Ten years ago, I would have turned my nose up at the idea that we already understood how to get machines to think. In the 2010s, my team at Google Research was working on a wide variety of artificial-intelligence models, including the next-word predictor that powers the keyboard on Android smartphones. Artificial neural networks of the sort we were training were finally solving long-standing challenges in visual perception, speech recognition, game playing and many other domains. But it seemed absurd to me that a mere next-word predictor could ever truly understand new concepts, write jokes, debug code or do any of the myriad other things that are hallmarks of human intelligence.
‘Solving’ this kind of intelligence would surely require some fundamentally new scientific insight. And that would probably be inspired by neuroscience — the study of the only known embodiment of general intelligence, the brain.
Science and the new age of AI: a Nature special
My views back then were comfortably within the scientific mainstream, but in retrospect were also tinged with snobbery. My training was in physics and computational neuroscience, and I found the Silicon Valley hype distasteful at times. The cultish view that Moore’s law — the observation that computing power increases exponentially over time — would solve not only every technological problem, but also all social and scientific ones, seemed naive to me. It was the epitome of the mindset “when you have a hammer, everything looks like a nail”.
I was wrong. In 2019, colleagues at Google trained a massive (for the time) next-word predictor — technically, a next-token predictor, with each token corresponding to a word fragment — codenamed Meena1. It seemed able, albeit haltingly, to understand new concepts, write jokes, make logical arguments and much else. Meena’s scaled-up successor, LaMDA, did better2. This trend has continued since. In 2025, we find ourselves in the rather comical situation of expecting such large language models to respond fluently, intelligently, responsibly and accurately to all sorts of esoteric questions and demands that humans would fail to answer. People get irritated when these systems fail to answer appropriately — while simultaneously debating when artificial general intelligence will arrive.
How to detect consciousness in people, animals and maybe even AI
Large language models can be unreliable and say dumb things, but then, so can humans. Their strengths and weaknesses are certainly different from ours. But we are running out of intelligence tests that humans can pass reliably and AI models cannot. By those benchmarks, and if we accept that intelligence is essentially computational — the view held by most computational neuroscientists — we must accept that a working ‘simulation’ of intelligence actually is intelligence. There was no profound discovery that suddenly made obviously non-intelligent machines intelligent: it did turn out to be a matter of scaling computation.
Other researchers disagree with my assessment of where we are with AI. But in what follows, I want to accept the premise that intelligent machines are already here, and turn the mirror back on ourselves. If scaling up computation yields AI, could the kind of intelligence shown by living organisms, humans included, also be the result of computational scaling? If so, what drove that — and how did living organisms become computational in the first place?
Over the past several years, a growing group of collaborators and I have begun to find some tentative, but exciting answers. AI, biological intelligence and, indeed, life itself might all have emerged from the same process. This insight could shed fresh light not just on AI, neuroscience and neurophilosophy, but also on theoretical biology, evolution and complexity science. Moreover, it would give us a glimpse of how human and machine intelligence are destined to co-evolve in the future.
Predictive brains
... continue reading