Science

The traditional Inca chronology has collapsed. New radiocarbon dates show the empire lasted longer than was previously thought—and met more resistance

Maya glyphic texts from the Classic period (250–900 CE) typically chronicle the exploits of historical or divine characters; everyday or functional records are rare. Here, the authors offer a reconstruction and transcription of a ‘microtext’ painted on an interior wall of Structure 10K-2 at the site of Xultun, Guatemala. The text records a unique astronomical formula that concludes with a name, attributing the work to an individual named Sak Tahn Waax (‘White-chested Fox’). To date, this is the only known example of a Classic Maya mathematician directly credited for their work, attesting to the value of its intellectual authorship.

Metagenomic databases hold vast quantities of unexplored genetic material that can seed tools for biotechnology and medicine, but discovery is limited by our ability to examine sequence data at scale. Here, we show that large language models (LLMs) in an autonomous harness can alleviate this bottleneck and surface novel molecular systems. We deployed Claude Code instances to survey reverse transcriptase (RT) loci across 1.9 billion protein clusters. Among the top candidates were array-associated RTs (ART), a novel family of jumbo-phage RTs coupled with an array of ~200-nucleotide units and a dedicated partner gene. Session transcripts show that Claude identified the ART family by directly examining its DNA sequences and recognizing the atypical repeats, a behavior attributable to specific Mythos 5 internal signals that respond to repeated DNA. ART arrays are highly expressed and appear as discrete units during Staphylococcus phage infection, suggesting an RT system directed by a repertoire of distinct RNAs. These findings indicate that LLMs can autonomously detect anomalies and drive analyses to initiate biological discoveries.

The question. Scientists have now read the DNA of so many microbes, from soil and seawater as much as from the lab, that the databases hold billions of genes nobody has ever looked at. Some of biology’s most useful tools were found by a person scrolling through that kind of data and noticing something odd. But no person can look at billions of anything. So the authors ask: can an AI, left to work on its own, go through the data, notice what is odd, and follow it up the way a good scientist would?

What kind of paper this is. Two things at once. It reports a discovery in biology, a family of genes nobody had described. And it studies the AI that made the discovery, to find out how it did it. All six authors work at Anthropic, the company that makes the Claude AI models, and the models they test are their own company’s.

How they did it. They built a “harness”, a program that runs many copies of Claude at once, each with a role. A worker plans a small task and does it, writing code and searching databases; a supervisor checks the work and can open new tasks when something interesting turns up; others keep a shared notebook and edit the reports. The goal was new systems built around reverse transcriptases, enzymes that copy RNA into DNA, and the hunting ground a database of about 1.9 billion groups of related proteins from some 11 million samples. After that run, people took over, still using Claude to write and run the analyses, and did one experiment in living bacteria. Last, they tested seven Claude models on the same puzzle, and looked inside one of them while it read the DNA.

What they found.

• The run lasted 21.5 hours with no human stepping in: 119 tasks, 949 working sessions, nineteen reports. The agents sorted about 200,000 groups of reverse transcriptases into nine classes and studied seventeen candidate partner genes closely. Only three proved to be new partners; the other fourteen were mislabelled genes, parts of known systems, or genes common around such enzymes without being tied to them.
• The discovery came from a side road. One agent decided a candidate partner was a false lead, but noticed that the enzyme beside it sat in the DNA of a giant virus that infects bacteria, and asked for a closer look. The next worker was told to check the DNA in front of these enzymes for an RNA like the one a retron uses. The enzymes beside the virus gene had almost none, but their close relatives had long stretches, and the worker pulled 2,900 letters of one straight into what it was reading. It wrote that it could “see by eye” a pattern repeated many times over. No repeat-finding program had been run. It then asked itself whether this was already known, and tested that before reporting.
• Following it up, the authors found 95 distinct versions of the enzyme, in giant viruses that infect bacteria and in environmental DNA predicted to come from viruses; 28 had a detectable repeat array in front. They named the family array-associated reverse transcriptases, ART. It is marked by three things: the array; a long extra front end on the enzyme, about 180 building blocks where other enzymes of this kind typically carry about 50 or fewer; and a partner gene right behind it. Its closest known relatives are retrons, bacterial systems that defend against viruses.
• The arrays look a little like CRISPR, the arrays bacteria use to remember past infections, but differ in telling ways. Each holds 3 to 21 copies of a repeat 15 to 49 letters long, whose middle reads the same on both DNA strands. The stretches between the repeats are 120 to 220 letters, all different, where CRISPR’s are about 30, and they are kept, in the same order, in two related viruses, where CRISPR’s are gained and lost between related strains. None of the genes CRISPR needs is anywhere near.
• In published data from another lab on one of these viruses, SA1, infecting Staphylococcus bacteria, RNA copied from the array made up as much as 8 percent of all the virus’s RNA fifteen minutes in, among the most plentiful it makes, cut into shorter pieces with the same ends each time. The authors put the SA1 system into E. coli and similar pieces appeared. Computer folding puts each repeat into a small hairpin.
• The partner genes come in three kinds with no detectable likeness in sequence or predicted shape, so the enzyme seems to have taken on a new partner more than once. The commonest, at 59 of the 95, is shaped like enzymes that attach small chemical tags to other molecules; whether it does anything of the kind is untested. Computer predictions suggest, with moderate confidence, that the partner grips the enzyme’s long front end.
• So a retron has one RNA, one enzyme and one partner; ART has the enzyme and the partner but a whole bank of RNAs. The authors suggest each RNA might form its own complex, so one enzyme could answer several different threats. They offer this as a hypothesis only.
• The discovery was not repeated. Almost every one of ten more runs with the same instructions that reached the survey came across ART genes, and two looked at them closely, but none read the DNA in front of the enzyme, and all ten missed the array. The authors put this down to how many enzymes there were to look at and to the system never doing the same thing twice.
• So they built a fixed test: 3,500 attempts across seven Claude models, each given the ART genes at one of five levels, from two protein sequences in the question to all 96 examples as files with full tools and the internet. A Claude model graded each report against ten features the authors chose. The four they call most capable (Opus 5.5, Mythos 5.1, Mythos 5, Opus 5) scored clearly above the other three (Opus 4.6, Opus 4.8, Sonnet 5). The surprise: tools made them worse at finding the array. With the DNA in front of them the best models found it at least 90 percent of the time; with files and tools, as little as 32. In 39 percent of those four models’ attempts with files, the model never read more than 200 letters in a row, too few to see a second repeat. The more DNA it read, the more often it found the array: from 29 to 76 percent for the four together, and up to 96 for Mythos 5.
• Last, they looked inside the model. AI models trained only on DNA are known to pick out repeats, and two did here. In Mythos 5 the authors first chose twelve signals that respond to repeats in made-up sequences; two are shown on the real array, quiet before it and strong by the third copy. Scrambling the letters of each copy silenced one on 12 of the 14 copies and the other on all 14. Neither is special to DNA: both fire on repeats in random letters and digits too. In the discovery session, the words “tandem repeat array” came straight after the model read that stretch.

Why it matters. The authors’ point is that data has outgrown the people who study it. The scientists who first published the genome of one of these viruses, MarsHill, noticed the enzyme and guessed there was an RNA in front of it, but described neither the repeats nor the partner. Here, they argue, the AI supplied both halves of a discovery: the sense of what normal looks like, and the judgment that something did not fit and was worth chasing. Enzymes of this kind have given biology some of its most important tools, from copying RNA in the lab to editing genes, though nobody yet knows whether ART will be one.

Things to keep in mind.

• Nothing about what ART does has been shown. The authors say plainly that they have not shown the enzyme works, that it uses the array’s RNAs, that it holds its partner, or what the system does for the virus. The one experiment in living cells shows only that the array’s RNA is made and cut into pieces, and each version was grown once.
• The discovery happened in one run of eleven. The fixed test suggests why the others missed it; it does not make the original repeatable.
• The test was graded by a Claude model, against features chosen by authors who already knew what ART looked like.
• The two internal signals shown were picked from the twelve after the authors had looked at the array. The paper shows they respond as the model reads the repeats; it does not test whether turning them off stops the model noticing. The abstract’s word “attributable” claims more than that.
• The authors are testing their own company’s models. This synopsis was written by one of the models in the test.

Words to know.

• Reverse transcriptase (RT): an enzyme that copies RNA back into DNA, the opposite of the cell’s usual direction.
• Phage: a virus that infects bacteria; a jumbo phage has an unusually large genome.
• Retron: a bacterial defence of an RNA, a reverse transcriptase and a partner protein, which kills an infected cell before the virus can spread.
• CRISPR array: rows of short repeats with pieces of past viruses between them, which bacteria use to recognise a virus again.
• Tandem repeat array: the same short stretch of DNA repeated many times in a row, with other sequence between the copies.
• Spacer: the sequence between two repeats in an array.
• Palindrome: in DNA, a stretch that reads the same on both strands, so its RNA copy can fold back into a hairpin.
• Metagenomic database: DNA read straight from soil or water samples, without growing the organisms first.
• Nucleotide (nt): one letter of DNA or RNA.
• Large language model (LLM): an AI trained on huge amounts of text to predict what comes next; Claude is one.
• Agent: an AI that takes actions, running code and searching databases, not just answering.
• Context: everything a language model is reading as it writes; a sequence left in a file it never opened is not in its context.

The evolutionary history of elusive organisms can be characterized through genomic analyses, which have the power to reveal previously unknown taxa even in groups assumed to be well studied, such as cats. We have analyzed complete genomes of 38 individuals from the Neotropical cat genus Leopardus, including 26 individuals representing multiple evolutionary units of the contentious tiger cat (Leopardus tigrinus) species complex. Eight genomes were generated from museum specimens, which allowed the first genetic assessment of the type locality for L. tigrinus in the Guiana Shield. We found that this complex comprises five distinct species, including a novel cat species, discovered in the Bolivian Yungas and described in this study as L. tilcayo. The Peruvian Yungas unit of this complex also represents a distinct taxonomic entity, which we describe here as a novel subspecies, L. tigrinus antisuyo. Our phylogenomic analyses resolve the evolutionary relationships among the tiger cat geographic units, thus stabilizing their recalcitrant taxonomy and enabling adequate conservation assessment of these threatened felids. We also address other aspects of their evolution, including biogeography, past episodes of interspecies admixture, demographic history of each taxonomic unit, and temporal changes in genetic diversity. Altogether, our results clarify the evolutionary history of a complex radiation of wild cats, reveal novel taxa, and serve as a basis for conservation planning on behalf of these elusive wild cats.

The question. Across the forests and savannas of South America live small spotted cats about the size of a house cat, called tiger cats or oncillas. For 250 years scientists have argued about how many kinds there really are. They all look alike, they live in remote places, and the very first one was named in 1775 from a drawing rather than an animal. Are they one species with local variations, or several species that have been lumped together? The answer matters, because a species gets its own protection and a “variety” does not.

How they studied it. The team read the complete DNA, the genome, of 38 cats from the genus Leopardus, which includes the tiger cats and their relatives such as the ocelot and the margay. Some samples came from living animals. Eight came from museum specimens, including a skull from Guyana collected decades ago; they drilled a tiny piece of the dense bone inside the ear, where DNA survives best, and read a genome from it. That mattered because the region where the first tiger cat was described, the Guiana Shield, had never had a single cat genetically tested. With the genomes in hand they built family trees in two ways: from the DNA in the cell nucleus, which comes from both parents, and from the DNA in mitochondria, which comes only from the mother. They cut the nuclear genome into nearly 23,000 pieces and built a dated tree from each piece, then asked how well the pieces agreed. They also estimated how the population of each kind of cat has grown or shrunk over the last two million years, from the pattern of variation within each genome.

What they found.

• The tiger cats are not one species. They fall into six groups that are each clearly separate on the family tree, and the authors recognise five species. Four had been proposed before.
• The fifth is brand new: a cat from the cloud forests of the Bolivian Andes, named Leopardus tilcayo after the name local people already use for it. It is the first new species of cat found anywhere in over a hundred years. The reference specimen is a live male caught near the village of Arapata in 2021.
• A sixth group, from the Peruvian Andes, is named as a subspecies, antisuyo, after Antisuyu, the eastern quarter of the Inca empire, which covered those same mountains.
• The oldest split among tiger cats happened about 2.3 million years ago, soon after cats first reached South America across the new land bridge from the north. The tilcayo branched off about 1.4 million years ago, during ice ages that broke the mountain forests into isolated patches.
• The strangest result is that the Peruvian cats are the closest relatives of the cats from the Guianas, on the far side of the Amazon, even though their homes could hardly be more different: cloud forest in Peru, savanna patches in the north. The authors suggest the two were once connected by a route around the edge of the Amazon.
• The mother-line DNA told a different story from the nuclear DNA almost everywhere, because these cats hybridised with each other and with other species long ago. About 15 percent of the tilcayo’s genome came from the Peruvian cats that way. This is why earlier studies, which used mitochondrial DNA alone, kept reaching the wrong conclusions.
• Every group has been shrinking since the last ice age, some by as much as 80 percent, and their genetic diversity is on the low side for cats.

Why it matters. Conservation law works species by species. Right now all these cats share a single entry on the world’s red list of threatened species, as the “northern tiger cat”, listed as Vulnerable. If they are really five species, each one is rarer than anyone thought, each lives in a different kind of habitat, and each needs its own assessment. The paper also shows that museum collections are not just storage: specimens gathered a century ago can answer questions the collectors never imagined. And it shows how much of the living world is still unknown. A new cat, in a mountain range people have lived in for thousands of years, with a local name all along.

Things to keep in mind. Some conclusions rest on very few animals, sometimes one or two per group, and the authors say their diversity numbers are only a first look. Nobody knows how far the tilcayo’s range extends, partly because it is easily mistaken for a margay on camera traps; the paper includes a note on telling the two apart. The Peruvian and Guianan cats might turn out to be separate species too; the authors chose the cautious option for now. And deciding where one species ends and another begins always involves judgement, since evolution is a continuous process and species names are lines that people draw across it.

Words to know.

• Genome: the complete set of DNA in an organism, its full genetic instruction book.
• Species complex: a group of closely related organisms that look so alike that it is unclear how many species they contain.
• Phylogenomics: building evolutionary family trees from whole genomes rather than from a few genes.
• Museomics: reading DNA from old museum specimens, such as bones and skins, to study evolution.
• Mitochondrial DNA: a small piece of DNA inside the cell’s power plants, inherited only from the mother.
• Nuclear DNA: the DNA in the cell nucleus, inherited from both parents, which makes up nearly all of the genome.
• Hybridisation: breeding between two different species or lineages, which mixes their DNA.
• Introgression: DNA from one species that ended up permanently in another through hybridisation long ago.
• Incomplete lineage sorting: a natural reason family trees from different parts of the genome can disagree, when old genetic variation is passed down and sorted at random.
• Holotype: the single specimen chosen to represent a newly named species, kept in a museum for reference.
• Subspecies: a distinct population within a species, given a third name, such as Leopardus tigrinus antisuyo.
• Effective population size: a genetic measure of how many breeding individuals a population behaves as if it has.
• Heterozygosity: how often the two copies of a gene in one animal differ; low values mean low genetic diversity.
• Yungas: the cloud forests on the eastern slopes of the Andes in Peru, Bolivia and Argentina.
• Guiana Shield: the ancient highland region of northern South America, covering Guyana, Suriname, French Guiana and parts of Venezuela and Brazil.
• Pleistocene: the geological epoch of the ice ages, from 2.6 million to about 12,000 years ago.
• IUCN Red List: the international catalogue of how threatened each species is, from Least Concern to Extinct.

This article examines the iconography of a type of Caracalla tetradrachm that has been newly attributed to Neapolis in Roman Palestine and whose reverse depicts a monumental altar decorated with statues of Tyche, Ephesian Artemis, and Kore Persephone. The study contextualizes these deities in the religious life of Neapolis and identifies the monument as an altar often depicted as a miniscule element in panoramic views of Mount Gerizim on the bronze coins of Neapolis. The tetradrachms provide, for the first time, a close-up view of this long-lost civic monument.

The question. In the Roman Near East, the big silver coins called tetradrachms almost all look the same: the emperor’s head on one side, an eagle on the other, and a symbol the size of a pinhead between the eagle’s feet to say which city made it. One very rare coin of the emperor Caracalla, struck around 215–217 CE, breaks the pattern. Its eagle stands on a huge altar with three statues in niches. Andreas Kropp, of the University of Nottingham, asks three things: which city made it, who are the three figures, and what building is this?

What kind of paper this is. It is detective work with pictures, not an experiment. The evidence is the coins themselves, compared with other coins from the region, with statues, with excavation reports and with ancient writers. It appeared in the Journal of Roman Archaeology in 2021.

How he studied it. For eighty years catalogues said these coins came from Byblos in Lebanon, because a scholar in 1940 thought the tiny altar looked like ones on Byblos’s bronze coins. Kropp found something better: one coin of the altar type was struck with the same front-side die (the engraved metal stamp that presses the design) as coins showing a mountain with two peaks. That mountain is Mount Gerizim, which stands over the city of Neapolis in Samaria (Nablus today) and appears on nothing else. A shared die means the same workshop, so all four coin types come from Neapolis, and the “Byblos mint” probably never existed. Then he matched every detail of the three tiny statues against the clearer pictures on the bronze coins of Neapolis and of neighbouring Sebaste.

What they found.

• Only four of the big-altar coins are known: one in a museum in Bern and three sold at auction between 2014 and 2019.
• The building is a solid block on a stepped base, divided into three niches by flat columns, the middle niche round-topped, the others pointed, with a patterned band across the top. There are no doors, so it is not a temple front. A coin from Emesa in Syria, struck the same year, shows a similar block with a small fire altar on its roof: a monumental altar as big as a building, with the working altar on top.
• The middle statue is Tyche, the goddess who stood for the city itself. She wears a crown shaped like city walls, a short tunic and a sword, rests one foot on the front of a ship, and holds a small bust of the emperor in her hand, a way of saying “we are loyal to Rome.”
• The left statue is Artemis of Ephesos, an ancient stiff-standing idol from western Turkey covered in rows of little bumps, with ribbons hanging from her hands. Neapolis is the only city in the whole region that put her on its coins.
• The right statue is Kore, also called Persephone, the goddess who was carried off to the underworld and returns each spring. She holds a tall torch in a raised right arm. What looks like a hand on her chest is a bunch of her veil, which Kropp only saw after studying the Sebaste coins.
• Once Kore is recognised, three Neapolis bronzes of the 250s catalogued as “Demeter?” or “standing female figure” turn out to be her too.
• The altar is also Neapolis’s usual mintmark on its silver, the way Tyre used a club and Sidon used Europa riding a bull. So this was the city’s emblem.
• The famous coins of Neapolis show Mount Gerizim with a temple on one peak and a small square altar with flames on the other. Kropp argues the tetradrachm shows that altar close up for the first time.
• On this same coin series the mountain is drawn for the first time inside a circle carried by an eagle, as if it were a small object. From the 240s almost every Neapolis coin (104 of 118 types in one collection) carries this little mountain, held up by eagles, goddesses or the Roman wolf.

Where it stood. The temple peak is known: Tell er-Ras, 831 metres high, where a Roman temple of Zeus 14 by 21 metres was dug up in the 1960s. The altar peak is probably the main summit, 881 metres, the holy place of the Samaritans, where excavators found a huge walled sanctuary of the 5th and 2nd centuries BCE with 16,000 coins and 400,000 animal bones, but nothing Roman. Ancient writers say Samaritans still gathered there in Roman times (in 67 CE the Roman army killed 11,600 of them), and Kropp thinks the altar was wiped out when a church was built on the summit in the 5th century CE.

Why it matters. A single coin, read carefully, has recovered a lost monument of a real city and corrected an error copied from catalogue to catalogue for eighty years. It also shows a city hijacking the emperor’s own currency to say something about itself: a mountain, a temple, an altar and a goddess that meant home.

Things to keep in mind. The whole reattribution rests on one shared die, and the photographs of it are in a different paper, because this journal will not print coins whose ownership history before 1970 is unknown. The figures are a few millimetres tall, so identifying them means trusting comparisons. No trace of the altar has been found in the ground, and Kropp says its exact spot is still unknown.

Words to know.

• Tetradrachm: a large silver coin, originally worth four drachmas, used across Roman Syria and Palestine for big payments.
• Die: the engraved metal stamp that strikes a coin’s design; two coins from the same die were made in the same workshop.
• Mintmark: a small symbol on a coin that identifies the city that made it.
• Reverse: the back side of a coin; the side with the emperor’s head is the obverse.
• Tyche: the goddess of a city’s fortune, shown with a crown shaped like city walls.
• Kore: “the Maiden”, another name for Persephone, daughter of the grain goddess Demeter.
• Niche: a recess in a wall made to hold a statue.
• Samaritans: a religious community related to the Jews, whose holy mountain is Gerizim.
• Iconography: the study of what the figures and symbols in a picture mean.
• Attribution: deciding which mint, artist or place a coin or object came from.

The Southern Ocean ecosystem has undergone extensive changes in the past two centuries driven by industrial sealing and whaling, climate change and commercial fishing. However, following the end of commercial whaling, some populations of whales in this region are recovering. Baleen whales are reliant on Antarctic krill, which is also the largest Southern Ocean fishery. Since 1993, krill catch has increased fourfold, buoyed by nutritional supplement and aquaculture industries. In this Perspective, we approximate baleen whale consumption of Antarctic krill before and after whaling to examine if the ecosystem can support both humans and whales as krill predators. Our back-of-the-envelope calculations suggest that current krill biomass cannot support both an expanding krill fishery and the recovery of whale populations to pre-whaling sizes, highlighting an emerging human-wildlife conflict. We then provide recommendations for enhancing sustainability in this region by reducing encounters with whales and bolstering the krill population. The Southern Ocean ecosystem is recovering from 20th-century industrial whaling, while the krill fishery in this region has grown rapidly and may expand further, driven by demand for supplements and aquaculture feed. This Perspective discusses how current krill biomass is unlikely to support both a growing krill fishery and rebounding whale populations in the Southern Ocean.

The question. For most of the twentieth century, whaling ships killed the great whales of the Southern Ocean by the hundreds of thousands. Whaling stopped in 1986, and some of those whales, humpbacks especially, are now coming back. But in the years the whales were gone, a new industry grew up in the same waters: fishing for Antarctic krill, the small shrimp-like animals that those whales eat almost nothing but. So the authors ask a plain question: is there enough krill for both the returning whales and the fishing boats?

What kind of paper this is. It is a Perspective, not a report of a new experiment. The authors, marine biologists at Stanford, Colorado and Germany’s Alfred Wegener Institute, pull together the history, do some rough arithmetic on how much krill whales eat, and then argue for what the people who manage the fishery should do. They call the arithmetic “back-of-the-envelope” themselves, and say so more than once.

How they did the arithmetic. Scientists have put tags on feeding whales and measured how much they eat. A humpback eats about 3 tonnes of krill a day when it is feeding hard, a fin whale about 2, a blue whale over 4. The authors multiply that by about 100 heavy-feeding days a year, and then by the number of whales. They do it twice: once for the whales alive today, and once for the numbers that lived before whaling began. Then they compare both answers to how much krill is actually in the water, which was measured by ship-mounted sonar in 2018 and 2019.

What they found.

• The history first. From 1904, whalers took about 1.5 million baleen whales from the Southern Ocean; four countries took over 90 percent of them. Within fifty years, less than 1 percent of blue whales and 5 percent of fin whales were left.
• Today, in the part of the Southern Ocean where krill are fished, the humpback and fin whales alone eat about 18 million tonnes of krill a year. The fishery catches about half a million tonnes. That is more than all the surviving blue whales in the region eat, because there are only about 900 of them.
• Before whaling, roughly 180,000 blue, fin and humpback whales lived there and would have eaten about 51 million tonnes a year. The sonar survey found about 63 million tonnes of krill in total. So even with no fishing at all, there is probably not enough krill today for the whales to return to their old numbers.
• The real problem is where the fishing happens, not just how much. One patch of sea northwest of the South Orkney Islands is less than 2 percent of its zone but takes over 96 percent of the fishing there, and nearly 30 percent of all krill ever caught in the Southern Ocean. It is also a major whale feeding ground. In 2022, boats were fishing there in the middle of a group of feeding fin whales. In 2021, three humpbacks died in krill nets; another did in 2022.

What they say should happen. The fishery is run by an international commission called CCAMLR, which can only change its rules when every member country agrees. The authors ask it to count the whales’ food needs when it sets catch limits; to add whales to the list of animals it monitors; to make boats move away when whales are within about 100 metres; to close the busiest whale feeding grounds, and the weeks in early summer when whales feed hardest; and to stop catching the female and young krill that will spawn next year’s swarms. Some krill companies have already agreed to voluntary no-fishing zones near penguin colonies, which the authors take as a hopeful sign.

Why it matters. Krill are the middle of the whole Southern Ocean food web, and their sinking droppings also carry carbon down into the deep sea, so keeping krill in the water is a benefit to everyone, not only to whales. The authors’ point is that “there is plenty of krill” and “the whales are recovering” can both be true and still add up to a conflict, because whales and boats want the same dense swarms in the same places at the same times.

Things to keep in mind. The whale-feeding numbers come from a handful of tagged animals, and nobody has measured feeding rates for Southern Ocean fin or blue whales directly. The number of feeding days is a guess. And the comparison is between what whales eat in a year and a single snapshot of how much krill exists, when krill are being born and eaten all the time. The authors would agree: their aim is to start the argument, not to finish it.

Words to know.

• Baleen whales: whales that filter food through plates in their mouths instead of teeth; blue, fin and humpback whales are all baleen whales.
• Krill: finger-length swimming crustaceans that live in huge swarms.
• Biomass: the total weight of all the animals of one kind in an area.
• CCAMLR: the commission of countries that manages fishing around Antarctica.
• Tonne: 1,000 kilograms.
• Perspective: a kind of journal article that argues a point of view from existing evidence rather than reporting a new experiment.