Racing Death: Four Forks in Modern Medicine

In September 1928, Scottish bacteriologist Alexander Fleming returned from holiday to find a forgotten culture dish contaminated by mold—and the staphylococcus around it dead. He recorded and named "penicillin," but could not purify it into a drug, and within a few years all but gave up. Discovery was only the opening.

The real breakthrough came twelve years later. In 1940, Oxford's Howard Florey and exiled Jewish biochemist Ernst Chain revived the forgotten paper, proving in mice that it cured lethal infection. Purification was brutal—they used bathtubs and milk churns as "factories." Their first patient, a policeman, improved in 1941 but died when the supply ran out. Wartime Britain couldn't mass-produce it, so Florey went to America, where deep-tank fermentation raised yields a thousandfold. By the 1944 Normandy landings, penicillin was saving wounded soldiers in bulk.

Countless papers hold "promising but unpursued" findings; the "valley of death" from lab to bedside persists. AI-driven drug screening aims at exactly the translation problem Florey once brute-forced with bathtubs.

In the early 1950s, the structure of the genetic material was biology's greatest mystery. Several teams raced: Caltech's chemistry titan Linus Pauling, Rosalind Franklin at King's College London using X-ray diffraction, and two unknowns at Cambridge, James Watson and Francis Crick—the latter ran no experiments at all, only built models.

The turning point was a crisp diffraction image Franklin captured—"Photo 51"—which pointed almost directly at a helix. In early 1953, her colleague Maurice Wilkins showed it to Watson without her consent. Watson later wrote that "my pulse began to race." Combining it with Franklin's unpublished data, the pair published the double-helix model in Nature that April—a single page that rewrote biology. Franklin died of cancer in 1958, never sharing the 1962 Nobel.

Who owns data, who may use it, whose name goes on it—from authorship disputes to the copyright wars over AI training data, Franklin's dilemma is magnified without limit in the data age.

Launched in 1990, the Human Genome Project was a public science endeavor led by the US NIH with international partners under geneticist Francis Collins, on the principle that data be released daily and shared freely. By 1998, the maverick scientist Craig Venter broke away to found Celera, claiming his "shotgun" method was faster and cheaper—and hinting he might patent certain genes.

A public-versus-private race ignited. The public consortium feared that if human genes were privatized and patented, all future diagnostics and research would owe a toll. The competition, paradoxically, accelerated progress. On June 26, 2000, President Clinton brought both sides to the White House to jointly announce a "working draft," deliberately proclaiming the genome "belongs to all humanity." It was completed in 2003. Celera never managed to lock up gene patents; the principle of open access held.

Whether a foundational capability should be open or private is exactly today's central fault line in AI: open model weights vs closed APIs. The genome fight was a rehearsal for that very question.

Hungarian immigrant scientist Katalin Karikó believed mRNA could become medicine—instructing the body's own cells to make therapeutic proteins. But through the 1990s, her grants were rejected again and again; she was demoted and nearly lost her job, because injected mRNA always triggered a violent immune rejection. Most peers thought the path was a dead end.

The turn came from a chance meeting by a photocopier—a collaboration with immunologist Drew Weissman. In 2005, they discovered that modifying a single nucleoside in the mRNA would fool the immune system and avoid rejection. The paper drew almost no attention at the time. Fifteen years later, when COVID struck, BioNTech and Moderna stood on exactly this finding to build highly effective vaccines in under a year—the fastest in history. In 2023, Karikó and Weissman shared the Nobel Prize.

Today's AI boom likewise rests on decades of neural networks on the bench—the backpropagation the Hintons kept alive through the "AI winter" (see Day 10). The compound interest of basic research always accrues quietly where no one is watching.