Death is an Engineering Problem | Originals
OOriginals
Science just discovered the antifreeze that makes reversible human cryonic freezing possible. The only problem is that it kills the patient. Dr. Alex Mathiasen, PhD, a 32-year-old mathematician with no lab, no chemists, an Oxford physicist co-founder and a stack of GPUs is betting his career that he can compute the poison out of existence using quantum simulation. Cryonic toxicity is the last wall. And his parents are sixty.<br>Subscribe now
Alex Mathiasen was seven the first time he understood that everyone he loved would die. He sat on the floor and cried, and his mother gave him the only comfort any parent has for this particular arithmetic: don’t think about it.<br>He ignored her.<br>He is thirty-two now, a Danish mathematician with a PhD in algorithmic optimization. For his entire adult life he has aimed that one skill at a single target, and it has carried him here — to a company called Vitrify Labs, and a proposition that sounds, on first hearing, insane.<br>Scale beats clever. Even in biology.
Mathiasen wants to build a pause button for human biology. Stripped to the studs, the idea runs like this. Take a person who has run out of time — late-stage cancer, an organ failing, the disease that took his grandmother — and freeze them in suspended animation, like Han Solo in Star Wars. Hold them there, unchanging, at 196 degrees below zero, for as long as it takes medicine to invent the cure they need. Then warm them up. The same person, resumed.<br>He does not think this is science fiction. He thinks it is an engineering problem the world has been too unimaginative to take seriously — and that the last obstacle between us and that pause button is a poison he can compute his way around.<br>Scale Is All You Need<br>You would be within your rights to stop reading. The dream of freezing the body and waking it later has a long and faintly disreputable history — frozen heads, desert dewars, promises that conveniently cannot be checked for a century. But Mathiasen is not a man you wave off.<br>Strip away the titles and the whole of him is a gift for the shortcut. During his PhD at Aarhus University he built a parallel algorithm for one of the slowest operations in machine learning — decomposing a dense matrix — and made it run, in the best case, 27.1x faster than the standard method. The paper went to NeurIPS, the field’s most selective conference, in 2020. Then, at the chip company Graphcore – an AI chip company bought by Softbank for $600M to compete with Cerebras and Groq – he ran into the lesson that now drives his company: in machine learning, scale wins.<br>Big Pharma buys drugs, not platforms.
Feed a simple model a mountain of data and it beats a clever model fed a molehill. In biology the mountain doesn’t exist, because nature is expensive to measure — so he built it, simulating molecules on Graphcore’s chips until he had a dataset of 1B examples. The previous record was 20M. It had taken two years on conventional supercomputers; he produced 50x more in days.<br>That is the thesis: scale beats clever, even in biology.<br>And here is the tell. He could have sold that dataset as the definitive tool for chemical machine learning. Instead he published its flaws — in his own hand, in the documentation, he warned other researchers that his shortcuts had introduced errors and that they must not use his data to rank their models. In a field full of people overselling, he undersold his own billion-molecule monument. Hold onto that. It is rarer than the math.<br>Elegant math wins academic prizes. Functional math ships products. Mathiasen builds functional math.<br>Why Not Just Start a Company?<br>During a seven-month stint at Charm Therapeutics in London, he learned the broken economics of drug discovery startups up close: Big Pharma buys drugs, not platforms. So the bio teams built narrow specialists — one deep-learning model in service of one drug against one cancer — not GPTs, not the general-purpose transformers powering Anthropic and OpenAI.<br>The bio people were ignoring the scaling laws.<br>Build a GPT that understood biology the way GPT-5 understands language, and drug discovery becomes a query against it. The only open question was the cost of the data, which — unlike writing on the internet — does not yet exist in digitized form.<br>Cryo scaling is beating Moore’s Law. 100x in 3 years.<br>Subscribe now
But every dollar at a biotech chases the next molecule, because the next molecule is what gets the company acquired. A biology GPT would cost at least as much as starting two more drug programs — so nobody starts one. Single molecules are the business model.<br>“There’s a 100% chance it works,” laughed Etminan. “It’s just we can’t tell you how much it’s going to cost. It could cost a million bucks. It could cost a trillion.”
Mathiasen felt stuck.<br>Then a venture capitalist asked him the question founders spend their lives waiting for — why don’t you just start your own company?
Thus was born...