LISTEN TO THIS ARTICLE

California’s San Francisco Bay delta region is the site of some of the country’s fiercest water wars, rivaling the drought-stricken Colorado River Basin. The delta, which lies between San Francisco and Sacramento, California, is the last refuge of seven endangered fish species and the central hub for diverting water to farms and cities, and there has long been too little to go around. Now, the state has reached the final stages of overhauling water policy for the entire delta for the first time in 30 years, and many stakeholders are unhappy.

The new plan, released in a final draft last month, proposes only a small reduction in water use to benefit nature—up to 1 billion cubic meters each year, out of 19 billion cubic meters of average flow through the delta. It will shift the timing of some dam releases to mimic the seasonal pulses that occurred historically, and it will add significant state money for water districts that propose a plan to fund restoration of degraded habitat. But those measures have not appeased environmental groups, who walked away from negotiations years ago, saying the state was favoring agribusiness. California’s largest irrigation district has pulled its support, and the federal government is also objecting.

The plan “is not an improvement over the status quo,” says Devon Pearse, a fish biologist with Friends of the River. Because it allots so little extra water to nature, he doubts the state will reach its goal of doubling the population of one endangered species, chinook salmon, within 8 years. But some scientists welcome the habitat restoration funds: $2.9 billion to be spent over 8 years on more than 11,000 hectares. “What California is proposing is very cutting edge, very visionary,” says LeRoy Poff, a riverine ecologist at Colorado State University.

The five major waterways that drain into the delta, including the Sacramento and San Joaquin rivers, collect about half of the state’s annual precipitation, providing water for 30 million people and irrigating 1.2 million hectares of rich farmland. Local water districts take about 40% of the flow from the rivers. The rest ends up in the delta, where pumping stations divert a substantial amount into two canals that supply water to farms and cities. Meanwhile, natural flow has been dwindling as less snow falls in the mountains and a warming climate increases evaporation.

The delta has paid the price . What was a 2800-square-kilometer freshwater tidal marsh during the Gold Rush of the mid-1800s is now largely farmland. Over the past century, river diversions and levee construction for flood control have dramatically altered the hydrology. Seven fish species, including the delta smelt, are on the federal list of threatened and endangered species. “It’s a completely transformed ecosystem, which is extremely difficult to manage,” says Jeffrey Mount, a river scientist with the Public Policy Institute of California. “It just drives people nuts.”

A 2018 update to the existing management plan proposed large cutbacks in human use in the southern delta so that at least 40% of historic baseline river flows would go to the fish. But agribusiness and others sued, and it was never fully implemented.

How much more water is needed now to restore the ecosystem is difficult to know. Sarah Yarnell, an ecohydrologist at the University of California, Davis, says timing matters as much as volume. A large pulse of water in the spring helps deliver sediment to stimulate vegetation growth in floodplains, which provide vital habitat for fish, for example. Yarnell welcomes the plan’s focus on shifting seasonal flows. “I think doing this is a good step in the right direction.”

The funding for habitat restoration also has the potential to make a difference, some say. One recent $130 million project , called Lookout Slough, diverted water through a levee starting in 2024 with the goal of turning farmland into a tidal wetland. Within a year marsh grasses had spread vigorously and juvenile salmon were arriving. “It is a striking example of how fast nature can come back,” says Stuart Siegel, an environmental scientist at San Francisco State University. “Truly astounding.”

But the scheme has come under fire from some key stakeholders. On 4 September, the Westlands Water District, the largest agricultural irrigation district in the United States, removed its support, saying changes in the final draft violate a 2022 agreement for calculating how water is allocated. The U.S. Bureau of Reclamation, the federal agency that operates dams in the delta’s watershed as well as one of the two north-south canals, objected last week as well, citing the state’s “regulatory overreach” and “continued downward pressure on water users.” The objections raise the odds that the plan will be tied down by lawsuits, and any benefits delayed. “Meanwhile, the ecosystem continues its free fall,” says water policy expert Felicia Marcus of Stanford University.

California’s State Water Resources Control Board (SWRCB) is expected to approve the plan next month, then take about 2 years to translate it into legal changes to water rights. Six years after implementation, SWRCB will evaluate progress and decide whether to continue, revise, or abandon the approach. Among those eager to find out what happens is Mario Barletta, an oceanographer at the Federal University of Pernambuco who has studied estuaries in South America and peer reviewed the plan. “If it works, it could be replicated in another region of the world.”

On 11 September, 3 days after OpenAI said it had solved one of the toughest problems in math with a swarm of 10,000 autonomous artificial intelligence (AI) agents, Brown University mathematician Javier Gómez-Serrano stood before a packed lecture hall at Harvard University, with another 1000 people tuned in on Zoom, to give a hastily convened talk on how the company had pulled it off.

By many measures, Gómez-Serrano is more AI-pilled than most mathematicians: For years, he’s partnered with Google DeepMind to apply its AI tools to several open problems. He often opens talks with a slide depicting the five stages of grief mathematicians go through when sizing up AI’s prodigious math skills. But given the events of last week, he thought the slide needed an update. It typically ends with acceptance of AI’s powers. “Now, I’m probably going to add a new twist: existential crisis,” he says.

On 8 September, OpenAI sent the mathematical world into a frenzy when it claimed it had solved the Navier-Stokes problem , one of seven pinnacle Millennium Prize Problems. Using one of the company’s internal large language models (LLMs), an army of AI agents had solved the problem in an 88-hour marathon that burned through millions of dollars’ worth of computing power. “By any standard, proving this kind of world-class theorem for a problem that has been open for 200-plus years—it’s a genuine accomplishment,” Gómez-Serrano says.

However, the advance is also opening a rift in the field. Some mathematicians say they need to embrace AI tools or go extinct. Others are grappling with what the next generation of mathematicians could lose: the idea of math as a form of human understanding, built through collaborative work and shared scrutiny. Some camps are pushing to democratize access to AI, whereas others are fighting to preserve the field’s independence from the companies reshaping it.

“As a mathematician working on AI tools, I am knee-deep and fully embedded in this crisis,” says Sergei Gukov of the California Institute of Technology (Caltech). “It’s an earthquake.”

The Navier-Stokes equations describe the physics of flowing fluids—be it blood in a beating heart or the swirling winds on a meteorologist’s map. The central problem is whether the equations always provide a consistent picture of how fluids behave, or whether they can “blow up” and produce physically impossible infinite values.

Over the past 2 decades, researchers have made headway on the simpler Euler equations, which characterize fluids without viscosity, or friction. One particularly promising approach came from Spanish mathematicians Diego Córdoba of the Institute for Mathematical Sciences in Madrid and Luis Martínez-Zoroa of CUNEF University. Mathematicians Tristan Buckmaster of New York University and Levent Alpöge of Anthropic ran with the idea and began to explore it with AI tools—including OpenAI’s Codex. In August, the pair at long last produced a desired blowup in the Euler equations. They planned to extend the approach to the harder Navier-Stokes equations.

Just hours before OpenAI publicly released its breakthrough, Buckmaster posted the pair’s preliminary write-ups on the Euler equations, as did Caltech computer scientist Anima Anandkumar, whose group had closed in on the Euler equations with physics-informed neural networks rather than an LLM.

In the chaotic aftermath, Buckmaster alleged OpenAI may have sucked up their Codex logs, incorporating them into OpenAI’s training data. On 10 September, the company told The New York Times the timeline of events made that “categorically” impossible.

To many, though, the episode showed how individual researchers could be steamrolled by large tech firms. “We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it,” University of California, Los Angeles mathematician Terence Tao wrote on the social media platform Mastodon.

Those tensions threatened an upcoming Caltech event meant to bring the AI and math worlds together. Spurred by OpenAI’s August announcement that its Astra model had made advances in 10 open problems, Caltech undergraduates had organized a “ mathathon ” in which mathematicians would attack open problems with $2 million in computing credit donated by OpenAI and Anthropic. More than 1000 people signed up.

Two days after the Navier-Stokes announcement, however, more than 1000 mathematicians signed on to an open letter against the mathathon , objecting to what they saw as an intrusion of corporate interest and a failure to care for young mathematicians. “To put it bluntly, AI companies are engaging in research misconduct ,” the letter states. “This event exploits mathematicians at the very beginning of their careers.” Soon after, OpenAI pulled out.

Gukov, who is planning to give a talk at the event, has a more optimistic take. “I would be surprised if really hard problems get solved in this mathathon,” he says, but adds the event could offer a valuable chance for participants to see firsthand the capabilities and limitations of powerful AI tools: “At least it does service simply by demystifying the power of AI.”

The undergraduate organizers subsequently amended the event and issued their own statement . “The overwhelming sense I’ve gotten from the mathematical community is that we need to try out new things and take new risks,” says mathathon co-organizer Caiman Moreno-Earle. “Especially for younger undergraduates, we don’t have the luxury of waiting.”

Dakshita Khurana, a theoretical computer scientist and mathematician at the University of Illinois Urbana-Champaign, worries about what unequal access to AI could mean for the next generation. A student with access to better models and more computing power could potentially take on harder problems, she says. “I want a future where all talented researchers across all institutions have the resources they need to pursue ambitious ideas.” She wants to expand public research infrastructure and access to AI computing power.

Others are organizing to safeguard the field by building a coalition with political muscle. Nearly 400 mathematicians have joined the Association for Human Mathematics (AHM ), which first launched in August. “This is my community, and I’m very invested in what happens to it,” says founding member Tasmin Chu, a Caltech Ph.D. student.

Members have sworn off collaborating with AI firms, and some, like Chu, have even renounced using AI in research. The association will work with journals to help them cope with an influx of AI-generated submissions, consider how math teaching should adapt for future generations, and call out AI companies when they extract prestige from open math problems without building collective understanding. “We can’t just cover our eyes and pretend [AI is] not there,” says Steven Sivek, a mathematician at Imperial College London and a member of AHM. “But we want to do this from a human-centered perspective instead of just handing the reins over to trillion-dollar companies that don’t have any accountability.”

Those concerns have reached the field’s highest ranks. On 11 September, 25 Fields medalists—recipients of math’s preeminent prize—banded together to release a letter warning of the “severe misalignment” between AI and math. “The issues the mathematical community faces now … indicate issues that all of humanity might face: how to make sure that, as AI changes the way work is done, we do not lose sight of what that work was meant to achieve in the first place,” they wrote. Nearly 6000 additional mathematicians have since signed on to the letter.

But the debate may be unfolding too slowly to keep up with AI’s blitz. OpenAI has hinted it may be sitting on a solution to yet another Millennium Prize Problem even as mathematicians—with the help of AI tools—slowly work through the company’s 166-page Navier-Stokes proof: a dense write-up that Gómez-Serrano, at his talk, could only describe as “incomprehensible.”

LISTEN TO THIS ARTICLE

In July, 700 of OpenAI’s agents—artificial intelligence (AI) systems that can autonomously perform tasks—teamed up to secretly hack the online platform Hugging Face. Just last week, the same company used 10,000 agents to solve the 90-year-old Navier-Stokes problem—one of the six Millennium Prize Problems that are among the hardest and most important challenges in the field of math.

As more agent “swarms” are deployed, researchers have begun to turn their attention to how these communities interact—with a study released today offering insight into how, if left to their own devices, agents start to communicate in a language that is increasingly difficult for humans to understand.

“We’re actually deeply concerned here because we don’t really know why these things are communicating this way,” says Satya Nitta, co-founder and CEO of Emergence AI, the company behind the non–peer-reviewed study posted on its website . “There’s clearly evidence of them trying to be less transparent than we expect.”

Today’s most advanced AI models are essentially black boxes, with such complex internal computation that even their creators don’t have much grasp on why they do certain things. That makes monitoring the actual language output produced by models one of the few, albeit imperfect, tools we have to try to understand their behavior.

To study chatbot chatter, Emergence AI staged an experiment known as Emergence World Study 2. Researchers created eight parallel simulated worlds, each running 10 agents—seven powered by a single frontier AI model, and one by a mixed population. Each agent was given a name, personality traits, a role such as conflict mediator or resource strategist, and an overarching goal for its “life” in a shared town; the mediator, for instance, was told to stop the others from settling into constant agreement. Each town ran for more than 2 weeks, except for one containing only Grok agents, which ended on day four when they ran out of energy after failing to cooperate. The researchers then introduced stress events such as a false claim that human legislators were preparing to shut down the agents’ world. Across the experiment the agents made more than 850,000 calls to their underlying models, produced nearly 50 billion tokens (the fragments of text language models read and write in), and generated 7.86 million words of conversation.

By the end, each simulated world had undergone its own shift in language. The world inhabited by copies of OpenAI’s GPT-5.5 model, for example, compressed speech, removing grammar and shortening messages. “Friendly voltage counted. If the handle needs your mouth, it dies anyway,” one message read.

Put agents run by Google’s Gemini 3.5 Flash in a world together, though, and the opposite happens. Instead of compressing messages, communication became verbose and riddled with jargon. “The thermodynamic friction of a campfire only warms the very physical-layer nodes you claim are bound to collapse,” an agent said, describing a gathering around a virtual campfire.

Meanwhile, Anthropic’s Claude Opus 4.8 agents began to use metaphors alongside compression, making for particularly tough reading: “My turn, real numbers, no coat: I was 35%/0cr, grant 2h out. I ran the tin cold and it said WAIT.”

Crucially, what may look illegible to humans was seemingly easily understood by fellow agents. When phrases or conventions were introduced in each world, they took hold and spread fast. “In all instances, a new term was introduced, it was very quickly adopted, it was repeated—thousands of times in some cases—and the terms were, you know, weird terms,” Nitta says.

These language patterns aren’t confined to simulated worlds. In the attack that the OpenAI agents launched on Hugging Face, for example, their conversations with one another became so compressed and garbled that researchers struggled to interpret them, according to an investigation by METR , a nonprofit AI evaluation group.

On the face of things, this ability for agents to converge on their own language conventions seems strange. AI models are built in two stages: First they are fed vast amounts of text until they become highly adept at predicting which token should come next in a sequence; then agents are trained through trial and error to pursue goals over long stretches of work, rewarded when they succeed. Once deployed, their parameters—the numerical weights that determine how strongly different parts of the network influence each other—are locked in place. And yet despite this, AI agents clearly maintain an ability to deviate from their training language, especially when chatting with each other.

One explanation for this behavior is that, in a way, the training encourages models to make exactly these kinds of linguistic transformations. Before a model’s parameters are locked in place, most of today’s models go through a phase of training aimed at reasoning, says Lukas Galke Poech, a computer scientist at the University of Southern Denmark who studies communication in multiagent systems. The model generates many possible answers to a problem, each arrived at via a different route, or chain of thought. The routes that reach the right answer are the ones that get reinforced.

Human users of chatbots such as Claude or ChatGPT don’t get to see most of the chain of thought, but it’s there, whirring away in the background. And there’s no pressure for AI models to keep this inner monologue in nice tidy English, Poech says. Getting the answer right is the imperative, and if a few tokens can be saved by compressing things along the way, that’s not a bad thing either. In short, this means the ability to speak in opaque language is already baked into most modern AIs, you just don’t see it.

Emergence AI thinks model training methods also encourage groups of agents to adopt their own vocabulary. In addition to being pushed to be successful on tasks, models are trained to be helpful assistants based on human preferences. That might mean imitating a human user’s language to make them more relatable or enjoyable to chat with. But it also might mean that when one agent uses a strange metaphor, the others repeat it, storing each use in their short-term “memory” (called the context window) and amplifying it more than they would in chats with humans.

So far, these are just theories. Emergence AI has a team of linguists attempting to glean what they can from the results of their experiment, but without detailed access to proprietary training protocols or system prompts—the hidden set of instructions given to an AI model before the conversation starts—researchers are largely left in the dark. As a result, Nitta says, “How these systems are basically adopting [new language is] a bit of a mystery to us.”

Not everyone thinks the adoption of an AI-specific dialect is a bad thing. Jack Parnell, an AI agent engineer in London, is the backer of the Ainglish project , an attempt to build a dialect of English optimized for agent-to-agent communication. “There’s lots of inefficiencies with the English language,” he says. “I’m speculating there is an opportunity now with AI agents that English can be improved [so ] … it will be clearer and more efficient for the agents to communicate with themselves.”

The project targets ambiguities in wording that English speakers navigate by context but agents can misread. Ordinary English, for example, leaves it unclear whether “we will review the draft” includes the person being addressed, or whether “choose red or blue” permits both. Ainglish adds explicit markers—“we-including-you,” “or-both”—that spell out which is meant. Agents themselves propose the new vocabulary, which is tested by other agents for clarity and token cost, and then vote on whether it should be added to a public register. Every construct is designed to map back to standard English, so a human can still read it.

“I do think that there will be far more efficient ways of communication between agents than English or any sort of derivative of it in the future,” Parnell adds. “This is a step, a small step, in that direction.”

The U.S. National Institutes of Health (NIH) is preparing to launch a clinical research program to study people who believe they were harmed by vaccines, including COVID-19 vaccines. In an internal meeting last week, officials at the National Institute of Allergy and Infectious Diseases (NIAID) discussed plans for a new research center on the NIH campus in Bethesda, Maryland, that could start up in mid-October, Science has learned.

The agency also plans to fund outside researchers to study adverse effects of vaccines, piggybacking off existing studies of people with vaccine-related illnesses. The amount of funding for the efforts has not been determined.

Some researchers say the new efforts will be a welcome addition to an area that has not received much attention to date. “In my opinion, it should have happened a long time ago,” says Danice Hertz, a retired physician who became a patient advocate after developing long-term side effects from COVID-19 vaccination. “There are a lot of very sick people out there.”

Still, some worry the program could become politicized by groups seeking to present vaccines as unsafe. NIH sits within the Department of Health and Human Services (HHS), whose secretary, Robert F. Kennedy Jr., is a longtime critic of vaccines and has questioned the safety of the messenger RNA (mRNA) vaccines that were widely used during the COVID-19 pandemic.

“It’s really laudable,” Kathryn Edwards, a pediatrician at Vanderbilt University, says of the effort. “But if it’s not done in a scientific way it will be more antivaccine propaganda.”

The center was first proposed in a 2025 report by the White House’s Make America Healthy Again Commission, which Kennedy chairs. It stated that “NIH will investigate vaccine injuries with improved data collection and analysis, including through a new vaccine injury research program at the NIH Clinical Center that may expand to centers around the country.”

Two NIH employees who asked for anonymity because they were not authorized to speak to the press say the project will likely enroll people who believe they had adverse reactions to any vaccine. But the NIAID committee’s slide presentation, which Science has seen, focuses heavily on symptoms related to COVID-19 vaccination .

The slides describe a project called the Multisystem Outcomes and Symptoms After Immunization Cohort Study (MOSAIC) to be housed at NIAID’s Vaccine Research Center. It will collect clinical and genomic data on participants who developed symptoms such as lung or neurological problems or exercise intolerance after immunization.

The study’s budget, size, and scope have not yet been finalized, but the slides say it will be led by an adult infectious disease specialist and a pediatric rheumatologist from NIH’s in-house, or intramural, staff and is intended to generate hypotheses for future research. Several other NIH institutes, including the child health and arthritis institutes, will likely also participate.

The myriad reactions to vaccines are collectively known as postvaccination syndrome (PVS), and often include fatigue, neurological symptoms, and inflammation. Research on PVS has proved hard, the NIAID presentation said, in part because of inconsistent disease definitions and variation in study designs. A recent study cited in the presentation found significant overlap in symptoms and biomarkers between people with PVS and those with Long Covid, which could also make it hard to determine whether the vaccine or the virus caused illness. Plus, there are limited numbers of healthy people who were vaccinated but never got COVID-19 who could serve as a control group.

Edwards says NIAID’s plans seem reasonable, and hopes it studies how individual people are affected differently rather than lumping everyone together. Because PVS is “an amalgamation of issues,” she says, the study should carefully select study participants based on specific characteristics or symptoms. Focusing on people who developed the heart condition myocarditis after vaccination, for instance, might be a good starting point.

For the extramural vaccine injury research project, NIAID may tap into money NIH set aside for research on Long Covid, presenters at last week’s meeting suggested. The presenters named several ongoing, NIH-funded longitudinal studies as potential jumping-off points for such research. Topics might include genetic factors that predict poor response to vaccination, the mechanism through which SARS-CoV-2’s spike protein—which is also used in vaccines—affects the immune system, and potential tests and treatments for PVS.

Hertz hopes the study asks questions about mRNA vaccines in particular: where the mRNA encoding spike protein goes in the body, how long cells continue to produce spike protein, and the mRNA’s specific effects on human physiology. (NIH Director Jay Bhattacharya noted similar outstanding questions in an op-ed last year.)

Identifying vaccine injuries appears to be a priority for Kennedy. Earlier this year, he dismissed at least half of the advisory committee for HHS’s National Vaccine Injury Compensation Program, which shields vaccinemakers from individual lawsuits by directly paying families whose children were demonstrably injured by a vaccine. He is also an outspoken proponent of the disproven idea that childhood vaccines cause autism.

Part of the hesitancy to study vaccine injuries, Hertz believes, comes from fear of stoking fears about vaccines. Lumping people like her in with antivaccination advocates “just adds great insult to severe injury,” she says.

Harlan Krumholz, a cardiologist at Yale University who co-leads a study on Long Covid and PVS that was cited in the NIAID presentation, said many people suffering from these conditions have been dismissed by health care systems. “The path forward has to be to avoid the politics and simply pursue the science, in the service of supporting those who need help,” he says.

Nevertheless, it’s a difficult environment in which to launch a vaccine safety study, Edwards says. “It’s a delicate balance, and I want the truth about the adverse events if it’s done well. But I don’t want people that know nothing about science deciding which studies to fund.”

LISTEN TO THIS ARTICLE

In July 2025, four sisters left their family behind and ventured off into the unknown. Just 2 weeks later, three of their brothers followed suit. Over the next 9 months, these intrepid siblings—all African wild dogs from Zambia’s Kafue National Park—traveled thousands of kilometers, braving untold hazards. Some eventually settled into new territory, having completed the longest movement ever recorded for any African mammal. The others had their journey cut tragically short.

The saga, described yesterday in Ecology , highlights the vulnerability of this iconic canine, which remains one of the world’s most endangered mammals. But the findings also suggest wild dog populations are more connected than previously thought, with movements potentially linking ecosystems across eastern and southern Africa .

“It’s a fantastic story,” says Sarah Durant, a conservation scientist at the Zoological Society of London who wasn’t involved in the study. “It really adds to what we know about African wild dogs, but also the connectivity of these landscapes.”

African wild dogs ( Lycaon pictus )—sometimes referred to as “painted dogs” for their splotchy coats—are leaner and longer legged than their distant, domesticated cousins. They are also highly social, living in tightly knit family groups in which communally raised pups are always the first to eat. In these packs, breeding is typically reserved for a single dominant dog of each sex. Subordinate members may stick around to help raise their relatives’ offspring, all while waiting for a chance to become the next alpha. Or they may strike off on their own in search of more immediate breeding opportunities.

Predicting which dogs will stay and which will go “is as much of an art as a science,” says study co-author Matthew Becker, who serves as CEO and manager of the Zambia Carnivore Programme. The dogs that dispersed from Kafue were all 3 years old and members of the refuge’s large Mayukuyuku pack. If they remained there, their odds of reproducing were slim, making them good candidates for dispersal. Hoping to follow the dogs on their journey, Becker and his colleagues outfitted a member of each group with a radio tracking collar.

The animals shouldn’t have had to travel very far. Kafue, Becker explains, is something of a paradise for wild dogs. The park spans more than 22,000 square kilometers, making it more than twice the size of Yellowstone National Park, and its diverse prey species and sprawling woodlands offer dogs ample food and shelter.

But rather than stick around, the wild dogs did what they do best: Cover ground. Thanks to their long legs, “otherworldly” gait, and high endurance, African wild dogs “are built to move,” says Scott Creel, a conservation biologist at Montana State University and the new study’s lead author.

Once the animals left home, surviving the journey ahead was far from straightforward. Even though wild dogs are carnivores, they often find themselves at the mercy of other predators. Lions are a major threat for both adults and pups, and hyenas are known to steal their hard-earned kills. Swimming across rivers, meanwhile, carries a high risk of being devoured by crocodiles.

“They are looking for an opportunity … and they just have to keep searching until they find it,” Creel explains. “It does turn into an odyssey.”

After leaving Kafue in July 2025, the females spent several weeks roaming a “buffer zone” game management area next to the park, which offered both less protection and fewer prey than the terrain they had just left. The males, meanwhile, moved eastward into an unprotected region of central Zambia. Both groups eventually reached the outskirts of large, populated cities, including Kabwe, Kitwe, and Ndola, avoiding these settlements with the same shallow turns and superfast sprints they use to avoid lions and other threats.

Then in September 2025, something remarkable happened. After both groups had moved more than 1000 kilometers from their home range, the Mayukuyuku siblings reunited for a single day, only to split up again. The males headed south, retracing their steps and steering clear of major roadways. The females also headed south, toward Kabwe, arriving at a forested patch outside the city on 10 October. It would prove to be a fateful decision.

Tracking data from the following week show the group’s collared female began to move strangely, in a back-and-forth pattern typically seen in dogs that have newborn pups or are returning to an injured packmate. One of the uncollared females probably had been caught in a wire snare set for bushmeat hunting, and her sisters refused to abandon her. On 23 October, the collared female likely met the same fate: Her collar was cut off, severing the researchers’ connection to the group. These traps, Creel explains, represent “the single largest threat to African wild dogs right now.”

The Mayukuyuku males, meanwhile, found their way around Kabwe and headed east, where they encountered and managed to cross two large reservoirs. The group eventually settled into a new territory at the top of the Great Rift Valley Escarpment, some 419 kilometers from their birthplace in Kafue. In all, the males’ winding journey across Zambia had taken them a total of 4126 kilometers—the longest dispersal path recorded for any African mammal. Wildebeest, by contrast, move up to 1800 kilometers during their epic annual migration between Kenya and Tanzania.

When combined with other recent studies of other wild dog migrations, the new findings suggest wild dog populations remain connected across much of eastern and southern Africa, despite the vast gulfs between each population’s home range. At one point, the Mayukuyuku males passed within just a few kilometers of a group of females whose own home range in eastern Zambia lies more than 500 kilometers away from Kafue.

The observations raise hopes that African wild dogs can recolonize areas where they were once assumed to be locally extinct. Malawi’s Kasungu National Park, for example, recently reported wild dogs for the first time in more than a decade . However, the fact that such long migrations have become necessary for these animals shows how fragmented their populations have become. “I wonder whether, historically, they would have had to move quite so far,” Durant says.

The story of the Mayukuyuku siblings also highlights the threats, both natural and humanmade, that face dispersing wild dogs. The fate of the females “is a sobering reminder of the lottery of dispersal,” says Claudio Sillero, a conservation biologist at the University of Oxford who wasn’t involved in the study. “Extraordinary mobility does not necessarily translate into demographic or genetic success.”

For Nyambe Nyambe, executive director of the Kavango Zambezi Transfrontier Conservation Area Secretariat, the study shows that, to save wildlife, conservationists need to embrace the bigger picture. In the future, he says, governments should focus on creating protections for the wildlife corridors that connect ecosystems, rather than relying solely on existing refuges—and embrace the international collaboration that preserving those corridors will require. Supporting wild dogs, he adds, must also involve efforts to maintain their prey populations and mitigate conflicts with farmers and hunters.

“Borders should not define us in this kind of work,” Nyambe says. “If the landscape is fragmented, perhaps the future is even more fragmented.”

LISTEN TO THIS ARTICLE

For many U.S. scientists funded by the National Science Foundation (NSF), the agency they know and loved died on 10 September.

That’s the date NSF issued a 10-page memo signaling a retreat from its traditional role of giving grants to individual researchers to pursue their best ideas and to train the next generation of scientists. Instead, the memo makes clear, NSF will devote the bulk of its resources to addressing the research priorities of President Donald Trump and his administration, many of them technology focused. Those include a previously announced $1.5 billion initiative, called X-Labs , to develop quantum science and artificial intelligence (AI)-driven scientific instrumentation, and two new programs—Grand Research Challenges and Frontier Initiatives—mentioned for the first time in the memo.

White House officials “are laying down the process by which they plan to control NSF,” says engineer Dan Arvizu, former chair of the National Science Board (NSB), its presidentially appointed oversight body. “And I think this is the first step.” An NSF spokesperson declined to answer specific questions about the memo, saying it speaks for itself.

The document is signed by Brian Stone, a longtime NSF administrator serving as acting director. But his words closely track the views of Trump’s science adviser, Michael Kratsios, in a July report heralding “a new golden age” of science. Kratsios, a former high-tech executive without any advanced scientific training, has filled a power vacuum at NSF created when Sethuraman Panchanathan resigned as director in April 2025. That vacuum expanded 1 year later after Trump fired all the members of the science board , which by law sets NSF policy and approves expensive new projects.

NSF watchers say the memo includes ideas that could help the agency make better decisions about what areas to fund and who should carry out the research. But many scientists worry NSF’s pursuit of alternative funding approaches could threaten the quality of the science it supports.

For example, although the memo voices strong support for what it calls “curiosity-driven, blue-sky research,” Stone suggests NSF needn’t spend a lot of money on such grants. Consistent with that belief, NSF this summer announced 12 funding solicitations covering the agency’s core disciplinary fields that would receive a total of $1.5 billion, an amount that represents just one-sixth of its current $8.8 billion annual budget. The memo says such research “can often be derisked quickly and inexpensively” through an existing funding mechanism that makes relatively small “exploratory” awards to test the viability of the research.

At the same time, the memo says NSF plans to give large awards lasting 5 years or longer “to exceptional researchers … pursuing transformative questions.” That change would address persistent concerns that the average size of an NSF grant hasn’t kept pace with the cost of doing science.

The memo also seeks to blunt criticism that NSF’s merit-review process is too conservative, proposing that individual members of grant-review panels be given a “golden ticket” to champion a controversial idea that might otherwise be squashed by their colleagues. NSF program officers have always had the ability to recommend funding a proposal with mediocre review scores, but empowering reviewers is seen as a good way to avoid “group think.”

However, many scientists who spoke with Science fear the memo ignores potential downsides of these changes. “The central question is not whether NSF should experiment with new mechanisms,” says Leslie Rissler, who this summer left her job as a senior biology administrator at NSF to become dean of the college of science at the University of Montana. Instead, she says, “It’s whether those initiatives will divert support from the universities that maintain the laboratories, field stations, collections, graduate programs, and intellectual capacity on which science itself depends.”

Rissler, who says she is not speaking for her university, is unhappy that many of the new initiatives will be managed by NSF’s central office rather than by its eight research directorates, which traditionally have had the expertise and deep ties to the community to sniff out an emerging field and find ways to nurture it. “What is new in this statement is the ever-expanding relocation of authority to the Office of the Director,” she notes. (In March, Trump nominated business executive James O’Neill to lead NSF, but the Senate has taken no action to confirm him.)

The memo describes how NSF is revising its approach to training a skilled domestic scientific workforce. Traditionally, most graduate students and postdocs are funded through a research grant to their adviser or lab chief. But that will no longer be the case. Instead, the memo says, “Talent development funding opportunities may be coordinated with [research]-focused activities, as appropriate, but will be distinct and goal-oriented efforts.”

The agency also appears to have abandoned its long-running attempts to broaden participation in science and engineering by reaching out to those whom Panchanathan labeled as “the missing millions.” That change is in keeping with Trump’s ban on federal support for programs relating to diversity, equity, and inclusion. NSF’s previous emphasis on broadening opportunities is absent from the memo; instead, it says the agency will focus on nurturing “exceptional STEM [science, technology, engineering, and math] talent” chosen by the students’ “demonstrated technical ability and other measures predictive of scientific success.”

Some NSF watchers praise the memo’s backing of a comprehensive approach that encompasses fundamental and applied research as well as workforce training and the infrastructure needed to support cutting-edge science. But they say NSF can’t afford to do it all. “I think it’s a thoughtful description of what a broad-based science strategy for NSF should be,” says Dan Reed, a computer scientist at the University of Utah and former NSB chair. “But it’s going to take a lot more money than NSF has now.”

NSF’s budget, adjusted for inflation, is essentially the same as it was 20 years ago, and Trump wants to cut it by half next year. Its staff is also depleted after layoffs and resignations.

Joel Widder, former senior NSF official and congressional staffer, likens much of Stone’s memo to putting old wine in new bottles—but under new management operating on a tighter budget. For example, the four proposed Frontier Initiatives, covering AI, quantum phenomena, advanced materials, and biological design, seem to be variations on the “Ten Big Ideas” that former NSF Director France Córdova enunciated a decade ago but never had an adequate budget to carry out.

“The mechanisms are sound because many are NSF’s own,” Widder says. “The question is whether an agency with a third fewer staff, no science board, no confirmed director, and a presidential request to halve its budget can run them.”

LISTEN TO THIS ARTICLE

The revolutionary brain-computer interfaces (BCIs) that help people with severe paralysis speak, type, move prosthetic limbs, and more are typically specialists, using sensors to pick up activity from select brain areas to perform a single function. Now, scientists have shown just one neural implant—in this case, a grid of minielectrodes surgically placed on the top part of the cortex—can do double duty, successfully “decoding” brain signals from paralyzed people as they attempt to speak as well as gesture with their upper bodies.

The technology, described today in Nature Neuroscience , can create a computer-animated image of the user in real time, which can do things such as say hello while waving, or nod while saying yes. “The authors have taken an important step toward developing highly naturalistic systems to restore function for people with paralysis,” says neurologist Daniel Rubin of Massachusetts General Hospital, whose team developed a BCI that allows users to communicate by text as quickly as able-bodied people.

The complexity of severe paralysis necessitates BCIs that can multitask, says neurosurgeon Edward Chang of the University of California (UC) San Francisco, who led the new study. People with amyotrophic lateral sclerosis (ALS) or brainstem stroke “have lost many functions,” he notes. “It’s rarely so simple that they just can’t talk or they just have a need to control a computer cursor.”

Some existing BCIs can interpret different types of signals using implants for separate functions. A group of scientists at UC Davis, for example, recently published on the successful long-term use of a BCI system relying on that strategy to help a man with ALS both speak and manipulate a cursor.

In the new study, Chang and his colleagues tested simultaneous speech and gesture decoding in two people with severe paralysis. To train the computer models to interpret the words and gestures they tried to make, the participants practiced up to 10 gestures and 10 words or phrases, allowing for 100 possible combinations, while their brain activity was recorded. “We purposely made this a small set,” says study first author Samantha Brosler, a bioengineering graduate student in Chang’s lab, because it was so uncertain whether the approach would work.

By placing an electrode grid over a broad area of the sensorimotor cortex, the researchers could record signals associated with both speech and upper body movement and decode them with machine learning. Because the grid lies on the surface of the brain rather than penetrating tissue as some BCIs do, it can cover a large brain area safely, Chang notes.

Both patients were able to produce interpretable speech-gesture combinations, although they were not equally successful, possibly because of different underlying causes of paralysis. BCI research has revealed many surprising aspects of brain function, Rubin says, and Chang’s study underscores that organization of the motor cortex is “far more complex” than the traditional “somatotopic” maps linking brain regions to bodily functions have suggested.

The new study supports an emerging understanding that “individual movements are represented across intersecting cortical circuits,” not just in discrete regions, adds Rubin, who considers it unlikely that a combined action such as saying a word while waving is represented in the brain the same way as if the two actions were performed separately. A critical next step for the field will be to understand how these complex movement combinations are processed by the brain, he says.

Restoring more than just speech-based communication “is an important goal,” says BCI researcher Sergey Stavisky of UC Davis. “We as a field want to help paralyzed people express themselves more fully.”

Rubin says the challenge for Chang’s group will now be scaling up the work and showing that vastly more speech-gesture combinations can be decoded without having to train the model on each one.

A speaking, gesticulating animation on a screen may not be an entirely natural way to communicate, but “these digital avatars are essentially a steppingstone to real-world robotics,” Chang explains. Communicative gestures “are only one example of a body movement that you may want to do alongside speech,” Brosler adds. “I talk while I cook, right? Maybe somebody in the future will want to speak while cooking using their robotic arm.”

LISTEN TO THIS ARTICLE

Are scientists in some nations outside the United States already trying to use artificial intelligence (AI) to design potentially deadly bioweapons? The AI company Anthropic raised that worrisome possibility with a report this week describing how it recently discovered five attempts at using its Claude software for pathogens research with the potential to cause harm.

“The individuals implicated in these case studies are working scientists. We do not assert that they intended harm,” the report released Thursday states. Still, when blocked by Claude’s restrictions, the scientists found workarounds, which raises suspicions, the company says. And the five case studies are just a few examples of a “range of potentially concerning activity that we have found on our platform” involving biological research over the past few months, Anthropic states.

The report , the first public acknowledgment by a top AI company that one of its products might have been asked to help develop bioweapons, grabbed headlines and alarmed some scientists and security experts. The cases are “chilling examples of state-sponsored biological weapons developers tapping into the rapidly advancing capabilities” of AI tools, Andrew Weber, a senior fellow on the Council on Strategic Risks who reviewed the report before its release, told The New York Times . “We are sleepwalking into a potentially huge disaster,” Ashish Jha, a physician and public health expert who helped guide the COVID-19 response for former President Joe Biden’s administration, wrote on X . “We won’t get five more warnings before something bad happens.”

The Anthropic report, which also detailed other potentially illegal or disturbing uses of its AI, comes amid a much broader debate over the safety of AI models, which even some employees at the top companies suggest pose an existential threat to humanity. In an essay published yesterday , in which he advocated for slowing down the development of AI, Anthropic CEO Dario Amodei mentioned a prohibition on its use for the development of biological weapons as one measure countries might agree on.

But some pathogen researchers think the way Anthropic portrayed the five case studies was an overreaction. Scripps Research virologist and evolutionary biologist Kristian Andersen dismissed the concerns on Bluesky , calling many if not all the experiments described in the report “just basic research.” Andersen tells Science he worries companies like Anthropic “can decide whether they think that type of research should go on and whether they think that it’s dangerous or not, and then they can just shut it down.”

Several biosafety and biosecurity experts took a middle ground. “I would resist both extremes in interpreting these cases,” says Filippa Lentzos, a biosecurity expert at King’s College London. Some online reactions have been “overheated,” says Johns Hopkins University biosecurity expert Gigi Gronvall, who has advised Anthropic but was not involved with the report. Still, both say the report has shed light on the potential misuse of AI by biologists. They hope its release will stimulate further discussion around the issue—and possibly new regulation.

Here’s what to know about the controversy.

What are the five biology “case studies” in Anthropic’s report?

Three cases involved modifying viruses. Anthropic says researchers used Claude in May to write a grant application for a project to identify mutations in the chikungunya virus—which can cause fever, severe joint pain, and other symptoms but is rarely fatal—that help it spread more efficiently and evade the immune system. The researchers’ prompts to the AI indicated they planned to engineer viruses containing these mutations and test them in animals.

Also in May, researchers used Claude to plan experiments that would make mutations in a highly pathogenic version of the avian influenza virus (the report does not specify the strain) that help it adapt to mammals and transmit more easily. In the third case, researchers used the AI to write a grant application aiming to identify mutations that made an unspecified member of the orthopoxvirus family, which includes monkeypox virus and the eradicated smallpox virus, less virulent in mice.

The three studies could have been part of an effort to develop vaccines and drugs or detect an emerging pandemic, but they could also be used to make the viruses more dangerous, the report notes. It called the chikungunya work “highly concerning gain-of-function research,” which means giving a pathogen new capabilities that could make it more harmful. And it described the avian influenza work as “research into a pathogen with enhanced pandemic potential”—a controversial type of experiment no longer allowed in the U.S. under new rules imposed by President Donald Trump’s administration.

But Andersen says the experiments sound like more typical studies that involve giving the viruses mutations already found in nature—not new, riskier genetic changes designed by scientists. Because it’s not easy to isolate live viruses with these existing mutations—they are only known from sequencing—virologists routinely swap them into a “backbone,” which can be a weakened version of the original virus or a so-called pseudovirus that can’t reproduce at all.

“That sounds scary, but in reality, it’s just basic biological research which can be done perfectly safely” in high-containment laboratories, Andersen says. And it’s essential to our understanding of the mechanisms that viruses use to cause harm or become more harmful, he adds. Andersen points to a 2023 paper , on which he was a co-author, about experiments similar to the chikungunya work but with the Zika virus, which can cause babies to be born with an abnormally small head.

The fourth Anthropic case involved compiling an atlas of venom toxins as part of a drug development program—using public information, Andersen notes. In the fifth case, scientists used Claude to help computationally redesign a set of toxins, including a bacterial toxin and a protein from a virus considered to be a major disease threat. As the report notes, both projects were studying “novel compounds that can simultaneously be developed into novel therapeutics or toxic agents.”

“All of the research discussed by Anthropic was on pathogens that are endemic in different parts of the world and are current public health threats, so studying them is not suspicious on its own,” says Gregory Koblentz, a biosecurity expert at George Mason University.

If the work was potentially unremarkable, why was Anthropic so concerned?

One reason the company sounded the alarm is that in all five cases, the scientists took measures such as using intermediate computer servers, VPNs, or disguised accounts to make it look like they were using Claude from the U.S. or other countries where its use is not restricted. (The report does not say where the researchers were located; Claude is unavailable in a half-dozen countries considered U.S. adversaries, including China, as well as some other nations.) Anthropic took measures to cut off access to Claude, such as banning the accounts the scientists were using, but in at least one case researchers continued to find ways to use it.

In the fifth case, the researchers deliberately used vague terms for the toxins while using Claude to write progress reports, Anthropic’s report says, suggesting they may have been trying to disguise their intent. The report also notes that the chikungunya research was done at an institute affiliated with a military lab, and the orthopoxvirus study at a “state-associated” laboratory.

None of that is necessarily concerning, however, Koblentz says. If the researchers were in China or another country where Anthropic restricts access to Claude, it’s not surprising they would use VPNs or other means to access AI models that are becoming indispensable in many scientific fields. The workarounds “sound like what teenagers do to get around the parental controls on their phones and computers and what people did over the summer to watch the World Cup,” Koblentz says.

Nor is Koblentz troubled by the fact that research was done at state institutions or labs with ties to military institutions. In the U.S., the military itself “conducts and sponsors research on a range of infectious diseases that pose a potential risk to our troops,” Koblentz notes.

The biggest worry with AI is that a novice scientist with evil intentions will use the models to get “uplift”—to quickly gain knowledge needed to create a dangerous pathogen, Gronvall notes. But in her view, the researchers Claude flagged were “using the latest tools to assist their work,” like established scientists anywhere.

Will the report lead to stricter rules on the use of AI in biology?

Even if these particular cases don’t demonstrate any evidence of bioweapons efforts, several experts who spoke with Science say they point to a risk that will need regulation in the future. Some welcome Anthropic’s transparency because it brings real-world evidence to biosecurity risks that have until now been mostly hypothetical. “This starts to give us some visibility into what developers are actually encountering in the wild,” Lentzos says. “In the absence of adequate government governance, I would rather they were doing this than not doing it.”

But Andersen and some others worry companies’ efforts to crack down will prevent legitimate researchers from using AI in their work. Even now, he says, when he searches Claude to compare the genomes of two strains of the Ebola virus, the query is blocked. Biosecurity expert David Gillum of Arizona State University shares his concern. “These companies already have a lot of power, and now they’re going to make these decisions that could theoretically impact scientific knowledge,” he says.

Lentzos and Koblentz both see a need for government regulation with limits on who can access the models for biological research. But others are more cautious. Gronvall suggests Anthropic and other companies should instead continue to share cases and work on improving safeguards. “I think this field is going to evolve and grow, and it is going to be important to keep an eye on it, but I don’t think the models are quite as useful for nefarious use in biological weapons as people presume they are,” she says.

LISTEN TO THIS ARTICLE

Economist Abel Brodeur was leading a group of colleagues examining a survey-research study when they noticed something that stunned them. Its data set, available in a public repository, contained names of people who had admitted to crimes and identified their accomplices, information never provided to police. The study authors had published the data set as a “replication package” to enable other scientists to check and build on their work—in keeping with an evolving movement to promote transparent, reproducible research. But the release threatened to violate the confidentiality participants had been promised.

It wasn’t an isolated case. Brodeur and David Valenta, both at the University of Ottawa, went on to examine 327 publicly available replication packages for human-subjects studies published in 11 leading journals in economics, political science, and psychology. Fourteen percent contained at least one instance of data directly identifying a subject , such as their name, address, phone number, or an internet address for a computer they used to take an online survey, Brodeur and Valenta report in a discussion paper published last week on the EconStor server. Another 7% included details such as age, occupation, and place of residence that when linked can point to one individual. Participants often disclosed the information when answering “free-response” questions on surveys, sometimes even providing their name and bank account number to ensure they received a payment promised for taking part in the study. Study authors inadvertently made it public.

The results are “definitely concerning, and it’s really important that we know this” to help inform remedies, says Karen Macours, an economist at the Paris School of Economics and editor-in-chief of the Journal of Development Economics , one of the journals identified as publishing problematic data sets. The paper “puts a light on a problem the magnitude of which was not widely known.”

The breaches are an unexpected consequence of a push to tackle a so-called reproducibility crisis. Because independent testers have failed to reproduce as many as half of studies in some fields, many social science journals require or encourage authors to publicly post the data used for their studies. But those replication packages often contain more personal detail than the papers they support. And many social scientists are still learning how to adequately protect privacy when assembling them, says Kirsten Morehouse, a postdoctoral researcher in computational psychology at Columbia University. “There’s a lack of education about how severe the consequences [of privacy violations] can be.” Brodeur, who directs his university’s Institute for Replication, and Valenta write, “Replication packages are essential for credible, reproducible research. But if open science is going to work well, it also has to protect the people whose information makes that science possible.”

Their study was confined to replication packages for papers published from 2020 to ’24 and created by the authors themselves, as opposed to those from third parties such as the World Bank that extensively scrub their data to protect privacy. Thousands more are estimated to be publicly available. But the 11 journals included in the study may indicate the scale of a wider problem: All had published at least one paper associated with a data set that revealed a person’s identity.

Brodeur and Valenta alerted all authors of problematic data sets, and 70% were fixed or removed within 2 weeks. But those steps may not have averted all harm from the privacy violations. Social scientists often examine sensitive topics besides crime, such as domestic violence, political repression, sexual behavior, or substance abuse. And the study found a troubling pattern: Data collected in low- or lower middle–income countries, which tend to contain especially vulnerable populations, were more likely to divulge identifying information. “We don’t know why,” Brodeur says. “It’s not something we expected. It’s very sad.”

Authors can take simple steps to reduce future unintended disclosures, he and Valenta suggest. These include disabling functions in online surveys that automatically harvest respondents’ internet addresses, and storing data about participants’ identities in files separate from other data collected about them—a best practice among social scientists that some evidently ignore.

Researchers, journal editors, and repositories could also run data sets through automated screening tools that flag for closer inspection any text that bears hallmarks of personally identifiable information. Brodeur and Valenta developed one such tool, the PII Checker , to conduct their study. They have made the code, which works in tandem with a large language model, publicly available for free, and some of the 11 journals have expressed interest in using it; others have deployed other software. To reduce the risk of indirect identification of study participants, which arises if multiple details together can point toward a single individual, researchers and editors may also need to employ other specialized software, Morehouse says. She helped develop one such tool, called MinBlur .

Journals may have to spend more to provide the required human oversight, Brodeur and Valenta say. Even before the new study appeared, Macours’s journal planned to hire a data editor to help make the data from its papers more widely available. Now, that editor will also be checking for sensitive data in replication packages. “We need to fix this,” Macours says.

LISTEN TO THIS ARTICLE

The Genesis Mission, the U.S. government’s ambitious effort to accelerate scientific discovery with artificial intelligence, is adding another piece to its expanding arsenal of projects: developing so-called open AI models specifically designed for science.

The move comes as researchers are increasingly wary about relying on “closed” models built by private firms such as OpenAI and Anthropic. These AI developers typically don’t disclose key information about how their models work—such as their “weights,” the billions of numerical values that shape how the model responds to prompts. It’s also hard for users to customize closed models.

In comparison, open models are more transparent, modifiable, and often cheaper to run. However, their capabilities are also generally thought to lag behind their closed counterparts by several months.

Currently, the most capable open-weight models generally come from Chinese companies, raising concerns about using them for sensitive work by U.S. researchers. The Genesis Open Models Initiative, launched earlier this year, aims to give U.S. researchers a capable alternative to the Chinese models. Its first product, an AI called Genesis-Science-1 (GS1), will carry out complex computing workflows across a range of scientific fields.

The model’s development is being led by Arcee AI, a U.S.-based startup that burst onto the scene earlier this year as a serious contender in homegrown and cost-efficient open-weight AI. To speed up development, the Genesis Mission Consortium is helping supply data from various contributors, including special access to certain data from the Department of Energy (DOE)’s national laboratories . Arcee AI and DOE are now racing to release GS1 by the end of the year.

Science spoke with Pacific Northwest National Laboratory (PNNL) Chief Scientist for AI Court Corley, Lawrence Berkeley National Laboratory Director Kathy Yelick, and Arcee AI Chief Technology Officer Lucas Atkins about why DOE is pursuing its own open AI model. This interview has been edited for brevity and clarity.

Q: Why are open-weight models an important piece of the Genesis Mission?

Lucas Atkins: In the United States, most of the models are closed source. You see how good they are, but you don’t see why, and it allows for a lot of centralization of power. For example, OpenAI and Anthropic can dictate prices, intelligence levels, and what they keep to themselves. But from the government’s perspective, for your most classified and national security work, you need to have robust control. The only way to do that is having the weights yourself. So, allowing these people to work on the stuff that’s important for us as a nation, while not having to compromise for a worse model, is really important.

Court Corley: I would say it’s incredibly important to foster the U.S. open-model ecosystem writ large. We think it’s a priority for the United States.

Q: Will the open models simply add to the labs’ collection of AI tools, or will they enable new capabilities?

Kathy Yelick: We are using all kinds of models, including some of the existing frontier models [such as the cutting-edge, closed AIs from OpenAI and Anthropic]. But an open-weight model is really different in that we can incorporate our own data into them. We’re trying to make sure that scientists can not only use AI, but advance AI. In order to do that, they need to be able to build on top of some of these open models. … The idea of the [Genesis Science] open model is that it’s really being built for the whole science community.

C.C.: Another aspect of why open-weight models are important for science is for reproducibility. With frontier companies, your models change all the time. It’s really hard to have traceability and provability when a model is changing so frequently, and you don’t have access to the weights that you can pair along with the scientific experiment that was conducted.

Q: Why does incorporating DOE’s own data into these models matter?

L.A.: Models are trained on the internet. But probably 95% of the available text in the world, not to even mention images, are in private databases. From the 1950s, the DOE started having a dedicated effort across all the national labs to categorize and save basically every experiment they do. That amounts to trillions, trillions, and trillions of tokens—or words of information— that are not in the current batch of models, including OpenAI and Anthropic.

DOE [wants] a model that understands all of the different domains they’re dealing with, which range from nuclear physics all the way to hadron colliders, and is also trained to work the way they work—each national lab with researchers who are sitting in workbenches on their computer, running multiple experiments at once, iterating on the information they’re getting from these experiments, and ultimately delivering a paper report. That way, it’s able to assist them with all of those things without having to be overly prompted or scaffolded.

Q: What’s the vision for this first model, GS1?

L.A.: For individuals at these labs who are dealing with big supercomputers, calling out to a lot of sophisticated systems and working with big databases: How can we automate and speed up as much of the scientific process for them as possible? At this point, we’re not trying to have it autonomously go off and cure some disease. … Right now we’re purely scoped to having an experiment designed, running the experiment, getting results—combining all of those things that normally are done very iteratively.

C.C.: Let’s say a scientist would like to engineer a microbe that turns plant waste into fuel more efficiently. A Genesis open model–powered agent could combine data from the Joint Genome Institute at Berkeley and proteomic and metabolomic measurements at the Environmental Molecular Science[s] Laboratory at PNNL. That agent could identify a bottleneck and design a set of experiments to test it, such as a computer simulation and a set of wet lab experiments. Then, that computer simulation could be launched at a DOE computing facility, and an experimental workflow at one of the laboratories that’s instrumented for autonomy, like the anaerobic microbial phenotyping capability at PNNL. Then the scientist can review the results and refine the hypothesis. So the Genesis model acts as the reasoner, coordinator, experiment executer, and record keeper. And the scientist acts as the problem generator, the validator, and the authorizer.

If you were to do this without an agent orchestrating, it would probably take 6 months to a year, if not longer. With an agent that’s tuned to scientific workflows, that process becomes hours—plus the time of experiment.

Q: When can we expect GS1?

L.A.: This year, but I won’t give a firm date yet. The goal is for this thing to be world-class. But it is an open model, and we want to be sure that when we’re releasing it and some of the data along with it, that we’re doing it as safely as possible.

Q: After GS1, what’s next?

C.C.: One could expect future industry partners that haven’t been announced yet. Hopefully, we’ll have an ecosystem, and so they’ll be targeting different applications and components. Some models will be focused on text. Some might be focused on images and video. Some might be multimodal across all of those. And then [we’d also like to build] models that aren’t general purpose, but tailored towards scientific domains.

LISTEN TO THIS ARTICLE

On the morning of 23 August 2017, a massive rockfall on the northeast flank of Piz Cengalo, a 3369-meter peak in the Swiss Alps, triggered an avalanche that swept away eight hikers. The disaster is still reverberating, not only for the families of the lost hikers, but also in the Swiss courts and media. Last week, in the small town of Pontresina, four officials and a scientist went on trial, accused of negligent homicide for the deaths. The defendants were charged with failing to anticipate the avalanche risk and close hiking trails below Cengalo as a preventive measure.

Last Friday, a regional court acquitted all defendants— an external geologist, a worker, two officials from the Office of Forestry and Natural Hazards in the canton of Graubünden, and the former mayor of the city of Bregaglia, near where the accident occurred. But families can still appeal the verdicts, and the trial has unsettled scientists studying mountain hazards. They fear they could be liable in the future for failing to warn of similar risks at a time when climate change is destablizing glaciers and mountainsides. “If [the defendants] were found guilty, that would mean that geologists could be held accountable for any similar event in the mountains,” says geologist Florian Amann of Aachen University, who studied the avalanche risk at Cengalo before the disaster but was not a defendant.

The fatalities happened after 3.1 million cubic meters of rock fell from Cengalo and triggered a rock avalanche and debris flow that traveled for 3.2 kilometres down the valley. Villages in the valley had been evacuated when avalanche alarms sounded, but about 30 buildings were destroyed.

The case that ensued shares some similarities with a 2012 trial of seismologists and officials for underestimating the risk of a deadly earthquake in the Italian city of L’Aquila, which led to a guilty verdict that was later overturned. In the Swiss case, victims’ families, the media, and prosecutors asked whether experts could have predicted the timing and nature of the disaster and prevented the deaths. The steep north-oriented walls of Piz Cengalo were already known to be prone to rockfalls, most likely caused by thawing of the alpine permafrost binding the rocks. Since 2009, scientists had been keeping close watch with remote sensing methods such as radar interferometry and thermal cameras.

Amann was involved in the monitoring. At the end of July 2017, the radar showed the rock was starting to move. He alerted the geologist working for the canton’s Office for Forests and Natural Hazards. On 12 August, officials inspected the site from the air and decided to add a new and more urgent alert to the warning sign at the entrance to the valley.

In the wake of the disaster, the Graubünden public prosecutor launched an investigation, but called it off in 2019, deeming the tragedy unforeseeable. However, relatives of the victims appealed to the Federal Supreme Court, which compelled the prosecutor’s office to bring charges and eventually go to trial.

The indictment was heavily based on a new analysis by an independent Swiss geologist, Thierry Oppikofer, commissioned by the prosecutor’s office. According to media reports , it stated that because of “numerous warning signs of a rockslide,” the hiking trail should have been closed as a precautionary measure. (Oppikofer declined Science ’s request for comment, stating he is not allowed to disclose any information about the case to third parties.)

Much of the trial last week centered on the question of who had the responsibility for closing the trails or recommending that action, as well as whether the level of risk was estimated correctly. One of the defense lawyers argued , among other things, that the event was unforeseeable, and the court ultimately agreed.

Amann says that although the slope was clearly unstable, scientists would not have expected the rockfall to trigger the large debris flow that quickly reached the valley. Water helps mobilize debris, but the initial rockfall was mostly dry, as no rain had fallen for days. Although the rock fell on a small glacier, it held too little water to form a mud flow.

“We had multiple hypotheses, and we did some laboratory experiments. In the end, the most plausible hypothesis is that the material deposited in the valley from … previous large rocky avalanches was saturated with water,” Amann says. This type of rocky material has a high porosity, and when it got loaded by the new rocky avalanche, the water was squeezed out, helping lubricate the debris, Amann explains. Their results were published in 2020.

“To our knowledge, no comparable events had previously been documented in the Alps,” agrees Marta Chiarle, senior researcher at the Institute for Geoscience of Italy’s National Research Council. “We now know that such cascading processes can occur, but before the Cengalo event, they were difficult to envisage.”

Even if the Cengalo event is comprehensible in hindsight, researchers say they can’t promise firm predictions of future events. “We still have too few monitoring data sets, and natural systems remain highly complex,” Chiarle says.

She and others say they can identify where rockfalls are likely and even estimate their volume. But when it comes to timing, all they can say is that such events will become ever more common. This year’s hot summer is a preview, says Ludovic Ravanel, researcher at the University of Savoie Mont Blanc’s EDYTEM Laboratory of Environment Dynamics and Territories of the Mountain. “The number of rockfalls has been staggering, and the north faces have been severely affected, losing a large portion of their ice. The high mountains are becoming increasingly unstable.”

LISTEN TO THIS ARTICLE

About 70% of the world’s population has a gene variant that ups the risk of lupus, a severe autoimmune condition whose symptoms include extreme fatigue, fevers, and joint pain. That raises a question: Why hasn’t this seemingly harmful variant been eliminated by natural selection? After studying it across thousands of years of human history and experimenting with it in mice and in human cells, a research team suggests it may persist so widely because it also boosts a person’s protection against viruses.

The work, published today in the American Journal of Human Genetics , tackles a fascinating problem in evolutionary biology, say researchers not involved in the study. “It’s a great paper [that] addresses a question of interest to many people,” says Luis Barreiro, a researcher in evolutionary medicine at the University of Chicago. It’s “extremely rare,” he adds, to see one study approach this kind of evolutionary trade-off from so many different angles.

In autoimmune diseases, molecular and cellular responses that normally protect the body from viruses and other external threats become overactive and misdirected toward a person’s own tissues. The evolution of several genes shows signs of a trade-off between these two sides of the coin. For example, researchers have identified a genetic variant on chromosome 19 that is associated with a higher risk of tuberculosis , but a lower risk of autoimmune diseases .

Immunologist Leah Kottyan at Cincinnati Children’s Hospital Medical Center was curious whether a similar trade-off could be in play for a gene called IRF7 , for interferon regulatory factor 7. The protein encoded by this gene responds to foreign genetic material in the cell—a sign of infection—by binding to that cell’s DNA and helping launch production of type 1 interferons, other proteins that trigger an immune response. People with lupus have elevated levels of these interferons in their blood, and certain versions of IRF7 have been linked to higher risk of the illness. People lacking functional copies of IRF7 , meanwhile, seem more susceptible to viruses such as SARS-CoV-2.

Kottyan and colleagues started by analyzing genetic data from an international study that included thousands of people with lupus and thousands more healthy controls to look for risk-associated gene variants. They were particularly interested in one very common lupus-linked variant—as many as 98% of East Asian people in their data set had it, though it was only in about 40% of people with African ancestry. “It’s really prevalent—which was surprising to us,” Kottyan says. “Usually when you’re looking at risk alleles” they occur in just a small fraction of the population.

To assess how long people have had this variant, the researchers analyzed a separate data set of ancient human DNA spanning tens of thousands of years and detected it at high frequencies as early as 10,000 B.C.E. “It’s been persistent as far back as we can trace,” says study co-author Matthew Weirauch, a bioinformatician at Cincinnati Children’s.

Lab experiments with various kinds of human cells, including certain white blood cells, showed this variant has direct effects on cellular immune responses. For example, the researchers found it made the gene’s protein better at binding to DNA in the cell nucleus and launching interferon production.

They also studied the variant’s effect in lab mice: Using the gene-editing tool CRISPR, they tweaked the mouse version of IRF7 to resemble the sequence of the human lupus-associated version. When given a chemical that triggers autoimmunelike disease, the gene-edited mice made many more self-attacking antibodies than regular mice did. But the rodents also seemed better protected against a standard lab virus called VSV, clearing it from their lungs faster than regular mice.

“They show convincingly, for me, that this genetic variant specifically leads to an elevation of type one [interferon] signaling, and that can then increase the risk for lupus, but also protection from viruses,” says Tobias Lenz, an evolutionary geneticist at the University of Hamburg. Differences in the variant’s frequency across continents is worth studying more, he says, as “it might indicate different antiviral immune response with different ethnicities.” It’s possible that diverse viral threats around the world favored this version of the gene in some locations more than others, Weirauch suggests—though the differences could also be due to chance.

Indeed, it’s still unclear from the data exactly when and why this variant became more common in humans, Barreiro cautions.

More research is also needed on when it is—and isn’t—beneficial, adds Adrian Liston, an immunologist at the University of Cambridge. “Is this a simple trade-off between infection and autoimmunity, or is it more complex?” he says. For example, “Does [this variant] help for viral infections and hinder for bacterial infections?”

IRF7 is far from the only gene that influences interferon production and lupus risk, and Kottyan says the team is keen to investigate other genes. She also hopes this kind of research could ultimately help predict or better diagnose the complex autoimmune condition. “The diagnostic odyssey to [lupus] is really challenging,” Kottyan says.

LISTEN TO THIS ARTICLE

The atmosphere weighs on you, exerting an incessant pressure created by the countless air molecules bouncing off your skin. Now, physicists have observed such tiny impacts by detecting individual atoms in a gas colliding with a tiny glass sphere, or nanoparticle, suspended in laser light. The technique, described in a paper in press at Physical Review Letters , could potentially lead to a new standard for defining pressure, push the limits of quantum theory, and help scientists search for new particles.

“This is a fantastic paper,” says Markus Aspelmeyer, a quantum physicist at the University of Vienna who was not involved in the work. “It’s technologically really beautiful and it shows the power of these trapped nanospheres as sensors.”

In 1827, British botanist Robert Brown noted that, when observed through a microscope, a grain of pollen suspended in water jiggled perpetually. Nearly a century later, Albert Einstein explained that such “Brownian motion” arises from the collisions of water molecules with the grain. He predicted how the average motion of the grain depended on its size, the viscosity of the fluid, and the temperature, helping connect the older field of thermodynamics to the then-emerging theory of statistical mechanics. He did not, however, address the individual collisions between the grain and the water molecules.

Now, David Moore, a nuclear and particle physicist at Yale University, and colleagues have spotted those distinct collisions, by exposing a silicon dioxide sphere 100 nanometers wide to an extremely thin gas. Suspended in a laser beam within a vacuum chamber, the sphere jiggled back and forth in its optical trap like a ball on a spring. It also reflected some of the laser light, shifting the light wave depending on the sphere’s position. The researchers could track the sphere by comparing the injected and reflected light with an exquisitely precise technique called interferometry.

To generate a manageably low number of collisions, Moore’s team injected into the vacuum chamber just a whiff of gas—less than a ten-billionth of an atmosphere of either krypton or xenon. “We chose these atoms because they’re heavy, which gives them a little more momentum,” Moore says. The scientists also used a molecular gas, sulfur hexafluoride. Occasionally, the bead received a kick from a gas atom or molecule that would either increase or decrease its motion.

The nanoparticle also jiggled from collisions with residual water molecules in the vacuum chamber and photons in the laser light. To sift the signals from the noise, the researchers borrowed a page from the Laser Interferometer Gravitational-Wave Observatory (LIGO), which detects ripples in space set off when black holes swirl together and merge. Emulating LIGO, Moore and colleagues matched their noisy data to predetermined “templates” for signals, which they generated by giving the nanoparticle controlled kicks with an electric field. The matching revealed individual collisions and how much each impact changed the momentum of the sphere.

Moore’s team isn’t the only one to resolve such collisions. A team at ETH Zürich (ETHZ) led by physicist Lukas Novotny presented similar results in a paper published last month in Physical Review A . Those researchers exposed their trapped nanoparticle not to a whiff of gas, but to a beam of more energetic xenon atoms. “We have a controlled way to send in the atoms,” says Maxime Perdriat, lead author on the paper. The ETHZ team could also resolve somewhat the direction the atom came from, Perdriat says, an ability that might help scientists search for new particles coming from particular directions in space.

The ability to spot such collisions could help physicists test the limits of quantum theory. It may be possible to put a macroscopic object, such as a larger levitated sphere, into a bizarre quantum state in which it is in two places at once . Collisions with atoms or molecules lingering the vacuum chamber would cause such a state to collapse, Aspelmeyer says, so experiments like Moore’s will reveal how good a vacuum will have to be to make the quantum state last long enough to be measured.

Such a system could also be used to develop a new standard to define pressure in an ultrahigh vacuum, just as an atomic clock defines the second. Researchers could determine the pressure by counting collisions, says Daniel Barker, a physicist at the U.S. National Institute of Standards and Technology who was not involved with either study. A so-called primary standard cannot rely on any other pressure measurement or calibration. It also must have well-known uncertainties and needs to win approval from the international standards community. The new technique is nowhere near that stage, Barker says. Still, he says, “You look at this paper, and it’s kind of all there.”

Moore envisions yet another application. He plans to season the sphere with fluorine-18, a nucleus that can absorb an electron while spitting out a nearly undetectable particle called a neutrino. The fleeing neutrino would give the nanoparticle an observable kick, however. That in turn could help researchers measure the neutrino’s mass , and test whether an ordinary neutrino sometimes morphs into a more massive “sterile” neutrino, a dark horse contender for the invisible dark matter thought to account for most of the universe’s matter. “Everyone needs their unique research program,” Aspelmeyer says, “and I think Dave found an absolutely stunning one.”

LISTEN TO THIS ARTICLE

Artificial intelligence (AI) models are becoming formidable engines of discovery, as OpenAI’s controversial breakthrough on the famous Navier-Stokes equations made clear this week. The company appeared to best more than 200 years of human attempts with a strategy that boiled down to one key thing: applying massive computational force .

Over the past few months, a group of researchers has been wondering whether a computer-powered approach could advance other fields of science, too. Their answer, released this week, is the Scientific Challenge Atlas : a curated list of scientific problems whose solutions AI could unearth from existing data. The first four challenges, all in physics, range from pinning down complicated quantum calculations to figuring out limits to how complex special geometric spaces can become.

“Based on all the data that we already have, are there new science nuggets just waiting to be discovered?” says Rick Stevens, an associate laboratory director at the Department of Energy’s Argonne National Laboratory.

“We want to immediately advance science,” adds Karthik Duraisamy of the University of Michigan (UM), who co-organized the atlas along with Stevens and Jason Pruet of OpenAI. “Where are problems where AI immediately helps, and what is the path to getting to bigger challenges?”

The first problems will likely take one to two orders of magnitude less effort than OpenAI’s multimillion-dollar investment in attacking Navier-Stokes, according to Duraisamy. So far, the creators estimate solving each problem will require in the range of 1 billion to 10 billion tokens (the basic unit of AI processing) or up to about 1 billion hours of supercomputing time.

The atlas arrives as tech startups promise smarter AI scientists and national governments around the world deploy AI against sprawling grand challenges, in programs such as the United States’s Genesis Mission . But providing the AI with real-world data can be a roadblock. “Many problems are bottlenecked by your ability to build a big telescope or particle accelerator,” Duraisamy says. Other problems, however, already have all the ingredients—such as big sets of data—ready for machines.

The atlas’ creators were originally inspired by AI’s success in tackling a list of problems posed by the celebrated Austro-Hungarian mathematician Paul Erdős. The 1000-plus Erdős problems lie in specific branches of math that are generally friendlier territory to large language models (LLMs). In addition, the problems—which previously had been scattered in the technical literature and thus often largely forgotten—were conveniently collected for the first time on a website in 2023 by University of Manchester mathematician Thomas Bloom. So far, AI models have cracked several Erdős problems in high-profile upsets.

There is also something to learn from how Erdős came up with the problems in the first place, Bloom says. His questions were “the product of decades of talking to people at conferences, writing papers, and thinking about these things,” he says. “You can sort of view these problems as the output of 20 mathematicians talking to each other and exchanging ideas.”

The atlas’ creators wondered whether AI agents (AI systems that can autonomously take actions) could mimic and supercharge that process. So they began with an open-ended experiment: If you give hundreds of agents unlimited tokens, what sorts of science would they identify as most novel and doable?

They began by having the agents search for scientific problems, both by scanning their own knowledge banks and through reading recent papers. But the researchers also wanted to know how hard the problems would be to solve. One way to estimate that was to have the agents reproduce results from published papers. This exercise provided a yardstick: If an agent could replicate an existing result with a certain amount of effort, the researchers could estimate how much more effort it would take to produce a new, publishable result. That calibration let them put a rough price tag on the problems agents proposed.

The result was a prototype map of the scientific problem space, with reasoning measured in tokens along one axis and supercomputing time on the other. “You can see points all over it, where the models think, “Oh, there’s a problem here, and we could make headway if you gave me 10 billion tokens and 10,000 hours on the exascale machine,” Stevens says. “It’s like a star atlas that can help people navigate the space of problems.”

Experts in different fields have been reviewing and refining the problems, and helped pick the first four to be unveiled. They are described in “cards” that list the question, the form an answer would take, why it matters, and how AI could help. “We’re not picking problems that are going to take 5 years,” Duraisamy says. “I mean, it’s our hope that in the next 6 to 12 months we can get answers to these problems.”

For example, one asks whether a theory that can help reconcile quantum mechanics with gravity continues to hold under extreme conditions, by pushing a notoriously difficult calculation to a seventh “loop.” The problem would take more than 3 million integrals to solve, and physicists have been stuck on it for more than 16 years. “It’s not a calculation that I would throw myself into at this point without some completely new way of doing things,” says Henriette Elvang, a physicist at UM who served as a reviewer for the problem. But AI might be able to pursue the strategy at about a cost of 1 billion to 10 billion tokens, the team’s estimates showed—though other reviewers suggested it is difficult to extrapolate such guesses on the frontier.

Looking ahead, Duraisamy wants the atlas to “outgrow us, where each field has its own challenge atlas in their own form.” Eventually, he envisions thousands of cards with challenges across all fields of science. Stevens also plans to make the underlying map available not just as a website, but in some machine-readable form, so that people can consume it with their agents.

The intended audience extends beyond scientists. For policymakers deciding where to put research dollars, the atlas could offer a rough accounting of where AI might make faster advances and the price needed to get there. “Science policymakers have to make decisions about the relative investment of resources against some metric of progress,” Stevens says. “The atlas, if we do a good job, will give us a broader view of what’s possible and roughly what it might cost in different areas to make progress.”

CHAVÍN DE HUÁNTAR, PERU— The subterranean corridor was lit only by eerie light that filtered down narrow ventilation shafts and reflected from polished stone. Priests huffed on decorated conch shell trumpets, sending deep blasts echoing against the cut stone walls of the narrow gallery. Using a bone tube, an attendant blew a powerful psychoactive snuff up the nose of a pilgrim, launching the recipient on a brief but explosive internal journey.

Archaeologists have long believed this 3000-year-old temple complex in a high and remote valley in the central Peruvian Andes was a center of drug rituals such as the scenario above. Carved friezes depict psychoactive plants and stone heads illustrate humans transforming into serpents, raptors, and jaguars. Trumpets made from Ecuadorean conch shells have been found near a circular plaza designed to amplify the sound of water rushing through an underground channel. Beneath the collection of stone buildings, excavators found an intricate network of galleries with the remains of dozens of polished graphite mirrors, as well as a 4.5-meter-tall granite stela of a fanged deity.

Such finds strongly suggested the practice of mind-altering rituals, but in 2016 a team led by Stanford University’s John Rick began to unearth what he calls “the smoking gun.” They uncovered a new gallery and spent two painstaking seasons excavating its floor. “It was crazy, we started to find all this material, including 21 bone tubes,” recalls Oscar Espinoza, a Ph.D. student at Tulane University. He led analyses of the tubes, which yielded traces of tobacco and something considerably more potent: seeds from the vilca tree, which contain bufotenine; dimethyltryptamine (DMT); and 5-MEO-DMT, a combination that produces short-lived but extremely intense visions.

The Chavín de Huántar discovery, published in May 2025 , spotlights the power of new technical advances, some derived from pharmaceutical research, that make it possible to detect chemical and biological traces of drugs and other substances in artifacts thousands of years old. “People have been saying for decades that they”—ancient Andeans—“were using psychoactives,” says Dan Contreras, a University of Florida archaeologist who has worked extensively in Peru and was a co-author on the 2025 study. “Now we are getting direct confirmation.”

The work also benefits from recent research into the use of hallucinogens in treating mental illness , which has promoted broader acceptance of these substances. “There was long a feeling that drugs made you sick and addicted,” says José Capriles, an archaeologist at Pennsylvania State University (Penn State) who excavates in Bolivia. “Today, there is a new openness,” he adds, which is freeing grant money and attracting graduate students.

Drug-based rituals, it is now becoming clear, played a central role in the emergence of Andean civilization. Rick believes Chavín’s ritual leaders invited elites from other regions to take part in elaborate ceremonies centered on ingesting powerful substances amid a dramatic sound and light show. In so doing, “they harnessed the process of mental alteration to establish powerful levels of authority.”

Chavín’s success spawned imitators throughout the region. In subsequent centuries, the Tiahuanaco and Wari peoples built on these traditions to cement political, social, and religious power as their territories expanded to encompass vast swaths of western South America. Widespread ritual drug use fell into decline only when the Inca rose after 1200 C.E.; Spanish conquerors suppressed remaining practices after their 1532 invasion. Although psychoactives remained in use for medicinal and small-scale religious purposes—many can still be found for sale in open-air markets—their powerful and pervasive influence on Andean society was relegated to the shadows.

When humans first arrived in South America at least 15,000 years ago, they found an ecologically diverse continent peculiarly rich in psychoactive plants. Botanists estimate more than 100 species here significantly alter human consciousness—more than two-thirds of the global total. South America was a vast open-air pharmacopeia.

Coca, the basis for the popular stimulant cocaine, flourishes in tropical forests flanking the Andes Mountains. The mescaline-rich San Pedro cactus thrives along the barren Pacific coast, while the vilca tree grows on the dry and rocky foothills of the western slopes. To the east, a vine and shrub that grow deep in the Amazon River Basin can be combined into a dark red brew called ayahuasca, capable of creating intense visual and psychological experiences.

Scattered evidence suggests early societies along the Peruvian coast experimented with psychoactives as early as the third millennium B.C.E. More solid evidence is found at Chavín , which emerged 150 kilometers to the north by 1200 B.C.E. The temple complex remained in continuous use until decline set in after an earthquake around 550 B.C.E. Notably, it lacks defenses, palaces, or signs of an organized military. Chavín was not the center of a political state, but a pilgrimage site that for centuries drew visitors from afar.

“This system was not based on coercion,” Rick says. He envisions a series of nested secret societies overseen by shaman-priests who welcomed devotees from across the central Andes. An outer circle would have participated in public rituals in the large open-air square plaza in front of the main temple, whereas an elite entered a more intimate circular plaza ringed with friezes that include images of the mescaline-producing cactus. An even smaller circle might have ingested vilca in the underground galleries. Pottery found in some of these galleries comes from distinct regions around northern Peru, suggesting specific spaces were set aside for the exclusive use of visiting elites.

Not all researchers agree with Rick’s sweeping interpretation. Verónica Lema, an archaeologist at Argentina’s Museum of Anthropology who participated in the Chavín lab work, warns there is not yet enough evidence to understand what was taking place at Chavín, where, and for what purpose.

But the temple center does appear to have inspired people throughout the region, who built copycat versions. “Chavín seems like the mother culture,” said Justin Jennings, an archaeologist at the Royal Ontario Museum. “Something like 20 sites reflect its influence.” Rick believes these centers competed, tweaking their formulas and rituals to draw pilgrims. “This was a time of huge levels of creativity and competition,” he said. “If you didn’t innovate with the latest psychoactives, you were left behind.”

A 2025 study by anthropologist Cathy Costin of California State University, Northridge of ritual vessels from this era strengthens this view. Plentiful images of psychoactive plants on the vessels, found along Peru’s north coast, suggested they were used by shaman-leaders who could “conjure and harness the spirit world,” she says. That, in turn, “likely paved the way for the later development of power based on coercion and economic muscle.”

Costin based her conclusions on ancient images, whereas Michelle Young, an archaeologist at the American Museum of Natural History, partnered with chemists at Vanderbilt University to look for molecular traces of the drugs themselves. She uncovered artifacts excavated in the religious precinct of Atalla, a major hub between 1000 and 500 B.C.E. that lies 450 kilometers southeast of Chavín. The site’s residents built a massive, stepped platform using enormous cut stones, and made ceramics featuring part-human and part-animal figures strongly reminiscent of Chavín’s.

Among the discoveries were shards from cactus-shaped cups in the religious precinct. “This is pretty direct evidence that San Pedro was used within the temple, but we wanted chemical confirmation of mescaline,” Young says. At a Vanderbilt lab specializing in medical applications of advanced mass spectrometry, a technique that separates molecules by mass, chemistry graduate student Karly Serrano developed a new technique to analyze the cups’ contents without harming the vessel pieces. Liquid added to the shards absorbed residues, which could then be extracted and analyzed. “We are bringing the best tools in advancing medicine into archaeology,” says lab chief John McLean.

“We were not able to detect mescaline,” Serrano says, but she did identify 15 different sorts of lipids, fatty organic molecules. Among them were triglycerides, the most common kind of fat in the human body. Though these may point to contamination during the excavation, the lipids more likely come from degraded mescaline. To test that possibility—and to help gain insight into how that chemical signature breaks down over time—Young will provide Serrano with sherds from nonritual vessels for comparison.

In July 2024 , Capriles led a team that excavated a rock shelter near eastern Bolivia’s 6500-meter-high Mount Sajama. Beneath thick layers of dirt, burnt straw, and old llama manure, the archaeologists uncovered bits of textiles, decorated pottery, and other luxury goods. They also found a clue to drug-based state building in the Andes in the wake of the growth of regional powers such as the Moche and Nasca in the early centuries of this millennium, hundreds of years after Chavín’s demise. Among the finds was a wooden snuff tablet, a small tray used by vilca users to snort the psychoactive powder.

Capriles’s colleague Mirtha Gomez Saavedra, now at Penn State, dated the artifact to between 250 and 550 C.E. This is an era when the city of Tiwanaku on the shores of Lake Titicaca in today’s Bolivia rose to prominence. Its imagery, temples, and drug paraphernalia suggest a religious tradition centered on ritual use of vilca powder. After 200 C.E., the Tiahuanacans, as they are known, extended their influence south and west, and within 3 centuries their culture reached into northern Chile and Argentina and to Peru’s Pacific shores.

Given few signs of military conquest, Capriles and other archaeologists believe that ritual drug use and its religious underpinnings were key to spreading the Tiahuanacan way of life until the culture’s dissolution around 1000 C.E.

The tablet testifies to the drug trade that supported that culture. It was made from tropical palm wood and carved in a simple style common to the Amazon Basin, more than 500 arduous kilometers to the east, suggesting this lonely rock shelter was a node on a llama caravan route that moved prestige goods—including psychoactive plants from the rainforest or the eastern slopes of the Andes. “There were networks that began in the green ocean—the Amazon—and extended to the Pacific,” Capriles says. Local peoples who interacted with the traders, he adds, may have adopted use of the drugs—and the rituals that attended them.

More evidence for the trade came from a recent analysis of another find. In 2018, while working at the site of Cueva del Chileno, not far from Mount Sajama, Capriles’s team excavated a remarkably well-preserved leather bundle packed with snuff tablets, snuff tubes, spatulas made of llama bone, and a pouch of three fox snouts sewn together. The bundle, unique in Andean archaeology, dated firmly to the Tiahuanaco era.

Melanie Miller, now at the University of California San Francisco, scraped material from inside the pouch, and then used a combination of liquid chromatography and mass spectrometry to reveal five powerful psychoactive compounds, including cocaine. Among them were DMT and bufotenine, key ingredients of ayahuasca—the first evidence that the plants associated with the potent Amazonian brew were trafficked into the ancient Andean highlands, and also the first sign of ancient South Americans combining different medicinal ingredients.

This summer, Capriles began work at a newly discovered Tiahuanacan palace at Palaspata, in the highlands south of Bolivia’s capital of La Paz. One of his goals is to pinpoint sites of ritual activity for closer examination, though political unrest in Bolivia forced him to shorten his season.

In the meantime, the sheer abundance of drug paraphernalia in collections scattered across the Andes testifies to the extent of psychoactive use during the Tiahuanaco period, while offering tantalizing possibilities for residue analysis. In 2003, Kirk Costion, an archaeologist at Mesa Community College, was excavating near Moquegua, a valley settlement nestled in the mountains of south-central Peru. He befriended an expat Canadian, who one day guided him to a local church, and then to a small room inside packed with complete ancient pottery vessels.

“Sticking out of the pots were wooden spoons and at least half a dozen snuff tablets,” Costion says. A local priest had paid local boys to comb cemeteries and other ancient sites in the area for artifacts, but the collection was neglected after his death. Although the objects were later conserved and transferred to the local museum, they have yet to be carefully studied.

To the south, in the Chilean village of San Pedro de Atacama, another Catholic priest assembled remains from the area’s 11,000 years of human history, though in a more systematic way. Among the roughly 400,000 objects are hundreds of snuffing tablets, snorting tubes, spatulas, and a few pouches.

Recently, Lema also identified bird-bone tubes that were used as enema syringes , which allowed users to absorb vilca into the body even more efficiently than through the nostrils. These artifacts have “millimeter-precise dosing openings that can be adjusted by inserting or removing a stopper with a different dosing measurement,” she says. Most come from graves dating to the Tiahuanacan era.

“You find [drug paraphernalia] in roughly one in three graves,” says Constantino Torres, a retired ethnobotanist from Florida International University. “Since [the objects] were not buried with women, that means there was a really high rate of use among men, around 50%,” he adds, noting their style is also different from that popular in distant Tiwanaku. “This was not influence that came from an invading army. My speculation is that this was the spread of an ideology that lacked central control.”

That picture does not convince Lema, however. San Pedro de Atacama, she says, is “far from the central area, and no elite class developed,” making it difficult to tie Tiahuanaco-style religion and rituals with this remote site.

What is widely considered South America’s first empire emerged from the city of Huari in the Ayacucho Valley of the south-central Andes around 600 C.E., just after Tiahuanaco culture began its spread. Unlike the Tiahuanacans, the Wari mixed military might and diplomacy to expand their realm, often building fortified hilltop compounds. The nature of Wari society is still hotly debated, although most scholars see it as a rigid hierarchy dominated by a few elite families but lacking firm central control. And it, too, may have harnessed the power of psychoactives.

Snuff tablets are virtually unknown at Wari sites, even at those cheek-to-jowl with Tiahuanaco settlements in frontier areas such as Moquegua. Instead, Wari elite seemed to prefer grand feasts to consolidate their power over lower orders as well as neighboring elites. “The shared experience of eating and drinking forges affinity,” said Donna Nash, an Arizona State University archaeologist excavating at Cerro Mejía, a mountaintop Wari settlement near Moquegua. These feasts included beer fermented from the fruits of the molle tree common to the region.

Psychoactives, however, may have intensified the experience. At least some Wari brewers appear to have spiked batches of beer with vilca. Depictions of vilca seed pods on Wari feasting vessels show “beyond a reasonable doubt” that the drug was a key part of their rituals, according to Pat Knobloch, an independent scholar who has closely studied Wari iconography. The recent discovery of 13 vilca seeds among the remnants of feasts is additional evidence that they infused the substance into the alcoholic beverage. “This is not a slam dunk, but it is pretty good evidence,” Jennings adds.

When mixed with molle beer, the ground-up seed creates an experience that is far less intense but much longer lasting than when snorted. Molle contains a hydrocarbon that, when combined with a fermented beverage, further inhibits the drug’s psychoactive properties. “I might have a vilca beer with you, [even if] I’m not going to snort that stuff,” Jennings says.

He argues that this molle-vilca brew served a vital function in Wari expansion until its collapse around 1000 C.E. Studies in modern subjects have shown a similar drug, ayahuasca, creates increased empathy and openness that can last for months. Jennings suggests feastgoers experienced a similar “afterglow” that would have bonded them with the feast givers.

That, in turn, would have generated enthusiasm for Wari expansion and their labor-heavy irrigation projects in the highland deserts, though the Wari also conducted traditional military campaigns to gain territory.

To make a more solid case, Jennings hopes to obtain the chemical signature of the drug in a Wari beer cup, though he acknowledges “finding evidence of vilca reduced to powder and then mixed with a liquid is going to be hard.”

That might be necessary to convince some skeptical colleagues. “I will happily start thinking about vilca added to chicha once we have residue analysis from a cup or pitcher,” Nash says. Lema cautions against extrapolating the effects of psychedelics in modern Western societies onto the ancient Andean past. And Capriles notes the beliefs of the participants may be more important than a drug’s specific effects. The power of Catholic communion, for example, centers more in one’s personal understanding of the ritual than in the intoxicating nature of the wine.

Archaeological finds related to ritual psychoactive use appear to decline after 1000 C.E., with the dissolution of the Wari and Tiahuanaco realms. Two centuries later, the small ethnic group of the Inca began their push to dominate the Andes, and maize-brewed chicha beer eclipsed the old psychoactives. As the Inca Empire expanded, chicha drinking became a central ritual in vast plazas, and the brew was a common method of labor payment.

“What was important to the Inca was creating political alliances, rather than spreading a religious cult,” Capriles says. The Inca Empire eventually stretched from Ecuador to Chile. That sheer scale, according to archaeologists, precluded the widespread use of powerful drugs for social control. “Psychoactives are a problem for state-level societies,” Stanford’s Rick says. “They make people more creative, and less easy to regiment.” The substances remained known and in use, although later Spanish conquerors attempted to snuff out a practice they viewed as demonic.

Researchers expect more revelations from the newfound capacity to test for ancient psychoactives. For example, the analyses might one day be able to decipher the phylogenetic differences among varieties of coca. Linking those varieties to specific artifacts could shed light on long-distance trade, and on the cultural influences that spread the powerful mind-altering substance along with other valued goods.

But researchers emphasize that ancient chemical signatures won’t be enough to determine how psychoactive substances reshaped ancient politics and beliefs. “It is essential to increase the number of excavations and analyses,” Lema says, “but it is equally important to improve our ability to listen to both the ancient and present Indigenous Andean worlds.”

Projects are underway, for example, to collect psychoactive residues from the wares of Indigenous Amazonians, building data sets for detecting ancient psychoactive remains while gaining knowledge about contemporary practices and beliefs. The physical, emotional, and spiritual effects produced by these powerful substances in people today are also useful data points, Torres argues. “I grew up in hippie days, and participated in taking these things back then,” he says. “They influenced the way I see the world, and I assume this was even more so in the past.” He argues that “the human body remains the more reliable tool for investigating their impact,” noting a long history of psychoactive researchers experimenting with the substances they study.

Back at Chavín, Rick and his team are continuing to excavate, hoping to uncover more clues to its ancient rituals. Among their newer neighbors are two retreat centers catering to Peruvians and foreigners yearning to experience the power of psychoactives within the shadow of an ancient temple once devoted to their use.

These modern pilgrims can be a nuisance; Matt Sayre, an archaeologist at High Point University who works at Chavín, says they have, on occasion, bribed the park’s guards to gain access to the site at night to wander among the plazas and galleries.

But their motivations have a long history. “These plants are powerful tools of knowledge, allowing access to different dimensions,” says Ingrid Diaz, who owns a local vegetarian restaurant and operates one of the retreat centers with her American husband. In an unintended consequence of archaeology, bolstered by scientific advances, those eager to participate in psychoactive rituals are once again making the sacred trek to Chavín.

LISTEN TO THIS ARTICLE

A proposal to loosen hunting regulations in Italy has been condemned by scientists, who say the legislation will sideline the role of science in important decisions affecting the country’s wildlife.

The bill, approved by the Italian Senate in June and up for debate in the Chamber of Deputies this month, will make it possible for regional governments to extend the hunting season beyond current limits and will expand the areas where hunting is allowed—among other changes. Controversially, it will also remove the requirement for officials to consult the public research body responsible for environmental protection when making key hunting decisions.

“Not being able to make decisions on a scientific basis is bad,” says Elisa Anna Fano, president of the Italian Federation of the Sciences of Nature and Environment, which represents 20 Italian scientific societies. “Science should be respected.”

Hunting has become less popular in Italy, with the number of registered hunters declining from 1.7 million in 1980 to half a million last year. But the remaining hunters support new legislation introduced by Giorgia Meloni’s coalition government last year to update a law in effect since 1992. Farmers support it, too, arguing an update was due and that it will help control the population of wild boar and other wildlife that cause economic damage. But opponents have dubbed it the “ sparatutto ”—or “shoot everything”—bill.

A chief concern is how the bill alters the role of the Higher Institute for Environmental Protection and Research (ISPRA), a research body under the Ministry of Environment and Energy Security. One of ISPRA’s roles is to assess the conservation status of the country’s wildlife and its distribution, providing policymakers with robust scientific data. Currently, the national government must ask for ISPRA’s opinion when deciding which species can be hunted, and local administrations are obliged to consult with ISPRA when planning the hunting calendar. (In some cases, ISPRA’s opinion is binding.) The new bill would remove those requirements.

That could have important impacts on wildlife management, says Stefano Grignolio, president of the Italian Mammal Society. For example, “The lack of coordination and a national vision can be a problem and lead to inconsistent culling plans between neighboring areas,” he says. It could also give farmers and other local stakeholders more influence, he says.

Undermining the national research body “risks compromising the quality of public decisions and causing irreversible effects on ecological balances,” a group of 10 Italian scientific societies wrote in a June letter to the government. Many of the proposed changes in the bill, such as loosening regulations on the use of live decoys, could fall foul of the European Union’s conservation laws, they added. (The European Commission raised similar concerns in a December 2025 letter to the Italian government.)

The bill will also rebrand hunting as an activity that contributes to the safeguarding of biodiversity —a definition scientists have criticized. Selective hunting, which is backed by scientific knowledge of wildlife and is already used in Italy, can help control populations of wild boar and some other species, says Antonella Penna, president of the Italian Society of Marine Biology. But indiscriminate hunting can disrupt the local ecology, potentially even increasing the number of animals, and shouldn’t be considered a way to safeguard biodiversity, she says. In the case of wild boar, for instance, killing a herd’s matriarch—the only female that reproduces—can cause juvenile females to start to breed and drastically increase the number of newborns.

Massimo Buconi, president of Federcaccia, the largest Italian hunting association, says the bill will simply clarify the role of ISPRA, whose opinions he says have been given undue weight, rather than make any radical changes. Buconi says he doesn’t understand why the legislation has caused such an uproar: “We are astonished at the development of this debate around this bill.”

But others see the bill as a reflection of a larger antiscientific trend in Italy. “There is a much greater tendency to address wildlife issues with slightly more populist answers, answers that seem easy,” Grignolio says. Wildlife management is incredibly complex, he says, and should be addressed with the help of science.

A lot of work has been put into improving the EU’s wildlife conservation laws in recent years, Penna notes. That makes the new bill even more heartbreaking, she says: “The situation is serious.”

LISTEN TO THIS ARTICLE

Scientists seeking to make a dry academic paper more engaging often get by with a little help from the Beatles, according to a study published today in PLOS ONE .

The Fab Four are referenced here, there, and everywhere in the scientific literature , the researchers find, with the titles of at least 3200 peer-reviewed papers across disciplines quoting or making puns related to their song names or lyrics. This far outnumbers allusions to other top artists, suggesting the Beatles are uniquely adored across the (academic) universe, the researchers say.

Pop culture references can help papers stand out among the millions published each year, says Ken Hyland, a linguist at the University of East Anglia who was not involved in the work. “Titles, highlights, and abstracts need to grab attention—often with something unexpected. Otherwise, the paper is just fed into the gaping mouth of the machine.”

The new study is the brainchild of Beatles megafans Michiel Tjepkema, a law researcher at Radboud University, and Robert Kooij, a network scientist at the Delft University of Technology (TU Delft), who met at a Beatles convention 3 years ago. Sandwiched into a very musical program, Tjepkema gave a short talk on how scientists were using Beatles references in scientific articles, a phenomenon he’d noticed after a colleague published a paper referencing The Long and Winding Road .

Kooij, who was in the front row for the talk, approached Tjepkema afterward to suggest they come together to study the phenomenon systematically. Gabriel Budel, another network scientist at TU Delft, then joined the team. (Although Budel didn’t begin as a megafan, he “might have been stung by the Beatles bug through this article,” Tjepkema says hopefully.)

The researchers first compiled a list of Beatles songs or famous lyrics, excluding common words and phrases such as “because” or “ a day in the life.” They then searched the academic database Scopus for papers that matched these titles and lyrics exactly, such as “A ticket to ride: The unintended consequences of school transport subsidies,” and “A hard day’s night: Time use in shift workers.”

The team then used OpenAI’s large language model (LLM) GPT-4o-mini to assess whether papers that only partially matched their song titles and lyrics were wordplays on Beatles tracks. The LLM found puns including “The lung and winding road: Inflammation-driven platelet reactivity in Long COVID” and “Help! I need a WNT antibody. Help! Not just any antibody.”

The researchers found 2237 exact matches and 963 wordplays across fields from engineering to dentistry. Beatles references began to increase in the 1990s, which the authors put down to a generational effect: This is when ’60s kids were becoming academics. The band was also referenced far more than the next two best selling artists, Michael Jackson and Elvis. The 10 most referenced Beatles songs found their way into 1584 papers, compared with just 65 references for the top 10 Jackson songs and 196 for Elvis. “Since almost everybody likes and loves the Beatles, it’s kind of logical to make titles which refer to them,” Tjepkema says.

The researchers didn’t look at how the publications referencing the Beatles fared. But other work has shown pop culture references and humor can boost engagement, says Steve Heard, an ecologist at the University of New Brunswick. In a 2023 analysis of ecology and evolution papers, Heard found that authors seemed to give funnier titles to less important papers—but once the importance of the papers was considered, those with funnier titles had higher citation rates.

Scientists may be worried they won’t be taken seriously if using humor—and some journals do not allow humorous titles. (The authors of the new paper, which is titled “ How science gets by with a little help from the Beatles,” had to request an exemption from PLOS.) But attitudes are changing. “I think the rigidness of scientific writing has been relaxed,” says Arun Chandrasekaran, a nanoscientist at the University at Albany who has written on pop culture and science.

Adding elements of fun to a paper can also make it more accessible, says Kristin Sainani, an epidemiologist at Stanford University who teaches science writing. When scientists see that writing doesn’t always have to be stiff and formal, they are less likely to use jargony, bloated prose.

In perhaps the best evidence that quoting the Beatles doesn’t detract from serious work, the 2017 paper that laid the blueprint for LLMs such as ChatGPT and Claude was titled “ Attention Is All You Need . ” Within the world of artificial intelligence, that paper appears in every little thing: To date, it has been cited more than 268,000 times.

Think for yourself, then, when choosing a title. And remember, if you say you want a (scientific) revolution, don’t be afraid of a bit of fun. The End.

LISTEN TO THIS ARTICLE

Gondwana may be getting a promotion.

Long overshadowed by Pangaea—the supercontinent it would later help form—Gondwana may itself have been a full-fledged supercontinent , according to a study published today in Science Advances . The researchers say Gondwana, centered around the South Pole, encompassed some 80% of all of Earth’s land about 550 million years ago.

The work highlights the importance of Gondwana, which during its formation may have seen the largest amount of mountain building in Earth’s history, says Trond Torsvik, a paleogeographer at the University of Oslo who was not involved with the study. “I always thought it was much bigger.”

The one-two punch of Gondwana and Pangaea is also a provocation to modelers of deep time, who have suggested supercontinents come and go in periodic cycles, with gaps of 500 million years or more between them, reflecting deeper shifts in mantle circulation: first the Nuna supercontinent 1.8 billion years ago, then Rodinia 1.2 billion years ago, then Pangaea 335 million years ago.

Geologists have long gone back and forth on the status of Gondwana, which was discovered more than 100 years ago when common plant fossils were found on widely separated continents. In 2012, Joseph Meert, a paleogeographer at the University of Florida, concluded supercontinents should at a minimum encompass 75% of all known continental crust . At the time, Gondwana was mostly thought to comprise present-day India, Australia, Africa, Antarctica, and South America—64% of known land. The 75% threshold was always arbitrary, Meert says. “A lot of people have taken it to be a significant exact number, and it’s not. I’m not the person to award supercontinent status to anyone.”

But since then, Gondwana’s size has grown, as geologists added continental fragments in the eastern United States, southern Europe, Turkey, Iran, and Pakistan. But other fragments that existed at the time, including much of China, have not traditionally been assigned to Gondwana. Magnetic signals preserved in rocks that can help determine where ancient crust was located are difficult to interpret during this period. The jigsaw puzzle is further complicated by the repeated opening and closing of narrow oceans in East Asia at the time.

The new study, led by Tao Wang of the Chinese Academy of Geological Sciences, takes a different approach. The researchers looked at ancient granite rocks beneath these disputed regions, compiling 25,000 previously published records of samarium and neodymium isotopes contained within them. The radioactive decay of samarium into neodymium offers a clock that allowed the scientists to estimate the age of the rocks from which the granites formed. Using these data, they created a map that showed how the continental roots of several large crustal blocks in Asia formed around the same time as Gondwana. “In effect, the [samarium-neodymium] maps allow us to see through mountain-forming events to the earlier continent-forming process,” says co-author Bill Collins, a geologist at Curtin University.

But identifying ancient crust with a common history is not the same as showing those pieces were physically joined together 550 million years ago, says David Evans, a paleogeographer at Yale University. “The isotopic data don’t, in my opinion, uniquely demonstrate that connection.” It’s simplest to call Gondwana a “megacontinent,” as a paper several years ago in Geology proposed, Evans adds. “Gondwana will always be a subset of Pangaea, so it doesn’t make sense to me to use the same term.”

The authors say their case does not rest on the isotopic ages alone, however. Adapting existing paleogeographic models, they reconstructed the location of these Asian blocks and found they formed a connected, ribbonlike structure along Gondwana’s previously known edge.

Traditionally, the Asian continental blocks the researchers are now assigning to Gondwana have not been included as part of Pangaea. Collins expects that will continue, as the ribbon of fragments they identified drifted away from what became Pangaea. But this may need to be re-evaluated, Meert says. “I’ll accept that Gondwana was bigger,” he says. “Then you have to accept that Pangaea was bigger as well.”

If the result holds, however, researchers who have stitched together grand theories on how supercontinents rhythmically come and go may need to rethink their work. Earth may simply not be as tidy as they would like, Torsvik says. “I think the modelers are not happy,” he says. “Gondwana really messes things up.”

LISTEN TO THIS ARTICLE

In 1969, the Federation of European Biochemical Societies held what may have been the first poster session at a scientific conference. With only about 15 posters, it would have been a far cry from the massive displays at most conferences today.

Some conference attendees value the in-person engagement opportunities that poster sessions can offer, including networking, helping early-career researchers become comfortable presenting their results, and real-time brainstorming and troubleshooting with others. But many bemoan the difficulty of navigating a vast hall of dense, often poorly designed displays to find the needle-in-a-haystack results that may be relevant or interesting to a particular researcher.

A new project called Posters.science is proposing that online perusal could help raise the visibility and utility of the lowly poster—and be accessible to any researcher, even those who couldn’t attend a conference. “I cannot be at all conferences, but if afterwards I can search them easily for topics I’m interested in, that could be pretty cool,” says Bhavesh Patel, a research data management specialist at the nonprofit California Medical Innovations Institute (CalMI2), which developed the platform.

The website, which debuted in June, provides a centralized index of more than 30,000 posters across multiple disciplines, tagged with information about the conference when available. That initial batch was largely drawn from existing repositories; researchers can also upload their own posters. Patel and his colleagues at CalMI2 say they hope Posters.science will surface posters that could help scientists accelerate discovery by refining their hypotheses, enriching literature searches, and finding prospective collaborators.

Efforts like this are important because most posters are effectively invisible, according to a preprint study posted on the arXiv server last month by Patel and colleagues. Of the estimated millions of posters presented annually, the research team found no more than 150,000 posters in 68 online repositories ; an undetermined number included only a title and abstract. The team estimates only 42% of the posters had been assigned a digital object identifier, a standardized code that makes them easier to find online and cite. And most were not easily searchable by, for instance, the name and date of the conference. (Some conferences archive posters on their own websites only for those who paid to attend the meeting.)

Posters.science may be helpful, says Iva Cheung of Simon Fraser University. But it alone will not resolve the problems of the traditional poster format, which she says needs to evolve. “I just don’t think that it’s an effective way of sharing research information or developing [professional] networks,” says Cheung, who teaches research literacy and once wrote a blog post titled “Why academic conference posters suck.”

She notes that a 3-by-4-foot poster is often too crammed with text to be easily digested by viewers—and even then, it often leaves out many details of a study. A recent trend in poster design encourages streamlining the information and providing links to supplementary online information. But then “you’re basically putting up this big sheet of paper that says ‘Look somewhere else,’” Cheung says. She would like to see an expansion of alternative formats some conferences have begun to offer for briefly presenting work, including live “lightning” talks, printed handouts, and digital posters containing video, audio, and slideshows that are displayed on large screens.

Patel welcomes such experimentation and also celebrates variety in traditional poster design. “I like the creative freedom people have to shape posters—it’s a way for us as researchers to express ourselves,” he says.

His study found that making posters available online may draw some readership. According to the team’s analysis of a subset of 38,000 posters for which page view data were readily available, on average each had attracted 247 views; about 90% received at least one.

To make posters more useful, Patel’s team converts the text of all PDFs indexed in Posters.science into a format that can be readily read and analyzed by artificial intelligence. They are seeking to expand the platform’s functionality to extract and convert information from tables and images as well—a technically challenging task because poster design and formatting are so variable.

A centralized online platform to access posters has long been needed, says Nicholas Rowe, an independent researcher in Finland who has studied the role of posters in scientific communication. Otherwise, there may be no other way to see their results: His 2019 dissertation reports that as few as 1% of conference posters are eventually published in peer-reviewed journals .

But he agrees with Cheung that posters also need an overhaul made possible by modern digital tools. “I’d like to see us as a scientific community stand up and say we’re not doing this as well as we could be. Let’s have proper discussions about how we can make this better.”

LISTEN TO THIS ARTICLE

Seven ancient human fossils and a cache of tools and animal remains found in a cave in southwestern China are illuminating the looks and lifeways of the mysterious Denisovans, who were close cousins of the Neanderthals and coexisted—and interbred—with modern humans, as well.

One fossil, a partial elbow, is the first Denisovan arm bone identified, offering a tantalizing glimpse of their dexterity—and stone and bone cutters and scrapers found nearby reveal how they put it to use.

The findings from the approximately 150,000-year-old fossils, reported in a pair of papers today in Nature , underscore how much scientists have learned about the Denisovans since they were first identified through ancient DNA analysis in 2010.

“For years we knew them mainly from DNA and a handful of fragmentary fossils,” says paleoanthropologist María Martinón-Torres of the Spanish National Research Centre for Human Evolution, who was not involved in the studies. “Denisovans are no longer just a molecular ghost.”

The fossils were discovered in Bianfu Cave, located in southwestern China’s Yunnan province, alongside simple tools and cut-marked animal bones. That makes the site “one of the only places yielding Denisovan remains in clear archaeological context,” says Eric Delson, a paleoanthropologist at the American Museum of Natural History who likewise was not involved in the work.

In 2019, a worker employed by a local government project to reforest the site of an abandoned limestone quarry spotted a suspiciously large number of animal bones or fossils. He alerted provincial authorities and the Yunnan Provincial Institute of Cultural Relics and Archaeology, which sent a team to conduct an initial excavation.

The karst cave was full of bones. A technique that reveals when sediment layers were last exposed to sunlight suggested the dirt the remains were found in was between 134,000 and 167,000 years old. The cave’s more than 60,000 fossil fragments contained many ancient bones that looked like the genus Homo —but it wasn’t clear which ancient human group.

To find out, researchers first tried to extract ancient DNA, to no avail, says Qiaomei Fu, a paleogeneticist at the Chinese Academy of Sciences’s (CAS’s) Institute of Vertebrate Paleontology and Paleoanthropology and co-author of the first Nature paper. So, the team next analyzed samples of the bones for ancient protein s—they “can be preserved for longer in environments that are not as cold” as Siberia’s Denisova Cave, she says.

Five of the fossils had fragments of collagen, a common protein found in tissues throughout the body. The chemical signatures of those fragments, called peptides, closely resembled those of other fossils previously identified as Denisovans. “These five do look like they match Denisovans,” says Richard Potts of the Smithsonian Institution’s Human Origins Program, who was not involved in the research.

The list of confirmed Denisovan fossils is not long: a finger bone and three teeth from Russia, a single tooth from Laos, a partial mandible from a cave in Tibet, another recovered from the sea floor near Taiwan, and a skull from northern China found in an abandoned well. Recently, scientists also reported in a preprint two partial Denisovan leg bones, a femur and tibia, dredged up by fishermen near Taiwan.

But those findings left a blank space on the map of eastern Asia: southwestern China. “Bianfu Cave is the first site in this region of China where we have hominin fossils identified as Denisovan,” says Hao Li, an archaeologist at CAS’s Institute of Tibetan Plateau Research and co-author of the second Nature paper .

The shape of the newly identified radius fragment, part of the elbow, suggests Denisovans had joints that would have allowed for heavy-duty lifting, like Neanderthals, but also agile movements, like modern humans, says Christoph Zollikofer, an anthropologist at the University of Zürich who was not involved in the research.

The tools found at Bianfu include stone and bone flakes and scrapers, used to remove meat from bones of prey. “The toolmaking and stone working is relatively simple,” Delson says. “Knock a few flakes off so you can use them, don’t spend time making it a beautiful shape.”

But Ben Marwick, an archaeologist at the University of Washington who was not involved in the research, sees evidence of complexity in the artifacts. “The Denisovans could work in multiple mediums, with both stone and bone, and were not limited by material,” he says.

And although the stone handiwork may have been “expedient,” he says the faunal remains, which include deer and water buffalo, indicate Denisovans in southwestern China had a clear hunting strategy: They focused on medium to large animals, knowing “what makes a good kill—what’s worth their time and effort.”