AI is actually compressing is the discovery layer of drug development: identifying targets, screening compounds, predicting molecular interactions, designing candidates.
This is real.
At MIT, James Collins trained generative AI to design 36 million novel antibiotic compounds, synthesized 24, and found two that effectively kill drug-resistant MRSA and gonorrhoea through entirely new mechanisms. At Cambridge, Michele Vendruscolo can now screen billions of small molecules against Parkinson’s disease targets in days for a few thousand pounds, work that previously took six months and cost millions. In commercial realms of AI drug discovery, progress is even more real, as I described in an industry report last year. These are real advances that are published, validated, and progressing toward clinical testing.
But discovery is one layer.
What AI does not compress too much is everything that follows: GMP-grade manufacturing, formulation, safety testing, regulatory approval, clinical trials, and the logistics of getting a physical product into a patient’s body. Vendruscolo said it clearly: “AI is revolutionising drug discovery. But only in very specific ways.”
Now, I am an avid user of LinkedIn, and the hype about Rosie the dog’s vaccine “miracle” is overwhelming as much as it is misleading.
Say, there is this “Napster moment in pharma” analogy that circulated in the discourse; it is instructive precisely because it is wrong. Napster made distribution free for a product, a digital music file, that required zero manufacturing and clinical testing, risking human life. An MP3 is pure information; copy it and you have the product. A drug design is also information. But a drug is a physical object: synthesized molecules, lipid nanoparticles, cold chain, quality controls, and a trained professional to administer it. You cannot download a vaccine.
This distinction is where the real power story lives.
Where Power Migrates
When one layer of a value chain commoditizes, the layers adjacent to it don’t become less valuable, they become more valuable. This is what is happening in drug development, and almost nobody in the viral discourse on LinkedIn and elsewhere is talking about it.
Manufacturing is becoming the chokepoint.
Contract development and manufacturing organizations with the certifications to produce clinical-grade therapeutics, the CDMOs, are increasingly becoming the most strategically important players in the system. Conyngham needed UNSW’s RNA Institute. Sijbrandij needed a specialized German facility. During COVID, Moderna and BioNTech had the vaccine designs ready in a matter of hours or days, and still needed months to scale manufacturing despite virtually unlimited funding and political urgency. As more people and more AI systems can design therapies, the power shifts to whoever can make them. This is not a software problem. It is an atoms problem, and atoms don’t scale like code.
Regulatory frameworks are the new dividing line.
The FDA’s expanded access pathway worked for Sijbrandij, but it’s designed for individuals, not populations.
A new “plausible mechanism pathway,” proposed after the first personalized CRISPR therapy was administered to an infant in 2025, represents the earliest attempt at a regulatory framework for individualized medicine at scale.
Meanwhile, China has adopted a different posture toward experimental treatments — creating a geopolitical dimension to what looks on the surface like a procedural question. How fast a country’s regulatory system adapts to personalized therapies may determine where the next generation of biomedical innovation actually happens.
Proprietary data is the hidden gate.
The tools are opening up — AlphaFold is free, ChatGPT, Claude, Gemini, etc. are ubiquitous. But as Collins at MIT noted, the critical datasets on drug absorption, distribution, metabolism, and toxicity are held by pharmaceutical companies and are not publicly available. The front door to drug design is unlocking, but the back room where the real knowledge lives is not.
Capital is still the ultimate filter.
Conyngham needed $3,000 and a personal connection to a willing nanomedicine pioneer. Fajgenbaum was already inside the system as a Penn medical student. Sijbrandij needed millions, a full-time care team, and the ability to fly to Germany on short notice. The design layer is democratizing. The execution layer may be getting less equal because as design becomes abundant, the scarce resources (manufacturing, regulatory navigation, data, clinical access) command more leverage, not less.
And the AI tool providers are becoming something new entirely in the biomedical and healthcare fields.
This is the shift that is happening. When Conyngham used AlphaFold to model Rosie’s mutated proteins and ChatGPT to navigate bioinformatics pipelines, those tools weren’t passive infrastructure, like cloud servers or MS Excel used to be, the way a microscope or a sequencing machine is. They were making consequential analytical judgments: which protein structures to prioritize, which pipelines to recommend, which targets looked promising. The tool was doing work that used to be done by specialized teams inside pharmaceutical companies. A sequencing machine generates data, and AI interprets it. That’s a fundamentally different position in the value chain. DeepMind, OpenAI, Anthropic, Google, and the growing ecosystem of specialized biotech AI platforms — Insilico Medicine, Recursion, Isomorphic Labs — are not selling picks and shovels to the gold rush. They are becoming the analytical intelligence layer of drug development itself. Whoever controls the models controls the lens through which the next generation of therapies is conceived. Pharma’s old intellectual moat, which has been the institutional expertise to go from data to hypothesis, is leaking to pharma outsiders, like big tech or even individual biohackers with resources. It is a new phenomenon that is now enabled by more and more powerful AI tools.
The Real Question
Here is the thing that makes me most uneasy about the Rosie discourse.
The viral narrative — “biohacker beats pharma” — is factually wrong but politically powerful. Millions of people will read this story and may start believing, based on a single dog’s anecdotal medical case, that pharmaceutical development costs are mostly artificial, that the regulatory apparatus is mostly obstruction, and that AI has made the whole system optional. None of this is true. But it does not need to be true to reshape perception and, potentially, even policy.
The notorious $2.6 billion average development cost is not primarily science — it is the cost of proving that a drug is safe and effective at a population scale. Clinical trials, adverse event monitoring, manufacturing consistency, and post-market surveillance. This infrastructure exists because without it, drugs could harm people at scale — thalidomide in the 1960s, which caused thousands of severe birth defects, is the reason most of modern drug regulation exists in the first place.
Rosie’s vaccine was personalized, designed for one dog, nobody is proposing it be given to millions. But the regulatory system that governed the red tape Conyngham complained about is the same system that governs every drug on the market. When public contempt for that system grows, fed by stories where “the red tape was harder than the vaccine,” the consequences extend far beyond one dog or one biohacker. They reach into the framework that protects everyone.
At the same time, the institutional defense — “N-of-1 means nothing, move along” — also misses the momentum. The design layer is commoditizing, it is an obvious trend. Outsiders can now do analytical work that was recently an institutional monopoly. And every time a Fajgenbaum or a Sijbrandij breaks through the wall and survives, the question of why that wall is so high for everyone else gets harder to dismiss.
The tension is not between innovation and regulation. It is between the speed at which individuals can now identify what might save them and the speed at which institutions can verify whether it actually will. AI is accelerating the first, but not the second — not yet. And the gap between those two speeds is going to produce a decade of collisions, in courts, in legislatures, in hospital ethics boards, and in the lives of patients who can, with help of AI, theoretically understand what might help them and can’t reach it in practical life.
Sid Sijbrandij calls himself “the Kool-Aid Man breaking through the wall.” He can afford to be. The harder question is what happens for everyone else. How do you make personalized medicine accessible to people who aren’t billionaire founders, without abandoning the safety standards that exist?
Nobody has a good answer yet. But the pressure to find one is building faster than most institutions realize. AI is increasingly a player in the game, or should I say, the new generation of companies that possess the keys to the most powerful AI systems on the planet…
Thanks for reading. Please share it with those who may find it useful!
— Andrii