Daily briefing: OpenAI claims a huge maths breakthrough

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

OpenAI says that it has shown that the Navier-Stokes equations — which describe the motion of fluids — can break down. Whether or not this could happen is considered one of the trickiest problems in mathematics, and it was on a list of ‘Millennium Prize’ challenges. The US artificial-intelligence company published a preprint that describes a spinning fluid system that ramps up to infinite speeds in a finite amount of time — something that would be impossible in real life. Mathematicians have hailed the achievement, but some are getting weary of science-by-press-release. “I would much rather see a professional level preprint appear on a standard server such as the arXiv, or a talk presented by the authors at a conference,” says mathematician Terence Tao.

Meanwhile, the preprint kicked off a debate about who should get credit for the solution. OpenAI researcher Sébastien Bubeck was open about the fact that his team was spurred on by rumours that researchers at a competing company, Anthropic, were close to a solution of a Millennium Prize problem. Mathematician Tristan Buckmaster, who worked on the Navier-Stokes problem with Levent Alpöge, a researcher affiliated with Anthropic, as a “purely personal collaboration”, says he got in touch with OpenAI. The company offered to put his name on its preprint, he says — but only if Alpöge’s didn’t appear. The preprint was published with only the company listed as an author.

Nature | 7 min read & Science | 11 min read

Reference: OpenAI preprint (not peer reviewed)

They say Rome wasn’t built in a day, but the Moon might have taken just five hours to form from the debris of a collision between Earth and a Mars-sized space rock. Simulations have tended to focus on the size of proto-Earth and ‘Theia’, the body that evidence suggests once struck our home planet. If Earth and Theia were both hot and smooshy, the chunks that eventually formed the Moon would have stayed largely intact — making it form more quickly — than if the two smashees were cool.

Live Science | 8 min read

Reference: The Astrophysical Journal Letters paper

The AI company Google DeepMind has released an ‘atlas’ of the human genome that aims to guide scientists through our biological code. The AlphaGenome Atlas charts the effects of nine billion single-DNA-letter changes — every possible mutation of this kind — in the human genome using predictions generated by a DeepMind AI model. The atlas could help to diagnose rare, unexplained diseases and uncover the hidden biology of common illnesses and biological traits, say researchers. It could also reveal some of the hidden rules by which DNA sequences control gene activity.

Nature | 6 min read

Features & opinion

If a language model trained on the information Albert Einstein might have had — Western science up to 1911, say — could it reproduce creative breakthroughs such as his general theory of relativity? Demis Hassabis, the co-founder of Google DeepMind, suggested this could be “a good test for AGI”, referring to the nebulous concept of artificial general intelligence that is a goal for many in the AI industry. Scientists are giving this, and similar experiments, a go. But their early attempts reveal more about the limitations of current AI systems — and the data we have to train them — than about their strengths.

Nature | 15 min read

Scientists in the United States have bioengineered microorganisms that can eat plastic, detect landmines and extract rare metals, says biotechnology writer Niko McCarty. But the process of getting them out of the lab has been hamstrung by a 1976 US law, the Toxic Substances Control Act (TSCA), that was not written with biotechnology in mind. Some engineered microbes dodge the TSCA by falling under other regulators — as pesticides, or dietary supplements, for example. The rest go into “the black hole of biology” — but it doesn’t have to be that way, McCarty argues.

Works in Progress | 15 min read

Africa’s geological record is a trove of clues about ancient climates and the origins of humans, but much of it is poorly documented, writes geoscientist Kamaldeen Omosanya. Publications related to Africa are underrepresented in major geological databases. This can both leave sites of possible geological interest vulnerable to mining or construction and alienate African students from Earth sciences in their own countries. “The rocks already hold the stories,” Omosanya writes. “The task is to make sure people can find them, understand them and protect them.”

The Conversation | 6 min read