The world is awash in data about, well, the world — thanks to satellites and environmental sensors. But there’s still a lot we can’t see, and Fieldstone Bio thinks microbes can change that. “They’ve evolved to sense and respond to information. It’s just trillions of calculations going on at all times all around us,” Brandon Fields, Fieldstone Bio’s co-founder and chief science officer, told TechCrunch. “How do we take that and actually manipulate that to gain benefits for us?” Fieldstone’s technology emerged from that question. The startup was founded in 2023 after spinning out of MIT, where professor Chris Voigt’s lab had developed a way to turn microbes into sensors. The scientists programmed the microbes to change color when they encountered something of interest, whether it be nutrients in soil or landmines hidden in the dirt, and then figured out how to detect them. “The key technology out of Chris’ lab is this idea of, ‘How do we actually visualize these cells from really far away?’” Fields said. Fieldstone Bio recently raised $5 million in seed funding led by Ubiquity Ventures with participation from E14 and LDV Capital, the company exclusively told TechCrunch. The startup has been testing its technology in the lab, and the funding will let it test those microbes in the real world. Each strain is tailored to sense a particular compound, such as nitrogen on a farm field or TNT residue from a landmine.  “We isolate microbes from the environments we want to sense,” Fields said. “We build our sensors the DNA pieces, and we just drop them into these different ones and see which ones behave the best, which ones can last the longest.” Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW Once the microbes are ready, Fieldstone will broadcast them using drones. After the microbes have some time to sense their environment — several hours to days, depending on the target — the company will have another drone snap photos of the area. The images aren’t the usual aerial photography seen on Google Maps. Rather, they’re taken using what’s known as a hyperspectral camera, which divide visible and infrared light into as many as 600 different colors. Because Fieldstone’s microbes will reflect light at a very specific wavelength, it can train AI models to look for those signals amid a torrent of data. “That’s where the power of AI comes in, because we can start using that information to tease out these really faint signals to produce really cool heat maps of the microbe sensing the environment,” Fields said. In addition to agriculture and national security applications, Fieldstone is also programming microbes to detect environmental contaminants like arsenic, CEO Patrick Stone said. “Instead of going to do core soil samples over every 100 feet — and then you have 100 foot resolution — we could get a one-inch resolution and really map out exactly where they need to go clean up stuff,” he said. Gene edited microbial sensors broadcast over farm fields are sure to raise the eyebrows among people who oppose genetic modification. Fields said that the company has been in contact with the EPA to ensure that the company follows regulations. Fields said that, over time, he hopes the company’s database will become large enough that it can train models to associate other signals in the environment with whatever data is returned by the microbes. That would allow hyperspectral cameras to detect, say, arsenic contamination without needing to spread the engineered microbes. “Eventually, you don’t need to apply the microbe at all,” Fields said. “You have drones, planes, and satellites now collecting information about chemical information on a global scale.”

Should you care about the suffering of bugs? For most people, it’s a laughable question. But for those who really, really care about animal welfare, there’s a certain intellectual journey that might lead them to take it seriously. It goes something like this: First, they learn that the vast majority of the 84 billion birds and mammals raised for food are kept on factory farms, where animals are routinely mutilated and intensively confined. They become passionate advocates for these neglected and abused creatures. Then they learn that over 90 percent of those land animals are poultry birds — chickens and turkeys raised for meat, and hens raised for eggs — who are treated worse than pigs and cows, and have even fewer legal protections. They become, more or less, advocates for chickens. But then they might learn that fish and shrimp are farmed (or caught from the ocean) on an even greater scale — trillions annually compared to a measly 76 billion chickens. If their compassion for animals extends equally to marine life, they might come to advocate primarily for these sea creatures. Go even further, and they’ll discover the emerging industry of insect farming, which works much like chicken, pig, or fish factory farming, with the aim of producing as many animals as possible as cheaply as possible. On these insect factory farms, vast numbers of bugs are confined in trays or other containers until, at several weeks old, they’re frozen, cooked, shredded, or suffocated alive. Most are then sold as feed for farmed fish, poultry, and pigs, as food for pets, or to a lesser degree, direct human consumption. For the animal advocates who take this journey and wind up at the bottom of this animal suffering rabbit hole, a new report from the research organization Rethink Priorities will be pure nightmare fuel. According to the group, humanity is on track to farm and kill nearly 6 trillion animals annually by 2033, a near-quadrupling from 2023. And almost all of the growth in animal farming will come from tiny animals: shrimp, fish, and most of all, two insect species (mealworms and black soldier fly larvae). While humans farm and slaughter an astonishing 3 billion pigs, sheep, goats, and cattle each year, these animals are so dwarfed in numbers by farmed chicken, fish, and bugs that Rethink Priorities didn’t even include them in its calculation, nor did it include the 1 to 2 trillion wild fish scooped out of the ocean every year. The forecast starkly illustrates how a transformation in global agriculture patterns have ratcheted up animal suffering to mind-boggling proportions. The reason is that we’re increasingly eating really small animals. In the 1990s, chicken overtook beef as America’s meat of choice, and US chicken consumption continues to climb every year. And it takes about 127 chickens to produce the same amount of meat as one cow, because cows are enormous, while chickens weigh only about 6 pounds at slaughter. So, as Americans shifted toward eating animal species that are smaller in size, the total number of animals raised on US factory farms shot up. The same logic applies to an even greater extent to fish and shrimp — you’d have to kill about 28,500 shrimps to get the same amount of meat as you would from one cow. These animals are being farmed and eaten in increasingly massive numbers around the world, with both fish and shrimp typically confined in crowded, disease-ridden ponds or tanks that animal advocates liken to underwater factory farms. The world now eats more fish from these farms than from the ocean. Even worse, small animals, like chickens, fish, insects, and shrimp, tend to be treated worse and have fewer protections than larger animals like pigs and cattle.Concern for the welfare of insects — and even fish and shrimp — might bemuse or even offend many people. Humans already kill untold numbers of bugs annually by simply going about our daily business — driving, walking, exterminating ant infestations from our homes, and spraying pesticides on our crops. Americans eat tens of billions of individual shrimps each year with virtually no worry that they might feel pain. While farmed chickens and pigs have received the sympathetic Hollywood treatment, like the Chicken Run movies, Charlotte’s Web, and Okja, similar films about shrimp or mealworms don’t seem to be in the offing. But Rethink Priorities, along with a growing chorus of scientists and philosophers, believe that invertebrates like shrimp and insects could be sentient, meaning they possess the capacity for pain, pleasure, and other sensations. They’re not arguing that these animals are equivalent to a chicken, cow, or human, but that they may be worth some moral consideration given emerging research on their potential for sentience and the massive scale on which they’re farmed.History has long shown us that today’s laughable moral concern could be tomorrow’s tragedy. That could be the case for these tiny, unfamiliar, uncharismatic animals the more we come to understand who they are and what they might be capable of feeling. What can a shrimp or an insect feel?There had long been relatively little research into whether invertebrates like shrimp and insects are sentient, but that’s begun to change in recent years.“Evidence is building that there’s a form of sentience there in insects,” Jonathan Birch, a philosopher at the London School of Economics who leads the Foundations of Animal Sentience project at the university, told me. Historically, this line of inquiry has focused on bees, he said, who have demonstrated signs of sentience by engaging in wound-tending behavior, complex decision-making in weighing pain versus pleasure, and even play. Some research has shown that fruit flies may have the capacity to feel pain and enjoy play.According to Birch and several of his colleagues, adult flies and mosquitoes, along with cockroaches and termites, satisfy six of eight key criteria for sentience, while several other orders satisfy three to four. He’s now collaborating with researchers to study pain indicators in black soldier fly larvae and crickets.What little research has looked at shrimp sentience has found mixed results, and much of it has been conducted on Caridean shrimp, not penaeid shrimp, the group that’s most commonly farmed (shrimp are not a single species, but a massive category comprising more than 2,000 known species across several taxonomic groups).Fish thrash inside a tightening net on a fish farm. Havva Zorlu/We AnimalsSome research has shown that shrimp have nociceptors, sensory neurons that detect and respond to potentially harmful stimuli — an important indicator of sentience — but their efforts to avoid threats could be merely reflexive. In one study, shrimp engaged in wound-tending behavior when researchers poured acid onto their antennae, but when they treated it with an anesthetic, the shrimp increased that behavior. This suggests they may not feel pain the way other animals do because the pain relief should have caused them to reduce wound tending.But Birch believes there is strong evidence of sentience in another crustacean species: the crab.Despite our limited understanding of invertebrates’ capacity for pain, our understanding of other animals’ capacities can quickly evolve. Just a couple of decades ago, it was largely thought that fish (which are not invertebrates) couldn’t feel pain, but scientific consensus has significantly shifted toward the belief that they can — and that they could experience many other physical and mental states. Given everything humans have learned about animals’ capacities in recent decades, there’s a strong argument to be made in favor of assuming that other animals are sentient unless proved otherwise, rather than assuming that they aren’t as the starting point.And if shrimp and insects are sentient, it would exacerbate an already emergency situation for global animal welfare, raising the number of farmed sentient animals by well over a trillion creatures today, and potentially many trillions in the decade ahead. The proposition to include these animals in humanity’s moral circle can lead some, Birch said, to throw up their hands in exasperation. “We lack ethical frameworks that tell us how to think about them, but to me, that’s not an excuse for ignoring the issue,” Birch said. “I think people sometimes imagine, well, if the sentient world is so vast — if all ways of feeding ourselves cause harm — then there can’t be any ethical constraints. And I think that’s entirely wrong. I think we do need to take the harm seriously, and think about what we might do to conduct ourselves more ethically.”A question of strategyMost people who advocate for factory-farmed animals focus on pigs, cows, chickens, and turkeys. Only the most quantitatively minded number-crunchers, like Birch and the folks at Rethink Priorities, look at the data and focus on fish, shrimp, and insects.Some in the animal advocacy movement might consider this expansion of moral concern — especially for insects — a major strategic error, one that will make an already fringe movement seem even more strange and scolding. It’s something, if I’m being honest, I’ve felt myself. A worker disinfects crickets’ watering trays in the final grow room at Entomo Farms in Ontario, Canada. James MacDonald/Bloomberg via Getty ImagesBut the general public might be more open to having some ethical consideration for these animals than we might think. Hannah McKay, a research analyst at Rethink Priorities who co-authored the new report, pointed me to recent surveys from the UK and Brazil in which a majority of participants said that they believed that shrimp can feel pain. In the UK and the Netherlands, food companies are phasing out particularly cruel practices in their shrimp supply chains, “even in the absence of high public pressure,” McKay said. Last month, a post about eyestalk ablation — the common, disturbing practice of tearing out female shrimp’s eyes to make them breed faster — made it to the front page of Reddit, whose users were overwhelmingly horrified by it.In 2022, the UK passed the Animal Welfare (Sentience) Act, which included decapod crustaceans (shrimp, lobsters, crabs, and crayfish) among the animals that should be considered sentient, in large part on the advice of Birch and colleagues.The aim of Rethink Priorities, Birch, and their ilk isn’t necessarily to start a campaign for worldwide shrimp and insect liberation, but rather to, at the very least, secure some minimum welfare standards.Many animal advocates today wish that their predecessors in the 1950s, who were more focused on the welfare of pets, had devoted more attention to cows, pigs, and chickens. Instead, the quiet rise of factory farming in mid-20th-century America went largely unchallenged, and has now led to the confinement, abuse, and slaughter of tens of billions of mammals and birds each year. Sagar Shah, a senior researcher at Rethink Priorities who co-authored the report with McKay, feels the same way about fish and shrimp farming, telling me that if he could turn the clock back 30 years to when these industries were relatively small, he would’ve pushed for “more resources into thinking about these questions: Are [fish and shrimp] sentient? What does good welfare mean for these animals if we’re going to use them?”“Collectively, we missed the boat a bit, and we’ve already got a huge scale of farming for fish and shrimp, and we’re catching up now,” Shah said. “But for insects, the industry is in its infancy, and that means we’ve got an opportunity to figure out what good welfare means and shape the growth of the industry.” It’s a cause that few animal advocates are willing to dedicate themselves to, and it may not win them many allies anytime soon. But that doesn’t deter Shah, who says “that doesn’t mean we shouldn’t try.” You’ve read 1 article in the last monthHere at Vox, we're unwavering in our commitment to covering the issues that matter most to you — threats to democracy, immigration, reproductive rights, the environment, and the rising polarization across this country.Our mission is to provide clear, accessible journalism that empowers you to stay informed and engaged in shaping our world. By becoming a Vox Member, you directly strengthen our ability to deliver in-depth, independent reporting that drives meaningful change.We rely on readers like you — join us.Swati SharmaVox Editor-in-ChiefSee More:

The data quality used in pretraining LLMs has become increasingly critical to their success. To build information-rich corpora, researchers have moved from heuristic filtering methods, such as rule-based noise removal and deduplication, to model-driven filtering, which leverages neural classifiers to identify high-quality samples. Despite its benefits, this approach still faces key issues: it lacks efficient validation mechanisms to assess data quality promptly and often relies on manually curated seed datasets that introduce subjectivity. While early datasets like C4 and Pile laid the groundwork for model development, recent efforts like RefinedWeb, Dolma, and DCLM have scaled significantly, incorporating up to trillions of tokens. Model-driven filtering has gained traction in these newer corpora for its ability to refine massive datasets and enhance LLM performance across downstream tasks. Nevertheless, the effectiveness of model-driven filtering is limited by the high costs and inefficiencies of current validation methods and the absence of clear standards for seed data selection. Recent datasets, such as FineWeb-edu and Ultra-FineWeb, have demonstrated improved model performance by using multiple classifiers to cross-verify data quality. These datasets outperform previous versions on benchmarks like MMLU, ARC, and C-Eval, indicating that refined filtering methods can enhance English and Chinese understanding. To further optimize this process, some studies propose using LLMs for multi-dimensional data evaluation via prompts or leveraging token-level perplexity scores. These innovations aim to lower computational overhead while improving data quality, ultimately enabling more effective training with fewer tokens.  Researchers from ModelBest Inc., Tsinghua University, and Soochow University developed an efficient data filtering pipeline to improve LLM training. They introduced a verification strategy that uses a nearly-trained LLM to evaluate new data by observing performance gains during final training steps, reducing computational costs. A lightweight fastText-based classifier further enhances filtering speed and accuracy. Applied to FineWeb and Chinese FineWeb datasets, this method produced the Ultra-FineWeb dataset, containing 1 trillion English and 120 billion Chinese tokens. LLMs trained on Ultra-FineWeb showed notable performance gains, confirming the pipeline’s effectiveness in improving data quality and training efficiency.  The study outlines an efficient, high-quality data filtering pipeline to reduce computational costs while maintaining data integrity. It begins by using a cost-effective verification strategy to select reliable seed samples from a candidate pool, which are then used to train a data classifier. Positive seeds are sourced from LLM annotations, curated datasets, textbooks, and synthesized content, while negatives come from diverse corpora. Classifier training avoids over-iteration, focusing instead on high-quality seed selection. A fastText-based classifier is used for scalable filtering, offering competitive performance at significantly lower inference costs compared to LLM-based methods, with preprocessing steps ensuring balanced, clean data input.  The models were trained using MegatronLM with the MiniCPM-1.2 B architecture on 100B tokens. Evaluations used Lighteval across English and Chinese benchmarks. The results show that models trained on Ultra-FineWeb consistently outperformed those trained on FineWeb and FineWeb-edu, individually and in mixed-language settings. Ultra-FineWeb-en achieved the highest English average score, while Ultra-FineWeb-zh improved performance on Chinese tasks. Ablation studies revealed that Ultra-FineWeb maintains balanced token lengths and benefits from efficient filtering strategies, highlighting its superior quality and effectiveness in improving model performance.  In conclusion, the study presents Ultra-FineWeb, a high-quality multilingual dataset comprising about 1 trillion English tokens and 120 billion Chinese tokens. Built upon FineWeb and Chinese FineWeb, it leverages a novel, efficient data filtering pipeline featuring a fastText-based lightweight classifier and a low-cost verification strategy. The pipeline enhances filtering accuracy, reduces reliance on manual seed data selection, and ensures robust performance with minimal computational overhead. Experimental results show that models trained on Ultra-FineWeb consistently outperform those trained on earlier datasets, demonstrating improved performance across benchmarks. The methodology ensures reproducibility and offers valuable insights for optimizing data quality in future LLM training.  Check out theTwitter and don’t forget to join our 90k+ ML SubReddit. Sana HassanSana Hassan, a consulting intern at Marktechpost and dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a fresh perspective to the intersection of AI and real-life solutions.Sana Hassanhttps://www.marktechpost.com/author/sana-hassan/Coding Agents See 75% Surge: SimilarWeb’s AI Usage Report Highlights the Sectors Winning and Losing in 2025’s Generative AI BoomSana Hassanhttps://www.marktechpost.com/author/sana-hassan/Rethinking Toxic Data in LLM Pretraining: A Co-Design Approach for Improved Steerability and DetoxificationSana Hassanhttps://www.marktechpost.com/author/sana-hassan/A Step-by-Step Guide on Building, Customizing, and Publishing an AI-Focused Blogging Website with Lovable.dev and Seamless GitHub IntegrationSana Hassanhttps://www.marktechpost.com/author/sana-hassan/NVIDIA AI Introduces Audio-SDS: A Unified Diffusion-Based Framework for Prompt-Guided Audio Synthesis and Source Separation without Specialized Datasets

Google DeepMind has once again used large language models to discover new solutions to long-standing problems in math and computer science. This time the firm has shown that its approach can not only tackle unsolved theoretical puzzles, but improve a range of important real-world processes as well. Google DeepMind's new tool, called AlphaEvolve, uses the Gemini 2.0 family of large language models (LLMs) to produce code for a wide range of different tasks. LLMs are known to be hit and miss at coding. The twist here is that AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process, until it has produced the best algorithm it can. In many cases, the results are more efficient or more accurate than the best existing (human-written) solutions. “You can see it as a sort of super coding agent,” says Pushmeet Kohli, a vice president at Google DeepMind who leads its AI for Science teams. “It doesn’t just propose a piece of code or an edit, it actually produces a result that maybe nobody was aware of.” In particular, AlphaEvolve came up with a way to improve the software Google uses to allocate jobs to its many millions of servers around the world. Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google's total computing resources. That might not sound like much, but at Google’s scale it’s huge. Jakob Moosbauer, a mathematician at the University of Warwick in the UK, is impressed. He says the way AlphaEvolve searches for algorithms that produce specific solutions—rather than searching for the solutions themselves—makes it especially powerful. “It makes the approach applicable to such a wide range of problems,” he says. “AI is becoming a tool that will be essential in mathematics and computer science.” AlphaEvolve continues a line of work that Google DeepMind has been pursuing for years. Its vision is that AI can help to advance human knowledge across math and science. In 2022, it developed AlphaTensor, a model that found a faster way to solve matrix multiplications—a fundamental problem in computer science—beating a record that had stood for more than 50 years. In 2023, it revealed AlphaDev, which discovered faster ways to perform a number of basic calculations performed by computers trillions of times a day. AlphaTensor and AlphaDev both turn math problems into a kind of game, then search for a winning series of moves. FunSearch, which arrived in late 2023, swapped out game-playing AI and replaced it with LLMs that can generate code. Because LLMs can carry out a range of tasks, FunSearch can take on a wider variety of problems than its predecessors, which were trained to play just one type of game. The tool was used to crack a famous unsolved problem in pure mathematics. AlphaEvolve is the next generation of FunSearch. Instead of coming up with short snippets of code to solve a specific problem, as FunSearch did, it can produce programs that are hundreds of lines long. This makes it applicable to a much wider variety of problems.     In theory, AlphaEvolve could be applied to any problem that can be described in code and that has solutions that can be evaluated by a computer. “Algorithms run the world around us, so the impact of that is huge,” says Matej Balog, a researcher at Google DeepMind who leads the algorithm discovery team. Survival of the fittest Here’s how it works: AlphaEvolve can be prompted like any LLM. Give it a description of the problem and any extra hints you want, such as previous solutions, and AlphaEvolve will get Gemini 2.0 Flash (the smallest, fastest version of Google DeepMind’s flagship LLM) to generate multiple blocks of code to solve the problem. It then takes these candidate solutions, runs them to see how accurate or efficient they are, and scores them according to a range of relevant metrics. Does this code produce the correct result? Does it run faster than previous solutions? And so on. AlphaEvolve then takes the best of the current batch of solutions and asks Gemini to improve them. Sometimes AlphaEvolve will throw a previous solution back into the mix to prevent Gemini from hitting a dead end. When it gets stuck, AlphaEvolve can also call on Gemini 2.0 Pro, the most powerful of Google DeepMind’s LLMs. The idea is to generate many solutions with the faster Flash but add solutions from the slower Pro when needed. These rounds of generation, scoring, and regeneration continue until Gemini fails to come up with anything better than what it already has. Number games The team tested AlphaEvolve on a range of different problems. For example, they looked at matrix multiplication again to see how a general-purpose tool like AlphaEvolve compared to the specialized AlphaTensor. Matrices are grids of numbers. Matrix multiplication is a basic computation that underpins many applications, from AI to computer graphics, yet nobody knows the fastest way to do it. “It’s kind of unbelievable that it’s still an open question,” says Balog. The team gave AlphaEvolve a description of the problem and an example of a standard algorithm for solving it. The tool not only produced new algorithms that could calculate 14 different sizes of matrix faster than any existing approach, it also improved on AlphaTensor’s record-beating result for multipying two four-by-four matrices. AlphaEvolve scored 16,000 candidates suggested by Gemini to find the winning solution, but that’s still more efficient than AlphaTensor, says Balog. AlphaTensor’s solution also only worked when a matrix was filled with 0s and 1s. AlphaEvolve solves the problem with other numbers too. “The result on matrix multiplication is very impressive,” says Moosbauer. “This new algorithm has the potential to speed up computations in practice.” Manuel Kauers, a mathematician at Johannes Kepler University in Linz, Austria, agrees: “The improvement for matrices is likely to have practical relevance.” By coincidence, Kauers and a colleague have just used a different computational technique to find some of the speedups AlphaEvolve came up with. The pair posted a paper online reporting their results last week. “It is great to see that we are moving forward with the understanding of matrix multiplication,” says Kauers. “Every technique that helps is a welcome contribution to this effort.” Real-world problems Matrix multiplication was just one breakthrough. In total, Google DeepMind tested AlphaEvolve on more than 50 different types of well-known math puzzles, including problems in Fourier analysis (the math behind data compression, essential to applications such as video streaming), the minimum overlap problem (an open problem in number theory proposed by mathematician Paul Erdős in 1955), and kissing numbers (a problem introduced by Isaac Newton that has applications in materials science, chemistry, and cryptography). AlphaEvolve matched the best existing solutions in 75% of cases and found better solutions in 20% of cases.   Google DeepMind then applied AlphaEvolve to a handful of real-world problems. As well as coming up with a more efficient algorithm for managing computational resources across data centers, the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. AlphaEvolve even found a way to speed up the training of Gemini itself, by producing a more efficient algorithm for managing a certain type of computation used in the training process. Google DeepMind plans to continue exploring potential applications of its tool. One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person, such as lab experiments that are subject to interpretation.    Moosbauer also points out that while AlphaEvolve may produce impressive new results across a wide range of problems, it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding.   Even so, tools like AlphaEvolve are set to change the way researchers work. “I don't think we are finished,” says Kohli. “There is much further that we can go in terms of how powerful this type of approach is.”

Google DeepMind has once again used large language models to discover new solutions to long-standing problems in math and computer science. This time the firm has shown that its approach can not only tackle unsolved theoretical puzzles, but improve a range of important real-world processes as well. Google DeepMind's new tool, called AlphaEvolve, uses the Gemini 2.0 family of large language models (LLMs) to produce code for a wide range of different tasks. LLMs are known to be hit and miss at coding. The twist here is that AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process, until it has produced the best algorithm it can. In many cases, the results are more efficient or more accurate than the best existing (human-written) solutions. “You can see it as a sort of super coding agent,” says Pushmeet Kohli, a vice president at Google DeepMind who leads its AI for Science teams. “It doesn’t just propose a piece of code or an edit, it actually produces a result that maybe nobody was aware of.” In particular, AlphaEvolve came up with a way to improve the software Google uses to allocate jobs to its many millions of servers around the world. Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google's total computing resources. That might not sound like much, but at Google’s scale it’s huge. Jakob Moosbauer, a mathematician at the University of Warwick in the UK, is impressed. He says the way AlphaEvolve searches for algorithms that produce specific solutions—rather than searching for the solutions themselves—makes it especially powerful. “It makes the approach applicable to such a wide range of problems,” he says. “AI is becoming a tool that will be essential in mathematics and computer science.” AlphaEvolve continues a line of work that Google DeepMind has been pursuing for years. Its vision is that AI can help to advance human knowledge across math and science. In 2022, it developed AlphaTensor, a model that found a faster way to solve matrix multiplications—a fundamental problem in computer science—beating a record that had stood for more than 50 years. In 2023, it revealed AlphaDev, which discovered faster ways to perform a number of basic calculations performed by computers trillions of times a day. AlphaTensor and AlphaDev both turn math problems into a kind of game, then search for a winning series of moves. FunSearch, which arrived in late 2023, swapped out game-playing AI and replaced it with LLMs that can generate code. Because LLMs can carry out a range of tasks, FunSearch can take on a wider variety of problems than its predecessors, which were trained to play just one type of game. The tool was used to crack a famous unsolved problem in pure mathematics. AlphaEvolve is the next generation of FunSearch. Instead of coming up with short snippets of code to solve a specific problem, as FunSearch did, it can produce programs that are hundreds of lines long. This makes it applicable to a much wider variety of problems.     In theory, AlphaEvolve could be applied to any problem that can be described in code and that has solutions that can be evaluated by a computer. “Algorithms run the world around us, so the impact of that is huge,” says Matej Balog, a researcher at Google DeepMind who leads the algorithm discovery team. Survival of the fittest Here’s how it works: AlphaEvolve can be prompted like any LLM. Give it a description of the problem and any extra hints you want, such as previous solutions, and AlphaEvolve will get Gemini 2.0 Flash (the smallest, fastest version of Google DeepMind’s flagship LLM) to generate multiple blocks of code to solve the problem. It then takes these candidate solutions, runs them to see how accurate or efficient they are, and scores them according to a range of relevant metrics. Does this code produce the correct result? Does it run faster than previous solutions? And so on. AlphaEvolve then takes the best of the current batch of solutions and asks Gemini to improve them. Sometimes AlphaEvolve will throw a previous solution back into the mix to prevent Gemini from hitting a dead end. When it gets stuck, AlphaEvolve can also call on Gemini 2.0 Pro, the most powerful of Google DeepMind’s LLMs. The idea is to generate many solutions with the faster Flash but add solutions from the slower Pro when needed. These rounds of generation, scoring, and regeneration continue until Gemini fails to come up with anything better than what it already has. Number games The team tested AlphaEvolve on a range of different problems. For example, they looked at matrix multiplication again to see how a general-purpose tool like AlphaEvolve compared to the specialized AlphaTensor. Matrices are grids of numbers. Matrix multiplication is a basic computation that underpins many applications, from AI to computer graphics, yet nobody knows the fastest way to do it. “It’s kind of unbelievable that it’s still an open question,” says Balog. The team gave AlphaEvolve a description of the problem and an example of a standard algorithm for solving it. The tool not only produced new algorithms that could calculate 14 different sizes of matrix faster than any existing approach, it also improved on AlphaTensor’s record-beating result for multipying two four-by-four matrices. AlphaEvolve scored 16,000 candidates suggested by Gemini to find the winning solution, but that’s still more efficient than AlphaTensor, says Balog. AlphaTensor’s solution also only worked when a matrix was filled with 0s and 1s. AlphaEvolve solves the problem with other numbers too. “The result on matrix multiplication is very impressive,” says Moosbauer. “This new algorithm has the potential to speed up computations in practice.” Manuel Kauers, a mathematician at Johannes Kepler University in Linz, Austria, agrees: “The improvement for matrices is likely to have practical relevance.” By coincidence, Kauers and a colleague have just used a different computational technique to find some of the speedups AlphaEvolve came up with. The pair posted a paper online reporting their results last week. “It is great to see that we are moving forward with the understanding of matrix multiplication,” says Kauers. “Every technique that helps is a welcome contribution to this effort.” Real-world problems Matrix multiplication was just one breakthrough. In total, Google DeepMind tested AlphaEvolve on more than 50 different types of well-known math puzzles, including problems in Fourier analysis (the math behind data compression, essential to applications such as video streaming), the minimum overlap problem (an open problem in number theory proposed by mathematician Paul Erdős in 1955), and kissing numbers (a problem introduced by Isaac Newton that has applications in materials science, chemistry, and cryptography). AlphaEvolve matched the best existing solutions in 75% of cases and found better solutions in 20% of cases.   Google DeepMind then applied AlphaEvolve to a handful of real-world problems. As well as coming up with a more efficient algorithm for managing computational resources across data centers, the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. AlphaEvolve even found a way to speed up the training of Gemini itself, by producing a more efficient algorithm for managing a certain type of computation used in the training process. Google DeepMind plans to continue exploring potential applications of its tool. One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person, such as lab experiments that are subject to interpretation.    Moosbauer also points out that while AlphaEvolve may produce impressive new results across a wide range of problems, it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding.   Even so, tools like AlphaEvolve are set to change the way researchers work. “I don't think we are finished,” says Kohli. “There is much further that we can go in terms of how powerful this type of approach is.”

Google DeepMind has once again used large language models to discover new solutions to long-standing problems in math and computer science. This time the firm has shown that its approach can not only tackle unsolved theoretical puzzles, but improve a range of important real-world processes as well. Google DeepMind’s new tool, called AlphaEvolve, uses the Gemini 2.0 family of large language models (LLMs) to produce code for a wide range of different tasks. LLMs are known to be hit and miss at coding. The twist here is that AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process, until it has produced the best algorithm it can. In many cases, the results are more efficient or more accurate than the best existing (human-written) solutions. “You can see it as a sort of super coding agent,” says Pushmeet Kohli, a vice president at Google DeepMind who leads its AI for Science teams. “It doesn’t just propose a piece of code or an edit, it actually produces a result that maybe nobody was aware of.” In particular, AlphaEvolve came up with a way to improve the software Google uses to allocate jobs to its many millions of servers around the world. Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google’s total computing resources. That might not sound like much, but at Google’s scale it’s huge. Jakob Moosbauer, a mathematician at the University of Warwick in the UK, is impressed. He says the way AlphaEvolve searches for algorithms that produce specific solutions—rather than searching for the solutions themselves—makes it especially powerful. “It makes the approach applicable to such a wide range of problems,” he says. “AI is becoming a tool that will be essential in mathematics and computer science.” AlphaEvolve continues a line of work that Google DeepMind has been pursuing for years. Its vision is that AI can help to advance human knowledge across math and science. In 2022, it developed AlphaTensor, a model that found a faster way to solve matrix multiplications—a fundamental problem in computer science—beating a record that had stood for more than 50 years. In 2023, it revealed AlphaDev, which discovered faster ways to perform a number of basic calculations performed by computers trillions of times a day. AlphaTensor and AlphaDev both turn math problems into a kind of game, then search for a winning series of moves. FunSearch, which arrived in late 2023, swapped out game-playing AI and replaced it with LLMs that can generate code. Because LLMs can carry out a range of tasks, FunSearch can take on a wider variety of problems than its predecessors, which were trained to play just one type of game. The tool was used to crack a famous unsolved problem in pure mathematics. AlphaEvolve is the next generation of FunSearch. Instead of coming up with short snippets of code to solve a specific problem, as FunSearch did, it can produce programs that are hundreds of lines long. This makes it applicable to a much wider variety of problems.     In theory, AlphaEvolve could be applied to any problem that can be described in code and that has solutions that can be evaluated by a computer. “Algorithms run the world around us, so the impact of that is huge,” says Matej Balog, a researcher at Google DeepMind who leads the algorithm discovery team. Survival of the fittest Here’s how it works: AlphaEvolve can be prompted like any LLM. Give it a description of the problem and any extra hints you want, such as previous solutions, and AlphaEvolve will get Gemini 2.0 Flash (the smallest, fastest version of Google DeepMind’s flagship LLM) to generate multiple blocks of code to solve the problem. It then takes these candidate solutions, runs them to see how accurate or efficient they are, and scores them according to a range of relevant metrics. Does this code produce the correct result? Does it run faster than previous solutions? And so on. AlphaEvolve then takes the best of the current batch of solutions and asks Gemini to improve them. Sometimes AlphaEvolve will throw a previous solution back into the mix to prevent Gemini from hitting a dead end. When it gets stuck, AlphaEvolve can also call on Gemini 2.0 Pro, the most powerful of Google DeepMind’s LLMs. The idea is to generate many solutions with the faster Flash but add solutions from the slower Pro when needed. These rounds of generation, scoring, and regeneration continue until Gemini fails to come up with anything better than what it already has. Number games The team tested AlphaEvolve on a range of different problems. For example, they looked at matrix multiplication again to see how a general-purpose tool like AlphaEvolve compared to the specialized AlphaTensor. Matrices are grids of numbers. Matrix multiplication is a basic computation that underpins many applications, from AI to computer graphics, yet nobody knows the fastest way to do it. “It’s kind of unbelievable that it’s still an open question,” says Balog. The team gave AlphaEvolve a description of the problem and an example of a standard algorithm for solving it. The tool not only produced new algorithms that could calculate 14 different sizes of matrix faster than any existing approach, it also improved on AlphaTensor’s record-beating result for multipying two four-by-four matrices. AlphaEvolve scored 16,000 candidates suggested by Gemini to find the winning solution, but that’s still more efficient than AlphaTensor, says Balog. AlphaTensor’s solution also only worked when a matrix was filled with 0s and 1s. AlphaEvolve solves the problem with other numbers too. “The result on matrix multiplication is very impressive,” says Moosbauer. “This new algorithm has the potential to speed up computations in practice.” Manuel Kauers, a mathematician at Johannes Kepler University in Linz, Austria, agrees: “The improvement for matrices is likely to have practical relevance.” By coincidence, Kauers and a colleague have just used a different computational technique to find some of the speedups AlphaEvolve came up with. The pair posted a paper online reporting their results last week. “It is great to see that we are moving forward with the understanding of matrix multiplication,” says Kauers. “Every technique that helps is a welcome contribution to this effort.” Real-world problems Matrix multiplication was just one breakthrough. In total, Google DeepMind tested AlphaEvolve on more than 50 different types of well-known math puzzles, including problems in Fourier analysis (the math behind data compression, essential to applications such as video streaming), the minimum overlap problem (an open problem in number theory proposed by mathematician Paul Erdős in 1955), and kissing numbers (a problem introduced by Isaac Newton that has applications in materials science, chemistry, and cryptography). AlphaEvolve matched the best existing solutions in 75% of cases and found better solutions in 20% of cases.   Google DeepMind then applied AlphaEvolve to a handful of real-world problems. As well as coming up with a more efficient algorithm for managing computational resources across data centers, the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. AlphaEvolve even found a way to speed up the training of Gemini itself, by producing a more efficient algorithm for managing a certain type of computation used in the training process. Google DeepMind plans to continue exploring potential applications of its tool. One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person, such as lab experiments that are subject to interpretation.    Moosbauer also points out that while AlphaEvolve may produce impressive new results across a wide range of problems, it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding.   Even so, tools like AlphaEvolve are set to change the way researchers work. “I don’t think we are finished,” says Kohli. “There is much further that we can go in terms of how powerful this type of approach is.”

By Passant Rabie Published May 13, 2025 | Comments (1) | An illustration of a decaying neutron star. Daniëlle Futselaar/artsource.nl Around 13.8 billion years ago, a tiny but dense fireball gave birth to the vast cosmos that holds trillions of galaxies, including the Milky Way. But our universe is dying, and it’s happening at a much faster rate than scientists previously estimated, according to new research. The last stellar remnants of the universe will cease to exist in 10 to the power of 78 years (that’s a one with 78 zeros), according to a new estimate from a group of scientists at Radboud University in the Netherlands. That’s still a long way off from when the universe powers down for good—but it’s a far earlier fade-to-black moment than the previous 10 to the power of 1,100 years estimate. The new paper, published Monday in the Journal of Cosmology and Astroparticle Physics, is a follow-up to a previous study by the same group of researchers. In their 2023 study, black hole expert Heino Falcke, quantum physicist Michael Wondrak, and mathematician Walter van Suijlekom suggested that other objects, like neutron stars, could evaporate in much the same way as black holes. The original theory, developed by Stephen Hawking in 1974, proposed that radiation escaping near a black hole’s event horizon would gradually erode its mass over time. The phenomenon, known as Hawking radiation, remains one of the most surprising ideas about black holes to this day. Building on the theory of Hawking radiation, the researchers behind the new paper suggest that the process of erosion depends on the density of the object. They found that neutron stars and stellar black holes take roughly the same amount of time to decay, an estimated 10 to the power of 67 years. Although black holes have a stronger gravitational field that should cause them to evaporate faster, they also have no surface so they end up reabsorbing some of their own radiation, “which inhibits the process,” Wondrak said in a statement. The researchers then calculated how long various celestial bodies would take to evaporate via Hawking-like radiation, leading them to the abbreviated cosmic expiration date. “So the ultimate end of the universe comes much sooner than expected, but fortunately it still takes a very long time,” Falcke said. The study also estimates that it would take the Moon around 10 to the power of 90 years to evaporate based on Hawking radiation. “By asking these kinds of questions and looking at extreme cases, we want to better understand the theory, and perhaps one day, we unravel the mystery of Hawking radiation,” van Suijlekom said. Daily Newsletter You May Also Like By Isaac Schultz Published May 11, 2025 By Passant Rabie Published March 20, 2025 By Passant Rabie Published February 10, 2025 By Isaac Schultz Published February 2, 2025 By Margherita Bassi Published February 1, 2025 By Isaac Schultz Published January 28, 2025

Dr. James Kinross believes gut health is linked to colon cancer risk. Getty Images/ Justine Stoddart 2025-05-13T11:59:36Z Save Saved Read in app This story is available exclusively to Business Insider subscribers. Become an Insider and start reading now. Have an account? Our modern lifestyles are thought to damage our gut health in a way that raises the risk of colon cancer. Lifestyle changes can improve the gut microbiome, which could in turn lower the risk of colon cancer. As well as eating healthily, GI surgeon James Kinross eats enough vitamin D to prevent the disease. Colon cancer is rising in people under 50. James Kinross, a gastrointestinal surgeon who researches how the gut microbiome affects our risk of the disease, told Business Insider that poor gut health could be partly to blame.Research suggests the gut microbiome, the trillions of microbes that live in the digestive tract, has a wide-ranging effect on our health. But our sterile, urbanized lifestyles, appear to have made them less diverse overall and, therefore, weaker, according to Kinross, who is based at Imperial College London."You're seeing a generational loss in our internal ecology, which is being hammered with a series of environmental hits that it simply cannot adapt to," he said, referring to factors including microplastics, ultra-processed foods, and minimal access to nature. Many studies have linked But, the gut microbiome is changeable, meaning there are things we can do to increase its diversity, which in turn could help lower colon cancer risk. "It is an ecosystem which you can adapt, and you can modify," he said.Kinross previously shared with BI how he eats to boost his gut microbiome. He shared the three things he does aside from healthy eating to lower his colon cancer risk.Don't take antibiotics if you don't need to Kinross limits his antibiotic use where possible because the medication disrupts the gut microbiome. Trevor Williams/Getty Images As a surgeon, Kinross is acutely aware that antibiotics are often necessary and save millions of lives each year, but he believes we use them too liberally — particularly to treat viral infections they can't tackle."In my house, to qualify for antibiotics, you've really got to have a pathogen that you need treated," Kinross said.Kinross and his family limit their use of antibiotics as much as possible because taking them can disrupt the ecosystem of the gut microbiome, killing off good bacteria along with pathogens, and reducing diversity.He likened it to pouring weed killer all over your garden. "Your garden won't really grow into a lovely garden full of wild meadows and flowers and color. It will just be brown and lifeless," he said.Take vitamin DKinross takes a vitamin D supplement because evidence suggests that having enough of the nutrient is important for gut health. Vitamin D is crucial for several biological processes, including calcium absorption and cellular repair — but also keeps the lining of the small intestine strong. If the lining becomes weak, microbes can pass through it into the bloodstream and cause inflammation. Chronic inflammation is linked to a higher risk of chronic diseases, including cancer.Vitamin D is found in foods including oily fish, egg yolks, and red meat, as well as from sunlight. In countries with climates like the UK's, where Kinross lives, it can be difficult to get enough vitamin D in the colder months, and residents are advised to supplement from October to March.Socialize Social connections are a pillar of health. pixdeluxe/Getty Images Kinross goes cycling with friends regularly and makes sure to sit down to eat dinner with his family in the evenings. As well as being a pillar of overall health, research suggests that socializing positively impacts the composition of our gut microbiomes, Kinross said.We exchange microbes through physical touch, and studies have found that friends, family members, and spouses have similar gut microbes to one another.A 2024 study published in the journal Nature, based on 1,787 adults from 18 isolated villages in Honduras, found that people in the same social network shared more similarities in their gut microbiomes compared to those outside of it. This was regardless of diet, water sources, and medications. Spouses and people living together had the highest amount of microbial sharing, but the phenomenon still occurred among friends and even friends of friends."Our social interactions, our real-world social interactions, define so much of our health," Kinross said. "It's good for all aspects of our health. It's good for our mental health, it's good for our cardiovascular health. It's good for everything. But if you're not having real-world social interaction, you are not really optimizing your gut health, I believe." Recommended video