If you're paying attention to AI, then I recommend this DeepMind documentary.
It's especially suitable for family viewing, and you can casually discuss topics like AI, games, and thinking.
This documentary was screened at film festivals last year, but yesterday Google released the full film for free on YouTube.
https://www.youtube.com/watch?v=d95J8yzvjbQ
What's particularly interesting is the documentary's name: The Thinking Game. What is most admirable is not that Demis won the Nobel Prize, not Demis's IQ, but that perspective of 'treating the world as a thinking game'.
Feynman moment
If we view the entire human civilization as a long 'thinking game', this film tells the story of how a group of people disassembled this 'thinking game' itself and rewrote its rules.
For them, chessboards, arcade games, StarCraft, protein folding, even entire virtual worlds are merely props for training one thing: general intelligence.
Demis's life is like a trajectory strung together by the same question: teenage chess prodigy, enduring to the tenth hour in the hall, suddenly realizing that locking hundreds of top minds in one room just to determine a winner is wasting intellect; then turning to game design, using 'vomit chain reaction' in theme parks to realistically simulate human behavior; then studying the brain at Cambridge, drawing lessons from neuroscience, finally teaming up with Shane to found DeepMind in London, boldly declaring: we are going to build artificial general intelligence.
They chose a path that seemed 'unserious': first having the system learn to play Pong in Atari, dig side tunnels in Breakout (brick breaker), then play Go and StarCraft. Reinforcement learning, reward signals, the closed loop of environment and agent, repeatedly refined in these games: the system initially couldn't even block the ball, only knew higher scores are better, but eventually made moves in Go unimaginable even to pro players, and developed multi-task operations in StarCraft comparable to professionals. There's a brutal honesty here: don't tell the machine the rules, just give the goal, and it will find 'paths humans never thought of' to high scores on its own.
But what truly gives this route its weight is applying the same mindset to real scientific problems. Protein folding, a puzzle since the 1960s, tried by countless brilliant minds for decades, stuck on slow experiments and scarce data. In their first CASP competition, AlphaFold beat peers, but only 'solved slightly better a problem no one had solved well'. Reflecting later, the team was frank: we knew we were still poor, this was the highest rung, but the moon was still far away.
The turning point was the second 'all-out push': rewriting the data pipeline, truly feeding physics and biology structural knowledge into the model, letting the system not just mimic but internalize constraints. Finally in CASP14, their structural accuracy nearly equaled button-press experimental results; the host declared, after half a century, this problem is solved. The next step was more radical: since all known species' proteins can be predicted in a month, don't wait for submissions, fold all Earth's proteins, publish free, making it human biology's infrastructure.
Technology is never a neutral toy. The 'Sputnik moment' after AlphaGo, AI arms race metaphors, Manhattan Project parallels, explicit rejection of autonomous weapons, and contempt for 'move fast and break things' all remind us: creating a cognitive system stronger than humans isn't making a new app, but potentially rewriting history's dividing line. Even DeepMind's own people worry: perhaps what we'll do in future isn't convincing the world we've built intelligence, but explaining we haven't reached it yet.
The film's greatest value isn't idolizing DeepMind, but laying key questions on the table: what is 'general' learning ability? In a reward-only system, how are values encoded and amplified? Between scientific exploration and engineering sprints, how to pace without 'dying on the timeline'? With intelligence outsourced to machines, where should humans focus—to new problem awareness, new institutions and governance, or rewriting our values?
What we should truly emulate isn't some algorithm or product, but that 'world as thinking game' perspective: daring to admit we're still bad for now, adjust pace in failure, and in success, push results to all humanity without holding back.