In early 2016, when she became Calico Labs’ chief computing officer, the views and aspirations of Raymond Kurzweil did not inhabit the mind of Daphne Koller. This wasn’t because she shunned Kurzweil, or digits, algorithms, or computer code; far from it. She loved them all. It was just that Kurzweil wasn’t her favorite when it came to the stratagems of machine learning and artificial intelligence. Bridge Four wasn’t her thing. Her inspiration was Thomas Bayes, an 18th-century mathematician and Presbyterian minister. She had explored his work and theories in the 1990s, and then used them to develop some of the most advanced artificial intelligence algorithms in the world.
Before joining Calico, Koller knew she could write her own ticket at any of Silicon Valley’s great Cloud makers: Google, Facebook, Apple. But did she really want to create software that made better Twitter feeds, or cool faces on Snapchat, or yet another product that piled up more of the planet’s digital ad revenues? In the sprawling world of Google, she would be a blip: one more very smart human in a sea of geeks. But Calico was different. She had never forgotten Steve Jobs’s grand goal: “Make a dent in the universe.” At Calico, maybe she could make a dent—save lives, perhaps millions—even her own.
Thomas Bayes would have appreciated Koller’s interest in saving lives, being both a Presbyterian minister and a devoted supporter of the humanist philosophies that called for the rational improvement of the human race—also the foundation of transhumanism. The mathematician’s work had been largely forgotten when Koller first began exploring it. But she liked the way he thought. It delivered a human flexibility unusual in mathematics. Instead of using set rules and cold logic, it was designed to adapt.
A simple example was the big urn problem. Imagine a large urn filled with balls. Half are black, the other half white. What was the likelihood that when one of the balls was pulled from the urn, it would be black? Well, obviously, 50-50. But what were the odds of pulling the very next black ball from the urn? Now the situation had changed; every time a ball was retrieved it would change again, and again, depending on the situation. Bayesian probability took these differing possibilities into account, or tried to. In the jargon of computer scientists, it both explored information and then exploited it—not unlike a human mind.
Good news, too, when it came to computer science, because Koller knew the world was becoming an increasingly noisy and confusing place. The rigid rules created by most computer code simply would not get the job done. Not where biology was concerned, and especially not molecular biology, which was about as messy as it got. No algorithm written by the human mind—no matter how fast or logical—could resolve the complex and rapid molecular pathways at work in a genome or a biome or any “ome.” Not that humans were completely useless; the initial coding had to be set up correctly. But after that, the algorithms were cut loose do their stuff, independently, like a living thing solving each problem it faced as it faced it without constantly coming back for human instruction. This was the only way that the billions of genetic communications could be unsnarled.
When Koller developed her cancer gene-mapping techniques at UCSF after receiving her MacArthur grant, the program blasted through data on thousands of genes, and then tested the likelihood that changes created by one gene could be teased out by locating changes in the others. She also developed code that examined the rates at which specific genes within a cell created its corresponding protein, and how that creation depended on signals from proteins encoded by still more genes.
All of this sounded a lot like Kurzweil’s pattern-recognition nodes. Both needed to understand context and then react on their own in real time. Kurzweil’s AI nodes were meant to solve a broad range of problems that could eventually lead to immortality. Koller’s algorithms were more specific, delving way, way down into the noisy, and entirely invisible world of organic chemistry by letting the algorithms come up with their own solutions. But in this way, they too were thinking, artificially.
David Botstein, Calico’s chief scientific officer, was of the opinion that understanding these pathways was central to Calico’s hopes for success. He and Levinson agreed that crunching lots of numbers to figure out what the human genome was trying to say was crucial—but Botstein wanted to know more. What the hell were all of those molecules inside the genes actually doing? How did a gene turn some proteins on and some off? How did one gene affect others? If only you could rummage among them all, look them in the eye, and say, “Ah-hah! Now I understand what you’re up to.” That was the only real way anyone could hope to begin undoing all the damage that took the human body apart.
To achieve this, Botstein was having custom-made sequencers developed for Calico, and Calico alone, by two companies collaborating in California: Pacific Biosciences and Bionano Genomics.21 These hand-built sequencers, and these only, he felt, would provide the kind of resolution that revealed the invisible machinations of proteins and molecules.
ONE OF THE PROJECTS Calico began exploring when Koller arrived in 2016 illustrated just how numbingly complex the problems were that the company faced. Between 2006 and 2010, an organization called the UK BioBank recruited 500,000 subjects between the ages of 40 and 69 years old and tested the living daylights out of them. BioBank’s researchers could tell you just about everything about the project’s recruits: their inflammatory markers, cholesterol and hormone levels, saliva samples, urine and blood results. They could tally up grip strength, the volume of the recruits’ brains—even how quickly they could walk 100 meters. The genes of 100,000 participants had already been sequenced, and then those people were asked to wear 24-hour activity monitors for a week. Recently, fMRI machines had begun scanning a fresh batch of 100,000 people. All of this in addition to the vast database BioBank was already developing about recruits’ diet, cognitive function, work history, and digestive health.
Venter and the researchers at Human Longevity, Inc., had been gathering information at a blistering clip too, but it was hard to imagine a more robust aggregation of human biology than the findings the UK Biobank was compiling. In addition, the bank had also begun following up with those in the studies, plotting their progress as they marched forward. Already they had tabulated 9,000 phenotypes correlated with their unique DNA.
The numbers were mind paralyzing: trillions of cells, each interacting with billions of genes performing untold numbers of biological interactions. All of them unveiling the two faces of humanity. One that could reveal you and you alone; the other that represented the great biological database that made all humans possible. The questions were endless—and who could say which genetic alterations provided the final answers? Koller felt pretty sure Calico would find more than one, but less than 5,000. She wondered if maybe five major genetic pathways could eliminate aging. If so, perhaps Calico could then come up with five drug regimens that would intervene and repair the killing. That would solve beta outright.
One fact was clear: The ultimate answers weren’t going to come from a “who.” They would come from a “what.” Maybe not from Kurzweil’s neurobots, or human-machine hybrids. Not yet. But from increasingly intelligent machines that were thinking up new ways to think. Their very own ways. That made them the new oracles: software reading the burnt offerings of all that massive data accumulating at exponential speed as the contraptions taught themselves to mend humanity, rather than terminate it. Imagine: Machines solving the ultimate human problem.
The irony was almost cosmic.