Do neural networks learn to implement algorithms such as look-ahead or search "in the wild"? Or do they rely purely on collections of simple heuristics? We present evidence of learned look-ahead in the policy network of Leela Chess Zero, the currently strongest neural chess engine. We find that Leela internally represents future optimal moves and that these representations are crucial for its final output in certain board states. Concretely, we exploit the fact that Leela is a transformer that treats every chessboard square like a token in language models, and give three lines of evidence: (1) activations on certain squares of future moves are unusually important causally; (2) we find attention heads that move important information "forward and backward in time," e.g., from squares of future moves to squares of earlier ones; and (3) we train a simple probe that can predict the optimal move 2 turns ahead with 92% accuracy (in board states where Leela finds a single best line). These findings are an existence proof of learned look-ahead in neural networks and might be a step towards a better understanding of their capabilities.

People who think about the long-term existential risks of artificial intelligence sometimes discuss the notion of an "AI box." To prevent a superintelligent computer from starting a nuclear war or otherwise wreaking havoc, its minders would seal it off from direct interaction with the outside world by keeping it offline. The only output would be communication with its operators. But, people worry, it might still escape, not through hacking but through "social engineering"--manipulating someone into setting it free. Such a scenario dramatically played out in the 2014 sci-fi thriller Ex Machina, in which a wily imprisoned robot seduces a hapless human into helping it break out.

A new BBC game show is to pit families against one another with a series of brain tests designed to assess different types of intelligence - but are you up to passing any of the cognitive challenges? The Family Brain Games, hosted by Dara Ó Briain, will follow eight different families from all walks of life as they come together in a specially designed'games lab' to compete in the ultimate test of intelligence. Questions are designed to go beyond traditional ideas about how to measure brain power, such as individual IQ. They will instead explore other measures including verbal ability, memory ability and even group intelligence. Dr Adam Hampshire, a neuroscientist at Imperial College London, is one of the brains behind the tests.