Real-world problems have three properties that make them especially tricky: they are open-ended, uncertain and temporally extended. Open-ended problems are those for which the possible alternatives are virtually limitless. ... Uncertain problems can be blown off course by random events. … So real-world planning demands contingency measures.

Recommender systems are susceptible to a strange phenomenon called 'auto-induced distribution shift', whereby they can inadvertently manipulate the user as a side-effect of learning to maximize approval.

Remarkably, it turns out that you can make LLMs reason in a more goal-based way just by prompting them to think about a problem more deeply. A paper from 2022 introduced a new trick called 'chain-of-thought prompting', in which an LLM is shown a demo of how to 'think aloud' whilst solving a maths or reasoning problem.

When humans use language, the boundary between fact and fiction can be indistinct. Each of us sees the world through a different lens, so our words are inevitably grounded in a bespoke version of reality. A narrative that is fact to one person can seem like fiction to another, and vice versa. Even a supposedly impartial journalist describing an

During sleep or quiet downtime, memories that have been buffered in the hippocampus get replayed over and over - creating the sort of endless repetition that is needed to store information in the weights of a target neural network. In the biological case, that target is the neocortex, which is gradually 'trained' by replay activity to incorporate

the most important reason why AI systems are not like us (and probably never will be) is that they lack the visceral and emotional experiences that make us human. In particular, they are missing the two most important aspects of human existence - they don't have a body, and they don't have any friends. They are not motivated to feel or want like we

… when imagining worrisome AI capabilities, we don't have to think ahead to a time when 1.7 trillion parameters seem as dinky as 1.7 billion do today. Even current AI systems, equipped with diverse objectives and allowed to interact, have the potential to wreak havoc. When personal AI systems are deployed to buy and sell on eBay, send and receive

The intellectual histories of computer science and linguistics are thus studded with attempts to wrangle natural language into the straightjacket of formal language, all of which—like a flailing prisoner—it managed to resist.

The computational tricks we discussed above—non-linear transduction, compression, recurrent memory and attention—are ubiquitous tools in AI research. They have already proved their mettle in multiple non-linguistic domains, helping deep networks to recognize faces, generate videos and play board games at expert levels. Moreover, they are also