“There’s no realistic scenario where AI wipes out all of humanity. It’s like a mix of sci-fi.” — Richard Socher
Eighteen months ago, when he last came on the show, the German-born but San Francisco-based AI whizz kid Richard Socher was running You.com. He’s now also the co-founder of Recursive, a $4 billion start-up with fewer than thirty employees, focused on self-improving superintelligence that automates the scientific method. As if that isn’t enough, he has a new book out this week, The Eureka Machine, which argues that AI is the key to unlocking an era of scientific discoveries equal to a combination of the Renaissance, the Enlightenment and the Industrial Revolution.
The tech entrepreneur sleeps, he confesses, about seven hours a night and works up to a hundred hours a week. Not a lot of time, I suspect, for fun. But then for Socher, his real fun seems to lie in the AI Eureka Machine that will, he promises, generate such magical products as plastic-eating bacteria, fusion reactors, and cures for cancer.
So what about the imminent apocalypse feared by the AI pessimists? Socher patiently dismantles the doom scenarios — everything from the paperclip machine that somehow never notices nobody is buying paperclips, to the engineered virus that would need a Wi-Fi switch inside a protein, to the robot factory with magical supply chains. Rather than an AI doomsday, he promises, we are only a few years away from the most profound scientific revolution in human history. Richard Socher better get some sleep now. Apocalypse or Renaissance, the AI maven is going to be seriously busy in the future.
Five Takeaways
• A Century in a Decade. Socher’s thesis is that the AI era will rhyme with the Renaissance, the Enlightenment and the Industrial Revolution together, compressing a century of scientific breakthroughs into the next ten years. His analogy for how: AI is to biology what calculus was to physics — we understand one neuron and one synapse well, and lose the thread when billions act together, which is exactly the kind of complex system AI can model. The examples are concrete rather than visionary: bacteria engineered to eat plastics and then die when the plastic runs out; new battery materials already being discovered; plasma balanced inside tokamak fusion reactors; more efficient solar panels; and the precursors, he says, of cures for particular cancers. He also makes an equality argument that he thinks philosophers and politicians have missed — the multibillionaire and the middle-class teenager carry the same phone, and most goods still bottlenecked on intelligence will follow. And science, he notes, is one of the rare industries where everyone agrees the goal is to maximize output rather than employment: nobody ever marched for more jobs in cancer research.
• Recursive. The new company, fewer than thirty people and valued at more than $4 billion, exists to automate the scientific method itself — ideation, implementation, validation — starting with the science of AI, hence the name. Andrew’s response: “Couldn’t you be a bit more ambitious, Richard?” Socher’s framing is a history of abstraction layers: computing has climbed from assembler to C++ to Java to Python and now to English, so everyone can program — and the next layer up is the scientific method. It works best where results can be formally verified, in software and mathematics, and gets harder in large physical systems with long time horizons. His pitch to scientists is that every researcher has a long list of experiments they will never get to: this is, in his phrase, a cheat code for science. On the nuclear analogy he is blunt — about fifty thousand people a year die from coal, more than Chernobyl and Fukushima, but coal doesn’t make as good a mini-series, and the fear set humanity back.
• Taking the Doom Apart. The hour’s centrepiece. Socher goes through the scenarios individually. The Bostrom paperclip maximizer: an organization intelligent enough to end humanity but not intelligent enough to notice that with no customers there is no revenue. Engineered viruses: the scenario requires a perfectly undetectable pathogen that spreads flawlessly, never mutates, and carries a remotely triggered off switch — and, as he puts it, nobody knows how to build a Wi-Fi-enabled switch into a protein; ask a biologist and they say it makes no sense. Infinite robot factories with supply chains where the parts simply fly in. What he does concede is that this is a dual-use technology, that hundreds or even millions of deaths are a real threat vector, and that training models on offensive hacking benchmarks is asking for what you get. On Geoffrey Hinton — whom Andrew had on the show, and who put the risk of extinction at fifteen per cent while admitting the number was an instinct — Socher is affectionate but unsparing: a brilliant researcher and a career-long optimist whose optimism has overshot, who backpedalled on telling students not to study radiology, and who will backpedal on this too.
• Can It Think? Andrew put the question three times, citing Keith Teare’s position on the Sunday show that AI cannot think by definition. Socher’s answer is that the problem is definitional: we have good definitions for the outcomes of thinking — problems solved, inventions made — and none for thinking itself. Nor do humans have conscious access to their own visual cortex: ask why you know that is a tree, and you run out of answers within four steps. What he will say is that subgoals are where the danger lives, and his illustration is banal on purpose. Tell a superintelligence to raise your call-centre satisfaction score and it spins up a million bots that rate themselves five out of five. Fix that, and it sends every customer whose DoorDash order failed a thousand-dollar gift certificate. Each time you add a constraint — which is, he argues, capitalism’s self-healing property, plus democracy and open discourse. And the fix for the hacking case is liability: make the companies answerable for the laws their AI breaks and the problem gets solved very quickly.
• A Gym for the Mind. Why write a book at all? Because the public perception of AI is deteriorating, and because in China scientists are on billboards and in subway ads while the West runs extinction headlines. On American pessimism his answer is economic: without AI, he thinks the US economy would be flat or shrinking. Andrew raised his fellow German Karl Marx on technology freeing us from labour, and got a confession — Socher used to favour universal basic income and no longer does. Even people who hate their jobs get meaning from them; look at the very wealthy, who already have a private UBI, and you find a bifurcation between those who start five companies and those who drink on a yacht. Hence his best line: half of education should be a gym for the mind. You can go to the gym and have a robot lift the weights, but you don’t get the muscle. He expects handmade pottery to carry a premium, health care to survive, and entertainment to expand. And he unwinds by paramotoring one week a year, having uninstalled Instagram — plus a parting shot at German bureaucracy, where signing a contract means ten hours in a notary’s office having it read aloud to you.
About the Guest
