AI & robotics briefing: AI helps to reveal first passages of ancient charred scroll

Hello Nature readers, would you like to get this Briefing in your inbox free every week? Sign up here.

Vision Pro: the dawn of spatial computers?

Apple’s recently released Vision Pro headset could open up possibilities in accessibility and medical research. The US$3,499 headset can create virtual overlays on the real world that users can navigate to with their eyes and interact with using hand gestures. Its incredibly realistic, near-real-time display makes it unique, say scientists. The headset could allow new ways for people with disabilities to use computers and help surgeons to perform operations. The device’s eye-tracking technology might even be capable of picking up early signs of a stroke or dementia. If spatial computing headsets become more widespread, they could change our society in the same way mobile phones have, says computer vision researcher Dima Damen.

Nature | 5 min read

Chatbot aces chemistry predictions

A general-purpose AI system can be turned into a chemistry specialist with only a little tweaking. Researchers added literature information about chemical compounds or materials to the training data of GPT-3, an early iteration of the large language model that powers ChatGPT. This allowed the system to predict properties of similar materials and reaction yields as well as, or better than, more specialized algorithms. “This greatly reduces the barrier for other chemists to benefit from machine learning in their domains,” says chemical engineer Andrew White.

Nature | 5 min read

Reference: Nature Machine Intelligence paper

GPT informs — and disinforms — better

The large language model GPT-3 produces more convincing disinformation than humans do. Almost 700 people were asked to judge whether tweets on topics such as climate change and COVID were written by people or AI, and whether or not they were accurate. False information slipped by partipants more often when it was generated by GPT-3. At the same time, people also found it easier to correctly identify true information in AI-generated tweets. GPT-3’s text is more structured, condensed and easier to process than human-written prose, explains ethics researcher and study co-author Giovanni Spitale. Meanwhile, the social media company Meta says it will start detecting and labelling AI-generated images on Facebook, Instagram and Threads.

MIT Technology Review | 4 min read & BBC | 4 min read

Reference: Science Advances paper

Passages of ancient scroll revealed with AI

Student researchers have used machine learning to read text hidden inside charred, unopenable scrolls from the ancient Roman city of Herculaneum. The algorithm was trained on tiny differences in texture where the ink had been, based on three-dimensional computed tomography scans of the scrolls. The newly revealed passages discuss sources of pleasure including music, the colour purple and the taste of capers.

Nature | 7 min read

Features & opinion

‘ChatGPT corrupted the whole process’

Conservation scientist Lizzie Wolkovich was shocked when a reviewer for her new paper casually accused her of fraud: her carefully composed prose was “obviously ChatGPT”. Although Wolkovich’s text-change history could disprove the baseless assumption, the episode “captured something both inherently broken about the peer-review process and about how AI could corrupt science without even trying”, she says. She suggests that the scientific community needs explicit standards about when and how AI can be used in the writing process.

Nature | 6 min read

Five of the best AI books

Novelist Stephen Marche offers his pick of essential AI books and why you should read them:

• The Alignment Problem (Brian Christian): a close examination of the scientists who witnessed the birth of AI

• Artificial Intelligence (Melanie Mitchell): a necessary history of the technology’s development

• The Algorithm (Hilke Schellmann): a (human resources) case study on how people are using and abusing AI

• Progressive Capitalism (Ro Khanna): prescriptions for a serious political response to AIAI 2041 (Kai-Fu Lee, Chen Qiufan): self-aware speculations about the future of AI

The New York Times | 10 min read

How AI might unlock animal languages

Researchers are using AI to listen in on wild animals — and even begin to decipher what they are saying. Key to this is large language models’ ability to translate by observing the relationship between words, rather than understanding the meaning of them. What remains unclear is whether it’s really possible to translate elephant rumbles or whale clicks into anything that makes sense to a human. “We want to ask animals, how do you feel today? Or what did you do yesterday?” says zoology professor Yossi Yovel. “The thing is, if animals aren’t talking about these things, there’s no way [for us] to talk to them about it.”

Financial Times | 15 min interactive scroll

Image of the week

This sped-up footage shows an AI robot that can open almost any door — a surprisingly difficult task for machines. Researchers first trained the system on a few examples of how to open certain doors and drawers. Then, they let it loose on campus to figure out the various knobs, handles and crash bars, which it usually did in less than an hour. “You want the robots to work autonomously… without relying on humans to keep giving examples at test time for every new kind of scenario that you’re in,” says computer scientist and study co-author Deepak Pathak. (New Scientist | 3 min read)

Reference: arXiv preprint