Efforts to understand the universe could get a boost from an AI developed by Google DeepMind. The algorithm, which can reduce unwanted noise by up to 100 times, could allow the Laser Interferometer Gravitational-Wave Observatory (LIGO) to spot a particular type of black hole that has so far eluded us.

LIGO is designed to detect the gravitational waves produced when objects such as black holes spiral into each other and collide. These waves cross the universe at the speed of light, but the fluctuations they cause in space-time are extremely small – . Since its first observations 10 years ago, LIGO has recorded such signals produced by nearly 100 black hole collisions.

To do so, the experiment consists of two observatories in the US, each with two arms 4 kilometres long that are set perpendicular to each other. Lasers are beamed down each arm, reflected by precise mirrors at the end and then compared using an interferometer. The length of the arms is changed by a tiny amount as gravitational waves wash over them, and this is carefully recorded to build a picture of the origin of these signals.

The problem is that such demanding accuracy is required that even distant ocean waves or clouds passing overhead can affect measurements. This noise can easily drown out signals, making some observations impossible. Dozens of major adjustments need to be made to filter out the worst of this noise, tweaking the orientation of mirrors and other equipment.

at the California Institute of Technology in Pasadena, who worked with DeepMind to develop the new AI technology, says that attempting to automate these adjustments can ironically create more noise. “That controls noise has been bedevilling us for decades and decades – everything in this field has been blocked,” says Adhikari. “How do you hold the mirrors so still without inducing noise? If you don’t control them, the mirrors swing all over the place, and if you control it too much, then it sort of buzzes around.”

at the University of Portsmouth in the UK was one of the scientists who used to manually make these tweaks at LIGO. “As you move one thing, something else goes, and something else goes and something else goes,” she says. “You’d spend forever tweaking.”

Subscriber-only newsletter

Sign up to Lost in Space-Time

Sign up to newsletter

DeepMind’s new Deep Loop Shaping AI aims to reduce the level of noise from adjusting the mirrors at LIGO by up to 100 times. The AI was trained in a simulation before testing in the real world, and is effectively tasked with achieving two goals: reducing noise and minimising the number of adjustments it makes. “Over time, by repeatedly doing it – it’s like hundreds and thousands of trials that are running in simulation – the controller will sort of find what works and what doesn’t work and find a really, really good policy,” says Jonas Buchli at DeepMind.

at the University of Birmingham, UK, who wasn’t involved in the research but works on LIGO, says the AI is exciting, although there are many hurdles yet to overcome.

Firstly, the technology has only been run for an hour in the real world on LIGO, so it needs to be shown that it can operate for weeks or even months at a time. Secondly, the technology has so far only been applied to one aspect of control, helping to stabilise the mirrors, and there are hundreds if not thousands of aspects it could conceivably be applied to.

“It’s clearly just the first step, but I still think it’s a very intriguing one. And clearly there is plenty of room for enormous progress,” says Vecchio.

If similar improvements could be made across the board, then he believes we could spot so-called intermediate-size black holes – for example those with masses around 1000 times that of our sun – a class of objects without any confirmed observations. The improvements would tend to occur on lower-frequency gravitational waves, where the length of wave is more susceptible to noise, and which are created by larger objects.

“We know black holes up to 100 solar masses. We know the black holes in our galaxy that are a million solar masses and above. What’s in between?” says Vecchio. “People think there will be black holes at all these different mass ranges, but nobody has got uncontroversial experimental observational evidence.”

Nuttall says that the new approach could also provide more detailed observation of the types of black hole we have already seen. “This is looking pretty damn good,” she says. “I’m super excited by this.”

Mysteries of the universe: Cheshire, England

Spend a weekend with some of the brightest minds in science, as you explore the mysteries of the universe in an exciting programme that includes an excursion to see the iconic Lovell Telescope.

Find out more

Google has been using artificial intelligence watermarking to automatically identify text generated by the company’s Gemini chatbot, making it easier to distinguish AI-generated content from human-written posts. That watermark system could help prevent misuse of the AI chatbots for misinformation and disinformation – not to mention cheating in school and business settings.

Now, the tech company is making an open-source version of its technique available so that other generative AI developers can similarly watermark the output from their own large language models, says at Google DeepMind, the company’s AI research team, which combines the former Google Brain and DeepMind labs. “While SynthID isn’t a silver bullet for identifying AI-generated content, it is an important building block for developing more reliable AI identification tools,” he says.

Independent researchers voiced similar optimism. “While no known watermarking method is foolproof, I really think this can help in catching some fraction of AI-generated misinformation, academic cheating and more,” says at The University of Texas at Austin, who previously worked on AI safety at OpenAI. “I hope that other large language model companies, including OpenAI and Anthropic, will follow DeepMind’s lead on this.”

In May of this year, Google DeepMind that it had implemented its SynthID method for watermarking AI-generated text and video from Google’s Gemini and Veo AI services, respectively. The company has now published a paper in the journal NatureÌęshowing how SynthID generally outperformed similar AI watermarking techniques for text. The comparison involved assessing how readily responses from various watermarked AI models could be detected.

In Google DeepMind’s AI watermarking approach, as the model generates a sequence of text, a “tournament sampling” algorithm subtly nudges it toward selecting certain word “tokens”, creating a statistical signature that is detectable by associated software. This process randomly pairs up possible word tokens in a tournament-style bracket, with the winner of each pair being determined by which one scores highest according to a watermarking function. The winners move through successive tournament rounds until just one remains – a “multi-layered approach” that “increases the complexity of any potential attempts to reverse-engineer or remove the watermark”, says at the University of Maryland.

A “determined adversary” with huge amounts of computational power could still remove such AI watermarks, says at Harvard University. But he described SynthID’s approach as making sense given the need for scalable watermarking in AI services.

Free newsletter

Sign up to The Weekly

Sign up to newsletter

The Google DeepMind researchers tested two versions of SynthID that represent trade-offs between making the watermark signature more detectable, at the expense of distorting the text typically generated by an AI model. They showed that the non-distortionary version of the AI watermark still worked, without noticeably affecting the quality of 20 million Gemini-generated text responses during a live experiment.

But the researchers also acknowledged that the watermarking works best with longer chatbot responses that can be answered in a variety of ways – such as generating an essay or email – and said it has not yet been tested on responses to maths or coding problems.

Both Google DeepMind’s team and others described the need for additional safeguards against misuse of AI chatbots – with Huang recommending stronger regulation as well. “Mandating watermarking by law would address both the practicality and user adoption challenges, ensuring a more secure use of large language models,” she says.

Google DeepMind says its artificial intelligence has helped design chips that are already being used in data centres and even smartphones. But some chip design experts are sceptical of the company’s claims that such AI can plan new chip layouts better than humans can.

The newly named AlphaChip method can design “superhuman chip layouts” in hours, rather than relying on weeks or months of human effort, said and , researchers at Google DeepMind, in a . This AI approach uses reinforcement learning to figure out the relationships among chip components and gets rewarded based on the final layout quality. But independent researchers say the company has not yet proven such AI can outperform expert human chip designers or commercial software tools – and they want to see AlphaChip’s performance on public benchmarks involving current, state-of-the-art circuit designs.

“If Google would provide experimental results for these designs, we could have fair comparisons, and I expect that everyone would accept the results,” says at Binghamton University in New York. “The experiments would take at most a day or two to run, and Google has near-infinite resources – that these results have not been offered speaks volumes to me.”

Google DeepMind’s blog post accompanies an to Google’s 2021 Nature journal paper about the company’s AI process. Since that time, Google DeepMind says that AlphaChip has helped design three generations of Google’s Tensor Processing Units (TPU) – specialised chips used to train and run generative AI models for services such as Google’s Gemini chatbot.

The company also claims that the AI-assisted chip designs perform better than those designed by human experts and have been improving steadily. The AI achieves this by reducing the total length of wires required to connect chip components – a factor that can lower chip power consumption and potentially improve processing speed. And Google DeepMind says that AlphaChip has created layouts for general-purpose chips used in Google’s data centres, along with helping the company MediaTek develop a chip used in Samsung mobile phones.

“We really don’t know what AlphaChip is today, what it does and what it doesn’t do,” says , a chip design researcher at a competing firm. “We do know that reinforcement learning takes two to three orders of magnitude greater compute resources than methods used in commercial tools and is usually behind [in terms of] results.”

Free newsletter

Sign up to The Weekly

Sign up to newsletter

Markov and Madden critiqued theÌęoriginal paper’s claimsÌęabout AlphaChip outperforming unnamed human experts. “Comparisons to unnamed human designers are subjective, not reproducible, and very easy to game. The human designers may be applying low effort or be poorly qualified – there is no scientific result here,” says Markov. “Imagine if AlphaGo reported wins over unnamed Go players.” A Google DeepMind spokesperson described the experts as members of Google’s TPU chip design team using the best available commercial tools.

In 2023, an independent expert who had reviewed Google’s paper his Nature commentary article that had originally praised Google’s work but had also urged replication. That expert, at the University of California, San Diego, also ran a that tried to replicate Google’s AI method and found it did not consistently outperform a human expert or conventional computer algorithms. The best-performing methods used for comparison were commercial software or internal research tools for chip design from companies such as Cadence and NVIDIA. In a , Goldie and Mirhoseini disputed Kahng’s benchmarking results. They said his tests had not pretrained the AI method on specific chip designs – a crucial factor in its performance – and relied upon “far fewer compute resources” than Google DeepMind’s team to train the AI.Ìę

“On every benchmark where there’s what I would consider a fair comparison, it seems like reinforcement learning lags behind the state of the art by a wide margin,” says Madden. “For circuit placement, I don’t believe that it’s a promising research direction.”

An AI-generated recreation of the classic computer game Doom can be played normally despite having no computer code or graphics. Researchers behind the project say similar AI models could be used to create games from scratch in the future, just as they create text and images today.

The model, called , was made by Dani Valevski at Google Research and his colleagues, who declined to speak to żìĂš¶ÌÊÓÆ”. According to their on the research, the AI can be played for up to 20 seconds while retaining all the features of the original, such as scores, ammunition levels and map layouts. Players can attack enemies, open doors and interact with the environment as usual.

After this period, the model begins to run out of memory and the illusion falls apart.

The original Doom was released in 1993 and has become a popular subject for computer science projects in the years since, including attempts to such as toasters, treadmills and espresso machines.

But in all those cases, the hardware is simply running the original game’s code. What GameNGen does is fundamentally different: a type of AI called a neural network has learned by observation how to recreate the game without seeing any of its code.

The researchers first created an AI model that learned to interact with Doom as a human would. That model was then tasked with playing the game over and over again while a second AI model, based on the image generator, learned how hundreds of millions of inputs resulted in changes in the game state.

Free newsletter

Sign up to The Daily

Sign up to newsletter

That second model essentially then became a copy of the game, with all of the knowledge, rules and instructions from the original code encoded in the mysterious network of artificial neurons in its own architecture. In tests, human players were only slightly better than random chance at distinguishing short clips of the game from clips of the AI simulation.

GameNGen’s creators claim in their paper that it is a proof-of-concept for games being created by a neural network rather than lines of code. They suggest that games could be generated from text descriptions or concept art, which would make production less costly than using human programmers.

at the University of Surrey, UK, says the idea of getting a neural network to hallucinate a game environment, and the interactions a human has with it, is an interesting step forward, but not one that will replace human game designers.

“I don’t think it’s the end of those game studios. I think what the game studios have is the imagination, the skills, to actually create these worlds, to understand gameplay, to understand engagement, understand how to draw us into a story. It’s not just the nuts and bolts, the bits and bytes,” he says. “There’s something very human about creating engaging experiences that we as human beings enjoy that, at the moment, and for the foreseeable future, will largely come from other human beings.”

Google DeepMind claims to have made the first ever scientific discovery with an AI chatbot by building a fact-checker to filter out useless outputs, leaving only reliable solutions to mathematical or computing problems.

Previous DeepMind achievements, such as using AI to predict the weather or protein shapes, have relied on models created specifically for the task at hand, trained on accurate and specific data. Large language models (LLMs), such as GPT-4 and Google’s Gemini, are instead trained on vast amounts of varied data to create a breadth of abilities. But that approach also makes them susceptible to “hallucination”, a term researchers use for producing false outputs.

Gemini – which was released earlier this month – has already demonstrated a propensity for hallucination, getting even simple facts such as the . Google’s previous AI-powered search engine even made errors in the advertising material for its own launch.

One common fix for this phenomenon is to add a layer above the AI that verifies the accuracy of its outputs before passing them to the user. But creating a comprehensive safety net is an enormously difficult task given the broad range of topics that chatbots can be asked about.

at Google DeepMind and his colleagues have created a generalised LLM called FunSearch based on Google’s PaLM2 model with a fact-checking layer, which they call an “evaluator”. The model is constrained to providing computer code that solves problems in mathematics and computer science, which DeepMind says is a much more manageable task because these new ideas and solutions are inherently and quickly verifiable.

The underlying AI can still hallucinate and provide inaccurate or misleading results, but the evaluator filters out erroneous outputs and leaves only reliable, potentially useful concepts.

Free newsletter

Sign up to The Daily

Sign up to newsletter

“We think that perhaps 90 per cent of what the LLM outputs is not going to be useful,” says Fawzi. “Given a candidate solution, it’s very easy for me to tell you whether this is actually a correct solution and to evaluate the solution, but actually coming up with a solution is really hard. And so mathematics and computer science fit particularly well.”

DeepMind claims the model can generate new scientific knowledge and ideas – something LLMs haven’t done before.

To start with, FunSearch is given a problem and a very basic solution in source code as an input, then it generates a database of new solutions that are checked by the evaluator for accuracy. The best of the reliable solutions are given back to the LLM as inputs with a prompt asking it to improve on the ideas. DeepMind says the system produces millions of potential solutions, which eventually converge on an efficient result – sometimes surpassing the best known solution.

For mathematical problems, the model writes computer programs that can find solutions rather than trying to solve the problem directly.

Fawzi and his colleagues challenged FunSearch to find solutions to the cap set problem, which involves determining patterns of points where no three points make a straight line. The problem gets rapidly more computationally intensive as the number of points grows. The AI found a solution consisting of 512 points in eight dimensions, larger than any previously known.

When tasked with the bin-packing problem, where the aim is to efficiently place objects of various sizes into containers, FunSearch found solutions that outperform commonly used algorithms – a result that has immediate applications for transport and logistics companies. DeepMind says FunSearch could lead to improvements in many more mathematical and computing problems.

at the University of Birmingham, UK, says the next breakthroughs in AI won’t come from scaling-up LLMs to ever-larger sizes, but from adding layers that ensure accuracy, as DeepMind has done with FunSearch.

“The strength of a language model is its ability to imagine things, but the problem is hallucinations,” says Lee. “And this research is breaking that problem: it’s reining it in, or fact-checking. It’s a neat idea.”

Lee says AIs shouldn’t be criticised for producing large amounts of inaccurate or useless outputs, as this is not dissimilar to the way that human mathematicians and scientists operate: brainstorming ideas, testing them and following up on the best ones while discarding the worst.