Hello Nature readers, would you like to get this Briefing in your inbox free every week? Sign up here.
AI outdoes humans on maths problems
An LLM has helped mathematicians to break new ground on combinatorics problems inspired by the card game Set. In the game, players must spot combinations of cards with certain symbols. This is the first time that an LLM-based system has gone beyond solving mathematics problems with known solutions. One important feature is that the AI system isn’t a black box: people can see and learn from successful programs. “What’s most exciting to me is modelling new modes of human–machine collaboration,” says mathematician and study co-author Jordan Ellenberg.
Chatbot keeps online discussions cool
A chatbot can help to keep political discussions civil. Around 1,500 participants with differing opinions on gun control were invited to have online conversations on the topic. In some conversation pairs, one person was offered optional support from an AI tool on how to make their reply friendlier — suggestions that were accepted two-thirds of the time. People who had the chatbot’s help were rated as more understanding and respectful, by four percentage points. And both conversation partners in the chatbot group “felt like the experience was better, which is something valuable in and of itself”, says psychologist and study co-author Ethan Busby.
Features & opinion
What you need to know about the AI Act
European Union member states and the European Parliament have hammered out the world’s first comprehensive effort to legislate AI: the AI Act. When voted through, the law will define staggered rules based on risk.
• ‘Unacceptable risk’ systems that are a threat to people — for example, aspects of mass surveillance such as real-time facial recognition — will be banned (although there are exceptions for some police and military applications).
• AI systems must respect EU copyright rules and make public the content that they use to train generative models.
• General-purpose tools, such as ChatGPT, that could be used for good or ill are assessed by how powerful they are. Those that were trained using computing power above a certain threshold face more obligations for transparency and reporting.
The legally binding rules put the EU at the forefront of AI regulation and could influence how similar laws develop elsewhere. But critics worry that the rules will stifle innovation. And tech companies do not have to implement the rules for two years — so they could be obsolete by the time they’re in force.
Is a toddler a stochastic parrot?
ChatGPT is a stochastic parrot: it is incredibly efficient at stitching together words according to probability and generating convincing language, without any understanding of its meaning. The way in which LLMs learn is unnervingly similar to the way her son does, writes illustrator and cartoonist Angie Wang. “Aren’t we, after all, just a wetware neural network, a complex electrochemical machine?” In her beautifully illustrated essay, Wang explores the feeling of the vertigo that comes with the ever-evolving flood of AI-produced content, and what it means to be human.